Z-Image-Turbo ControlNet (#12792)

* init

---------

Co-authored-by: github-actions[bot] <github-actions[bot]@users.noreply.github.com>

src/diffusers/__init__.py

@@ -279,6 +279,7 @@ else:
             "WanAnimateTransformer3DModel",
             "WanTransformer3DModel",
             "WanVACETransformer3DModel",
+            "ZImageControlNetModel",
             "ZImageTransformer2DModel",
             "attention_backend",
         ]
@@ -670,6 +671,8 @@ else:
             "WuerstchenCombinedPipeline",
             "WuerstchenDecoderPipeline",
             "WuerstchenPriorPipeline",
+            "ZImageControlNetInpaintPipeline",
+            "ZImageControlNetPipeline",
             "ZImageImg2ImgPipeline",
             "ZImagePipeline",
         ]
@@ -1017,6 +1020,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             WanAnimateTransformer3DModel,
             WanTransformer3DModel,
             WanVACETransformer3DModel,
+            ZImageControlNetModel,
             ZImageTransformer2DModel,
             attention_backend,
         )
@@ -1377,6 +1381,8 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             WuerstchenCombinedPipeline,
             WuerstchenDecoderPipeline,
             WuerstchenPriorPipeline,
+            ZImageControlNetInpaintPipeline,
+            ZImageControlNetPipeline,
             ZImageImg2ImgPipeline,
             ZImagePipeline,
         )

src/diffusers/loaders/single_file_model.py

@@ -49,6 +49,7 @@ from .single_file_utils import (
     convert_stable_cascade_unet_single_file_to_diffusers,
     convert_wan_transformer_to_diffusers,
     convert_wan_vae_to_diffusers,
+    convert_z_image_controlnet_checkpoint_to_diffusers,
     convert_z_image_transformer_checkpoint_to_diffusers,
     create_controlnet_diffusers_config_from_ldm,
     create_unet_diffusers_config_from_ldm,
@@ -172,11 +173,18 @@ SINGLE_FILE_LOADABLE_CLASSES = {
         "checkpoint_mapping_fn": convert_z_image_transformer_checkpoint_to_diffusers,
         "default_subfolder": "transformer",
     },
+    "ZImageControlNetModel": {
+        "checkpoint_mapping_fn": convert_z_image_controlnet_checkpoint_to_diffusers,
+    },
 }
 
 
 def _should_convert_state_dict_to_diffusers(model_state_dict, checkpoint_state_dict):
-    return not set(model_state_dict.keys()).issubset(set(checkpoint_state_dict.keys()))
+    model_state_dict_keys = set(model_state_dict.keys())
+    checkpoint_state_dict_keys = set(checkpoint_state_dict.keys())
+    is_subset = model_state_dict_keys.issubset(checkpoint_state_dict_keys)
+    is_match = model_state_dict_keys == checkpoint_state_dict_keys
+    return not (is_subset and is_match)
 
 
 def _get_single_file_loadable_mapping_class(cls):

src/diffusers/loaders/single_file_utils.py

@@ -121,6 +121,8 @@ CHECKPOINT_KEY_NAMES = {
     "instruct-pix2pix": "model.diffusion_model.input_blocks.0.0.weight",
     "lumina2": ["model.diffusion_model.cap_embedder.0.weight", "cap_embedder.0.weight"],
     "z-image-turbo": "cap_embedder.0.weight",
+    "z-image-turbo-controlnet": "control_all_x_embedder.2-1.weight",
+    "z-image-turbo-controlnet-2.x": "control_layers.14.adaLN_modulation.0.weight",
     "sana": [
         "blocks.0.cross_attn.q_linear.weight",
         "blocks.0.cross_attn.q_linear.bias",
@@ -220,6 +222,8 @@ DIFFUSERS_DEFAULT_PIPELINE_PATHS = {
     "cosmos-2.0-v2w-2B": {"pretrained_model_name_or_path": "nvidia/Cosmos-Predict2-2B-Video2World"},
     "cosmos-2.0-v2w-14B": {"pretrained_model_name_or_path": "nvidia/Cosmos-Predict2-14B-Video2World"},
     "z-image-turbo": {"pretrained_model_name_or_path": "Tongyi-MAI/Z-Image-Turbo"},
+    "z-image-turbo-controlnet": {"pretrained_model_name_or_path": "hlky/Z-Image-Turbo-Fun-Controlnet-Union"},
+    "z-image-turbo-controlnet-2.x": {"pretrained_model_name_or_path": "hlky/Z-Image-Turbo-Fun-Controlnet-Union-2.1"},
 }
 
 # Use to configure model sample size when original config is provided
@@ -779,6 +783,12 @@ def infer_diffusers_model_type(checkpoint):
         else:
             raise ValueError(f"Unexpected x_embedder shape: {x_embedder_shape} when loading Cosmos 2.0 model.")
 
+    elif CHECKPOINT_KEY_NAMES["z-image-turbo-controlnet-2.x"] in checkpoint:
+        model_type = "z-image-turbo-controlnet-2.x"
+
+    elif CHECKPOINT_KEY_NAMES["z-image-turbo-controlnet"] in checkpoint:
+        model_type = "z-image-turbo-controlnet"
+
     else:
         model_type = "v1"
 
@@ -3885,3 +3895,17 @@ def convert_z_image_transformer_checkpoint_to_diffusers(checkpoint, **kwargs):
             handler_fn_inplace(key, converted_state_dict)
 
     return converted_state_dict
+
+
+def convert_z_image_controlnet_checkpoint_to_diffusers(checkpoint, config, **kwargs):
+    if config["add_control_noise_refiner"] is None:
+        return checkpoint
+    elif config["add_control_noise_refiner"] == "control_noise_refiner":
+        return checkpoint
+    elif config["add_control_noise_refiner"] == "control_layers":
+        converted_state_dict = {
+            key: checkpoint.pop(key) for key in list(checkpoint.keys()) if not key.startswith("control_noise_refiner.")
+        }
+        return converted_state_dict
+    else:
+        raise ValueError("Unknown Z-Image Turbo ControlNet type.")

src/diffusers/models/__init__.py

@@ -66,6 +66,7 @@ if is_torch_available():
     _import_structure["controlnets.controlnet_sparsectrl"] = ["SparseControlNetModel"]
     _import_structure["controlnets.controlnet_union"] = ["ControlNetUnionModel"]
     _import_structure["controlnets.controlnet_xs"] = ["ControlNetXSAdapter", "UNetControlNetXSModel"]
+    _import_structure["controlnets.controlnet_z_image"] = ["ZImageControlNetModel"]
     _import_structure["controlnets.multicontrolnet"] = ["MultiControlNetModel"]
     _import_structure["controlnets.multicontrolnet_union"] = ["MultiControlNetUnionModel"]
     _import_structure["embeddings"] = ["ImageProjection"]
@@ -181,6 +182,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             SD3MultiControlNetModel,
             SparseControlNetModel,
             UNetControlNetXSModel,
+            ZImageControlNetModel,
         )
         from .embeddings import ImageProjection
         from .modeling_utils import ModelMixin

src/diffusers/models/controlnets/__init__.py

@@ -19,6 +19,7 @@ if is_torch_available():
     )
     from .controlnet_union import ControlNetUnionModel
     from .controlnet_xs import ControlNetXSAdapter, ControlNetXSOutput, UNetControlNetXSModel
+    from .controlnet_z_image import ZImageControlNetModel
     from .multicontrolnet import MultiControlNetModel
     from .multicontrolnet_union import MultiControlNetUnionModel
 

src/diffusers/models/controlnets/controlnet_z_image.py

@@ -0,0 +1,824 @@
+# Copyright 2025 Alibaba Z-Image Team and The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import math
+from typing import List, Literal, Optional
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from torch.nn.utils.rnn import pad_sequence
+
+from ...configuration_utils import ConfigMixin, register_to_config
+from ...loaders import PeftAdapterMixin
+from ...loaders.single_file_model import FromOriginalModelMixin
+from ...models.attention_processor import Attention
+from ...models.normalization import RMSNorm
+from ...utils.torch_utils import maybe_allow_in_graph
+from ..attention_dispatch import dispatch_attention_fn
+from ..controlnets.controlnet import zero_module
+from ..modeling_utils import ModelMixin
+
+
+ADALN_EMBED_DIM = 256
+SEQ_MULTI_OF = 32
+
+
+# Copied from diffusers.models.transformers.transformer_z_image.TimestepEmbedder
+class TimestepEmbedder(nn.Module):
+    def __init__(self, out_size, mid_size=None, frequency_embedding_size=256):
+        super().__init__()
+        if mid_size is None:
+            mid_size = out_size
+        self.mlp = nn.Sequential(
+            nn.Linear(frequency_embedding_size, mid_size, bias=True),
+            nn.SiLU(),
+            nn.Linear(mid_size, out_size, bias=True),
+        )
+
+        self.frequency_embedding_size = frequency_embedding_size
+
+    @staticmethod
+    def timestep_embedding(t, dim, max_period=10000):
+        with torch.amp.autocast("cuda", enabled=False):
+            half = dim // 2
+            freqs = torch.exp(
+                -math.log(max_period) * torch.arange(start=0, end=half, dtype=torch.float32, device=t.device) / half
+            )
+            args = t[:, None].float() * freqs[None]
+            embedding = torch.cat([torch.cos(args), torch.sin(args)], dim=-1)
+            if dim % 2:
+                embedding = torch.cat([embedding, torch.zeros_like(embedding[:, :1])], dim=-1)
+            return embedding
+
+    def forward(self, t):
+        t_freq = self.timestep_embedding(t, self.frequency_embedding_size)
+        weight_dtype = self.mlp[0].weight.dtype
+        compute_dtype = getattr(self.mlp[0], "compute_dtype", None)
+        if weight_dtype.is_floating_point:
+            t_freq = t_freq.to(weight_dtype)
+        elif compute_dtype is not None:
+            t_freq = t_freq.to(compute_dtype)
+        t_emb = self.mlp(t_freq)
+        return t_emb
+
+
+# Copied from diffusers.models.transformers.transformer_z_image.ZSingleStreamAttnProcessor
+class ZSingleStreamAttnProcessor:
+    """
+    Processor for Z-Image single stream attention that adapts the existing Attention class to match the behavior of the
+    original Z-ImageAttention module.
+    """
+
+    _attention_backend = None
+    _parallel_config = None
+
+    def __init__(self):
+        if not hasattr(F, "scaled_dot_product_attention"):
+            raise ImportError(
+                "ZSingleStreamAttnProcessor requires PyTorch 2.0. To use it, please upgrade PyTorch to version 2.0 or higher."
+            )
+
+    def __call__(
+        self,
+        attn: Attention,
+        hidden_states: torch.Tensor,
+        encoder_hidden_states: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        freqs_cis: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        query = attn.to_q(hidden_states)
+        key = attn.to_k(hidden_states)
+        value = attn.to_v(hidden_states)
+
+        query = query.unflatten(-1, (attn.heads, -1))
+        key = key.unflatten(-1, (attn.heads, -1))
+        value = value.unflatten(-1, (attn.heads, -1))
+
+        # Apply Norms
+        if attn.norm_q is not None:
+            query = attn.norm_q(query)
+        if attn.norm_k is not None:
+            key = attn.norm_k(key)
+
+        # Apply RoPE
+        def apply_rotary_emb(x_in: torch.Tensor, freqs_cis: torch.Tensor) -> torch.Tensor:
+            with torch.amp.autocast("cuda", enabled=False):
+                x = torch.view_as_complex(x_in.float().reshape(*x_in.shape[:-1], -1, 2))
+                freqs_cis = freqs_cis.unsqueeze(2)
+                x_out = torch.view_as_real(x * freqs_cis).flatten(3)
+                return x_out.type_as(x_in)  # todo
+
+        if freqs_cis is not None:
+            query = apply_rotary_emb(query, freqs_cis)
+            key = apply_rotary_emb(key, freqs_cis)
+
+        # Cast to correct dtype
+        dtype = query.dtype
+        query, key = query.to(dtype), key.to(dtype)
+
+        # From [batch, seq_len] to [batch, 1, 1, seq_len] -> broadcast to [batch, heads, seq_len, seq_len]
+        if attention_mask is not None and attention_mask.ndim == 2:
+            attention_mask = attention_mask[:, None, None, :]
+
+        # Compute joint attention
+        hidden_states = dispatch_attention_fn(
+            query,
+            key,
+            value,
+            attn_mask=attention_mask,
+            dropout_p=0.0,
+            is_causal=False,
+            backend=self._attention_backend,
+            parallel_config=self._parallel_config,
+        )
+
+        # Reshape back
+        hidden_states = hidden_states.flatten(2, 3)
+        hidden_states = hidden_states.to(dtype)
+
+        output = attn.to_out[0](hidden_states)
+        if len(attn.to_out) > 1:  # dropout
+            output = attn.to_out[1](output)
+
+        return output
+
+
+# Copied from diffusers.models.transformers.transformer_z_image.FeedForward
+class FeedForward(nn.Module):
+    def __init__(self, dim: int, hidden_dim: int):
+        super().__init__()
+        self.w1 = nn.Linear(dim, hidden_dim, bias=False)
+        self.w2 = nn.Linear(hidden_dim, dim, bias=False)
+        self.w3 = nn.Linear(dim, hidden_dim, bias=False)
+
+    def _forward_silu_gating(self, x1, x3):
+        return F.silu(x1) * x3
+
+    def forward(self, x):
+        return self.w2(self._forward_silu_gating(self.w1(x), self.w3(x)))
+
+
+@maybe_allow_in_graph
+# Copied from diffusers.models.transformers.transformer_z_image.ZImageTransformerBlock
+class ZImageTransformerBlock(nn.Module):
+    def __init__(
+        self,
+        layer_id: int,
+        dim: int,
+        n_heads: int,
+        n_kv_heads: int,
+        norm_eps: float,
+        qk_norm: bool,
+        modulation=True,
+    ):
+        super().__init__()
+        self.dim = dim
+        self.head_dim = dim // n_heads
+
+        # Refactored to use diffusers Attention with custom processor
+        # Original Z-Image params: dim, n_heads, n_kv_heads, qk_norm
+        self.attention = Attention(
+            query_dim=dim,
+            cross_attention_dim=None,
+            dim_head=dim // n_heads,
+            heads=n_heads,
+            qk_norm="rms_norm" if qk_norm else None,
+            eps=1e-5,
+            bias=False,
+            out_bias=False,
+            processor=ZSingleStreamAttnProcessor(),
+        )
+
+        self.feed_forward = FeedForward(dim=dim, hidden_dim=int(dim / 3 * 8))
+        self.layer_id = layer_id
+
+        self.attention_norm1 = RMSNorm(dim, eps=norm_eps)
+        self.ffn_norm1 = RMSNorm(dim, eps=norm_eps)
+
+        self.attention_norm2 = RMSNorm(dim, eps=norm_eps)
+        self.ffn_norm2 = RMSNorm(dim, eps=norm_eps)
+
+        self.modulation = modulation
+        if modulation:
+            self.adaLN_modulation = nn.Sequential(nn.Linear(min(dim, ADALN_EMBED_DIM), 4 * dim, bias=True))
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        attn_mask: torch.Tensor,
+        freqs_cis: torch.Tensor,
+        adaln_input: Optional[torch.Tensor] = None,
+    ):
+        if self.modulation:
+            assert adaln_input is not None
+            scale_msa, gate_msa, scale_mlp, gate_mlp = self.adaLN_modulation(adaln_input).unsqueeze(1).chunk(4, dim=2)
+            gate_msa, gate_mlp = gate_msa.tanh(), gate_mlp.tanh()
+            scale_msa, scale_mlp = 1.0 + scale_msa, 1.0 + scale_mlp
+
+            # Attention block
+            attn_out = self.attention(
+                self.attention_norm1(x) * scale_msa, attention_mask=attn_mask, freqs_cis=freqs_cis
+            )
+            x = x + gate_msa * self.attention_norm2(attn_out)
+
+            # FFN block
+            x = x + gate_mlp * self.ffn_norm2(self.feed_forward(self.ffn_norm1(x) * scale_mlp))
+        else:
+            # Attention block
+            attn_out = self.attention(self.attention_norm1(x), attention_mask=attn_mask, freqs_cis=freqs_cis)
+            x = x + self.attention_norm2(attn_out)
+
+            # FFN block
+            x = x + self.ffn_norm2(self.feed_forward(self.ffn_norm1(x)))
+
+        return x
+
+
+# Copied from diffusers.models.transformers.transformer_z_image.RopeEmbedder
+class RopeEmbedder:
+    def __init__(
+        self,
+        theta: float = 256.0,
+        axes_dims: List[int] = (16, 56, 56),
+        axes_lens: List[int] = (64, 128, 128),
+    ):
+        self.theta = theta
+        self.axes_dims = axes_dims
+        self.axes_lens = axes_lens
+        assert len(axes_dims) == len(axes_lens), "axes_dims and axes_lens must have the same length"
+        self.freqs_cis = None
+
+    @staticmethod
+    def precompute_freqs_cis(dim: List[int], end: List[int], theta: float = 256.0):
+        with torch.device("cpu"):
+            freqs_cis = []
+            for i, (d, e) in enumerate(zip(dim, end)):
+                freqs = 1.0 / (theta ** (torch.arange(0, d, 2, dtype=torch.float64, device="cpu") / d))
+                timestep = torch.arange(e, device=freqs.device, dtype=torch.float64)
+                freqs = torch.outer(timestep, freqs).float()
+                freqs_cis_i = torch.polar(torch.ones_like(freqs), freqs).to(torch.complex64)  # complex64
+                freqs_cis.append(freqs_cis_i)
+
+            return freqs_cis
+
+    def __call__(self, ids: torch.Tensor):
+        assert ids.ndim == 2
+        assert ids.shape[-1] == len(self.axes_dims)
+        device = ids.device
+
+        if self.freqs_cis is None:
+            self.freqs_cis = self.precompute_freqs_cis(self.axes_dims, self.axes_lens, theta=self.theta)
+            self.freqs_cis = [freqs_cis.to(device) for freqs_cis in self.freqs_cis]
+        else:
+            # Ensure freqs_cis are on the same device as ids
+            if self.freqs_cis[0].device != device:
+                self.freqs_cis = [freqs_cis.to(device) for freqs_cis in self.freqs_cis]
+
+        result = []
+        for i in range(len(self.axes_dims)):
+            index = ids[:, i]
+            result.append(self.freqs_cis[i][index])
+        return torch.cat(result, dim=-1)
+
+
+@maybe_allow_in_graph
+class ZImageControlTransformerBlock(nn.Module):
+    def __init__(
+        self,
+        layer_id: int,
+        dim: int,
+        n_heads: int,
+        n_kv_heads: int,
+        norm_eps: float,
+        qk_norm: bool,
+        modulation=True,
+        block_id=0,
+    ):
+        super().__init__()
+        self.dim = dim
+        self.head_dim = dim // n_heads
+
+        # Refactored to use diffusers Attention with custom processor
+        # Original Z-Image params: dim, n_heads, n_kv_heads, qk_norm
+        self.attention = Attention(
+            query_dim=dim,
+            cross_attention_dim=None,
+            dim_head=dim // n_heads,
+            heads=n_heads,
+            qk_norm="rms_norm" if qk_norm else None,
+            eps=1e-5,
+            bias=False,
+            out_bias=False,
+            processor=ZSingleStreamAttnProcessor(),
+        )
+
+        self.feed_forward = FeedForward(dim=dim, hidden_dim=int(dim / 3 * 8))
+        self.layer_id = layer_id
+
+        self.attention_norm1 = RMSNorm(dim, eps=norm_eps)
+        self.ffn_norm1 = RMSNorm(dim, eps=norm_eps)
+
+        self.attention_norm2 = RMSNorm(dim, eps=norm_eps)
+        self.ffn_norm2 = RMSNorm(dim, eps=norm_eps)
+
+        self.modulation = modulation
+        if modulation:
+            self.adaLN_modulation = nn.Sequential(nn.Linear(min(dim, ADALN_EMBED_DIM), 4 * dim, bias=True))
+
+        # Control variant start
+        self.block_id = block_id
+        if block_id == 0:
+            self.before_proj = zero_module(nn.Linear(self.dim, self.dim))
+        self.after_proj = zero_module(nn.Linear(self.dim, self.dim))
+
+    def forward(
+        self,
+        c: torch.Tensor,
+        x: torch.Tensor,
+        attn_mask: torch.Tensor,
+        freqs_cis: torch.Tensor,
+        adaln_input: Optional[torch.Tensor] = None,
+    ):
+        # Control
+        if self.block_id == 0:
+            c = self.before_proj(c) + x
+            all_c = []
+        else:
+            all_c = list(torch.unbind(c))
+            c = all_c.pop(-1)
+
+        # Compared to `ZImageTransformerBlock` x -> c
+        if self.modulation:
+            assert adaln_input is not None
+            scale_msa, gate_msa, scale_mlp, gate_mlp = self.adaLN_modulation(adaln_input).unsqueeze(1).chunk(4, dim=2)
+            gate_msa, gate_mlp = gate_msa.tanh(), gate_mlp.tanh()
+            scale_msa, scale_mlp = 1.0 + scale_msa, 1.0 + scale_mlp
+
+            # Attention block
+            attn_out = self.attention(
+                self.attention_norm1(c) * scale_msa, attention_mask=attn_mask, freqs_cis=freqs_cis
+            )
+            c = c + gate_msa * self.attention_norm2(attn_out)
+
+            # FFN block
+            c = c + gate_mlp * self.ffn_norm2(self.feed_forward(self.ffn_norm1(c) * scale_mlp))
+        else:
+            # Attention block
+            attn_out = self.attention(self.attention_norm1(c), attention_mask=attn_mask, freqs_cis=freqs_cis)
+            c = c + self.attention_norm2(attn_out)
+
+            # FFN block
+            c = c + self.ffn_norm2(self.feed_forward(self.ffn_norm1(c)))
+
+        # Control
+        c_skip = self.after_proj(c)
+        all_c += [c_skip, c]
+        c = torch.stack(all_c)
+        return c
+
+
+class ZImageControlNetModel(ModelMixin, ConfigMixin, PeftAdapterMixin, FromOriginalModelMixin):
+    _supports_gradient_checkpointing = True
+
+    @register_to_config
+    def __init__(
+        self,
+        control_layers_places: List[int] = None,
+        control_refiner_layers_places: List[int] = None,
+        control_in_dim=None,
+        add_control_noise_refiner: Optional[Literal["control_layers", "control_noise_refiner"]] = None,
+        all_patch_size=(2,),
+        all_f_patch_size=(1,),
+        dim=3840,
+        n_refiner_layers=2,
+        n_heads=30,
+        n_kv_heads=30,
+        norm_eps=1e-5,
+        qk_norm=True,
+    ):
+        super().__init__()
+        self.control_layers_places = control_layers_places
+        self.control_in_dim = control_in_dim
+        self.control_refiner_layers_places = control_refiner_layers_places
+        self.add_control_noise_refiner = add_control_noise_refiner
+
+        assert 0 in self.control_layers_places
+
+        # control blocks
+        self.control_layers = nn.ModuleList(
+            [
+                ZImageControlTransformerBlock(i, dim, n_heads, n_kv_heads, norm_eps, qk_norm, block_id=i)
+                for i in self.control_layers_places
+            ]
+        )
+
+        # control patch embeddings
+        all_x_embedder = {}
+        for patch_idx, (patch_size, f_patch_size) in enumerate(zip(all_patch_size, all_f_patch_size)):
+            x_embedder = nn.Linear(f_patch_size * patch_size * patch_size * self.control_in_dim, dim, bias=True)
+            all_x_embedder[f"{patch_size}-{f_patch_size}"] = x_embedder
+
+        self.control_all_x_embedder = nn.ModuleDict(all_x_embedder)
+        if self.add_control_noise_refiner == "control_layers":
+            self.control_noise_refiner = None
+        elif self.add_control_noise_refiner == "control_noise_refiner":
+            self.control_noise_refiner = nn.ModuleList(
+                [
+                    ZImageControlTransformerBlock(
+                        1000 + layer_id,
+                        dim,
+                        n_heads,
+                        n_kv_heads,
+                        norm_eps,
+                        qk_norm,
+                        modulation=True,
+                        block_id=layer_id,
+                    )
+                    for layer_id in range(n_refiner_layers)
+                ]
+            )
+        else:
+            self.control_noise_refiner = nn.ModuleList(
+                [
+                    ZImageTransformerBlock(
+                        1000 + layer_id,
+                        dim,
+                        n_heads,
+                        n_kv_heads,
+                        norm_eps,
+                        qk_norm,
+                        modulation=True,
+                    )
+                    for layer_id in range(n_refiner_layers)
+                ]
+            )
+
+        self.t_scale: Optional[float] = None
+        self.t_embedder: Optional[TimestepEmbedder] = None
+        self.all_x_embedder: Optional[nn.ModuleDict] = None
+        self.cap_embedder: Optional[nn.Sequential] = None
+        self.rope_embedder: Optional[RopeEmbedder] = None
+        self.noise_refiner: Optional[nn.ModuleList] = None
+        self.context_refiner: Optional[nn.ModuleList] = None
+        self.x_pad_token: Optional[nn.Parameter] = None
+        self.cap_pad_token: Optional[nn.Parameter] = None
+
+    @classmethod
+    def from_transformer(cls, controlnet, transformer):
+        controlnet.t_scale = transformer.t_scale
+        controlnet.t_embedder = transformer.t_embedder
+        controlnet.all_x_embedder = transformer.all_x_embedder
+        controlnet.cap_embedder = transformer.cap_embedder
+        controlnet.rope_embedder = transformer.rope_embedder
+        controlnet.noise_refiner = transformer.noise_refiner
+        controlnet.context_refiner = transformer.context_refiner
+        controlnet.x_pad_token = transformer.x_pad_token
+        controlnet.cap_pad_token = transformer.cap_pad_token
+        return controlnet
+
+    @staticmethod
+    # Copied from diffusers.models.transformers.transformer_z_image.ZImageTransformer2DModel.create_coordinate_grid
+    def create_coordinate_grid(size, start=None, device=None):
+        if start is None:
+            start = (0 for _ in size)
+
+        axes = [torch.arange(x0, x0 + span, dtype=torch.int32, device=device) for x0, span in zip(start, size)]
+        grids = torch.meshgrid(axes, indexing="ij")
+        return torch.stack(grids, dim=-1)
+
+    # Copied from diffusers.models.transformers.transformer_z_image.ZImageTransformer2DModel.patchify_and_embed
+    def patchify_and_embed(
+        self,
+        all_image: List[torch.Tensor],
+        all_cap_feats: List[torch.Tensor],
+        patch_size: int,
+        f_patch_size: int,
+    ):
+        pH = pW = patch_size
+        pF = f_patch_size
+        device = all_image[0].device
+
+        all_image_out = []
+        all_image_size = []
+        all_image_pos_ids = []
+        all_image_pad_mask = []
+        all_cap_pos_ids = []
+        all_cap_pad_mask = []
+        all_cap_feats_out = []
+
+        for i, (image, cap_feat) in enumerate(zip(all_image, all_cap_feats)):
+            ### Process Caption
+            cap_ori_len = len(cap_feat)
+            cap_padding_len = (-cap_ori_len) % SEQ_MULTI_OF
+            # padded position ids
+            cap_padded_pos_ids = self.create_coordinate_grid(
+                size=(cap_ori_len + cap_padding_len, 1, 1),
+                start=(1, 0, 0),
+                device=device,
+            ).flatten(0, 2)
+            all_cap_pos_ids.append(cap_padded_pos_ids)
+            # pad mask
+            cap_pad_mask = torch.cat(
+                [
+                    torch.zeros((cap_ori_len,), dtype=torch.bool, device=device),
+                    torch.ones((cap_padding_len,), dtype=torch.bool, device=device),
+                ],
+                dim=0,
+            )
+            all_cap_pad_mask.append(
+                cap_pad_mask if cap_padding_len > 0 else torch.zeros((cap_ori_len,), dtype=torch.bool, device=device)
+            )
+
+            # padded feature
+            cap_padded_feat = torch.cat([cap_feat, cap_feat[-1:].repeat(cap_padding_len, 1)], dim=0)
+            all_cap_feats_out.append(cap_padded_feat)
+
+            ### Process Image
+            C, F, H, W = image.size()
+            all_image_size.append((F, H, W))
+            F_tokens, H_tokens, W_tokens = F // pF, H // pH, W // pW
+
+            image = image.view(C, F_tokens, pF, H_tokens, pH, W_tokens, pW)
+            # "c f pf h ph w pw -> (f h w) (pf ph pw c)"
+            image = image.permute(1, 3, 5, 2, 4, 6, 0).reshape(F_tokens * H_tokens * W_tokens, pF * pH * pW * C)
+
+            image_ori_len = len(image)
+            image_padding_len = (-image_ori_len) % SEQ_MULTI_OF
+
+            image_ori_pos_ids = self.create_coordinate_grid(
+                size=(F_tokens, H_tokens, W_tokens),
+                start=(cap_ori_len + cap_padding_len + 1, 0, 0),
+                device=device,
+            ).flatten(0, 2)
+            image_padded_pos_ids = torch.cat(
+                [
+                    image_ori_pos_ids,
+                    self.create_coordinate_grid(size=(1, 1, 1), start=(0, 0, 0), device=device)
+                    .flatten(0, 2)
+                    .repeat(image_padding_len, 1),
+                ],
+                dim=0,
+            )
+            all_image_pos_ids.append(image_padded_pos_ids if image_padding_len > 0 else image_ori_pos_ids)
+            # pad mask
+            image_pad_mask = torch.cat(
+                [
+                    torch.zeros((image_ori_len,), dtype=torch.bool, device=device),
+                    torch.ones((image_padding_len,), dtype=torch.bool, device=device),
+                ],
+                dim=0,
+            )
+            all_image_pad_mask.append(
+                image_pad_mask
+                if image_padding_len > 0
+                else torch.zeros((image_ori_len,), dtype=torch.bool, device=device)
+            )
+            # padded feature
+            image_padded_feat = torch.cat(
+                [image, image[-1:].repeat(image_padding_len, 1)],
+                dim=0,
+            )
+            all_image_out.append(image_padded_feat if image_padding_len > 0 else image)
+
+        return (
+            all_image_out,
+            all_cap_feats_out,
+            all_image_size,
+            all_image_pos_ids,
+            all_cap_pos_ids,
+            all_image_pad_mask,
+            all_cap_pad_mask,
+        )
+
+    def patchify(
+        self,
+        all_image: List[torch.Tensor],
+        patch_size: int,
+        f_patch_size: int,
+    ):
+        pH = pW = patch_size
+        pF = f_patch_size
+        all_image_out = []
+
+        for i, image in enumerate(all_image):
+            ### Process Image
+            C, F, H, W = image.size()
+            F_tokens, H_tokens, W_tokens = F // pF, H // pH, W // pW
+
+            image = image.view(C, F_tokens, pF, H_tokens, pH, W_tokens, pW)
+            # "c f pf h ph w pw -> (f h w) (pf ph pw c)"
+            image = image.permute(1, 3, 5, 2, 4, 6, 0).reshape(F_tokens * H_tokens * W_tokens, pF * pH * pW * C)
+
+            image_ori_len = len(image)
+            image_padding_len = (-image_ori_len) % SEQ_MULTI_OF
+
+            # padded feature
+            image_padded_feat = torch.cat([image, image[-1:].repeat(image_padding_len, 1)], dim=0)
+            all_image_out.append(image_padded_feat)
+
+        return all_image_out
+
+    def forward(
+        self,
+        x: List[torch.Tensor],
+        t,
+        cap_feats: List[torch.Tensor],
+        control_context: List[torch.Tensor],
+        conditioning_scale: float = 1.0,
+        patch_size=2,
+        f_patch_size=1,
+    ):
+        if (
+            self.t_scale is None
+            or self.t_embedder is None
+            or self.all_x_embedder is None
+            or self.cap_embedder is None
+            or self.rope_embedder is None
+            or self.noise_refiner is None
+            or self.context_refiner is None
+            or self.x_pad_token is None
+            or self.cap_pad_token is None
+        ):
+            raise ValueError(
+                "Required modules are `None`, use `from_transformer` to share required modules from `transformer`."
+            )
+
+        assert patch_size in self.config.all_patch_size
+        assert f_patch_size in self.config.all_f_patch_size
+
+        bsz = len(x)
+        device = x[0].device
+        t = t * self.t_scale
+        t = self.t_embedder(t)
+
+        (
+            x,
+            cap_feats,
+            x_size,
+            x_pos_ids,
+            cap_pos_ids,
+            x_inner_pad_mask,
+            cap_inner_pad_mask,
+        ) = self.patchify_and_embed(x, cap_feats, patch_size, f_patch_size)
+
+        x_item_seqlens = [len(_) for _ in x]
+        assert all(_ % SEQ_MULTI_OF == 0 for _ in x_item_seqlens)
+        x_max_item_seqlen = max(x_item_seqlens)
+
+        control_context = self.patchify(control_context, patch_size, f_patch_size)
+        control_context = torch.cat(control_context, dim=0)
+        control_context = self.control_all_x_embedder[f"{patch_size}-{f_patch_size}"](control_context)
+
+        control_context[torch.cat(x_inner_pad_mask)] = self.x_pad_token
+        control_context = list(control_context.split(x_item_seqlens, dim=0))
+
+        control_context = pad_sequence(control_context, batch_first=True, padding_value=0.0)
+
+        # x embed & refine
+        x = torch.cat(x, dim=0)
+        x = self.all_x_embedder[f"{patch_size}-{f_patch_size}"](x)
+
+        # Match t_embedder output dtype to x for layerwise casting compatibility
+        adaln_input = t.type_as(x)
+        x[torch.cat(x_inner_pad_mask)] = self.x_pad_token
+        x = list(x.split(x_item_seqlens, dim=0))
+        x_freqs_cis = list(self.rope_embedder(torch.cat(x_pos_ids, dim=0)).split([len(_) for _ in x_pos_ids], dim=0))
+
+        x = pad_sequence(x, batch_first=True, padding_value=0.0)
+        x_freqs_cis = pad_sequence(x_freqs_cis, batch_first=True, padding_value=0.0)
+        # Clarify the length matches to satisfy Dynamo due to "Symbolic Shape Inference" to avoid compilation errors
+        x_freqs_cis = x_freqs_cis[:, : x.shape[1]]
+
+        x_attn_mask = torch.zeros((bsz, x_max_item_seqlen), dtype=torch.bool, device=device)
+        for i, seq_len in enumerate(x_item_seqlens):
+            x_attn_mask[i, :seq_len] = 1
+
+        if self.add_control_noise_refiner is not None:
+            if self.add_control_noise_refiner == "control_layers":
+                layers = self.control_layers
+            elif self.add_control_noise_refiner == "control_noise_refiner":
+                layers = self.control_noise_refiner
+            else:
+                raise ValueError(f"Unsupported `add_control_noise_refiner` type: {self.add_control_noise_refiner}.")
+            for layer in layers:
+                if torch.is_grad_enabled() and self.gradient_checkpointing:
+                    control_context = self._gradient_checkpointing_func(
+                        layer, control_context, x, x_attn_mask, x_freqs_cis, adaln_input
+                    )
+                else:
+                    control_context = layer(control_context, x, x_attn_mask, x_freqs_cis, adaln_input)
+
+            hints = torch.unbind(control_context)[:-1]
+            control_context = torch.unbind(control_context)[-1]
+            noise_refiner_block_samples = {
+                layer_idx: hints[idx] * conditioning_scale
+                for idx, layer_idx in enumerate(self.control_refiner_layers_places)
+            }
+        else:
+            noise_refiner_block_samples = None
+
+        if torch.is_grad_enabled() and self.gradient_checkpointing:
+            for layer_idx, layer in enumerate(self.noise_refiner):
+                x = self._gradient_checkpointing_func(layer, x, x_attn_mask, x_freqs_cis, adaln_input)
+                if noise_refiner_block_samples is not None:
+                    if layer_idx in noise_refiner_block_samples:
+                        x = x + noise_refiner_block_samples[layer_idx]
+        else:
+            for layer_idx, layer in enumerate(self.noise_refiner):
+                x = layer(x, x_attn_mask, x_freqs_cis, adaln_input)
+                if noise_refiner_block_samples is not None:
+                    if layer_idx in noise_refiner_block_samples:
+                        x = x + noise_refiner_block_samples[layer_idx]
+
+        # cap embed & refine
+        cap_item_seqlens = [len(_) for _ in cap_feats]
+        cap_max_item_seqlen = max(cap_item_seqlens)
+
+        cap_feats = torch.cat(cap_feats, dim=0)
+        cap_feats = self.cap_embedder(cap_feats)
+        cap_feats[torch.cat(cap_inner_pad_mask)] = self.cap_pad_token
+        cap_feats = list(cap_feats.split(cap_item_seqlens, dim=0))
+        cap_freqs_cis = list(
+            self.rope_embedder(torch.cat(cap_pos_ids, dim=0)).split([len(_) for _ in cap_pos_ids], dim=0)
+        )
+
+        cap_feats = pad_sequence(cap_feats, batch_first=True, padding_value=0.0)
+        cap_freqs_cis = pad_sequence(cap_freqs_cis, batch_first=True, padding_value=0.0)
+        # Clarify the length matches to satisfy Dynamo due to "Symbolic Shape Inference" to avoid compilation errors
+        cap_freqs_cis = cap_freqs_cis[:, : cap_feats.shape[1]]
+
+        cap_attn_mask = torch.zeros((bsz, cap_max_item_seqlen), dtype=torch.bool, device=device)
+        for i, seq_len in enumerate(cap_item_seqlens):
+            cap_attn_mask[i, :seq_len] = 1
+
+        if torch.is_grad_enabled() and self.gradient_checkpointing:
+            for layer in self.context_refiner:
+                cap_feats = self._gradient_checkpointing_func(layer, cap_feats, cap_attn_mask, cap_freqs_cis)
+        else:
+            for layer in self.context_refiner:
+                cap_feats = layer(cap_feats, cap_attn_mask, cap_freqs_cis)
+
+        # unified
+        unified = []
+        unified_freqs_cis = []
+        for i in range(bsz):
+            x_len = x_item_seqlens[i]
+            cap_len = cap_item_seqlens[i]
+            unified.append(torch.cat([x[i][:x_len], cap_feats[i][:cap_len]]))
+            unified_freqs_cis.append(torch.cat([x_freqs_cis[i][:x_len], cap_freqs_cis[i][:cap_len]]))
+        unified_item_seqlens = [a + b for a, b in zip(cap_item_seqlens, x_item_seqlens)]
+        assert unified_item_seqlens == [len(_) for _ in unified]
+        unified_max_item_seqlen = max(unified_item_seqlens)
+
+        unified = pad_sequence(unified, batch_first=True, padding_value=0.0)
+        unified_freqs_cis = pad_sequence(unified_freqs_cis, batch_first=True, padding_value=0.0)
+        unified_attn_mask = torch.zeros((bsz, unified_max_item_seqlen), dtype=torch.bool, device=device)
+        for i, seq_len in enumerate(unified_item_seqlens):
+            unified_attn_mask[i, :seq_len] = 1
+
+        ## ControlNet start
+        if not self.add_control_noise_refiner:
+            if torch.is_grad_enabled() and self.gradient_checkpointing:
+                for layer in self.control_noise_refiner:
+                    control_context = self._gradient_checkpointing_func(
+                        layer, control_context, x_attn_mask, x_freqs_cis, adaln_input
+                    )
+            else:
+                for layer in self.control_noise_refiner:
+                    control_context = layer(control_context, x_attn_mask, x_freqs_cis, adaln_input)
+
+        # unified
+        control_context_unified = []
+        for i in range(bsz):
+            x_len = x_item_seqlens[i]
+            cap_len = cap_item_seqlens[i]
+            control_context_unified.append(torch.cat([control_context[i][:x_len], cap_feats[i][:cap_len]]))
+        control_context_unified = pad_sequence(control_context_unified, batch_first=True, padding_value=0.0)
+
+        for layer in self.control_layers:
+            if torch.is_grad_enabled() and self.gradient_checkpointing:
+                control_context_unified = self._gradient_checkpointing_func(
+                    layer, control_context_unified, unified, unified_attn_mask, unified_freqs_cis, adaln_input
+                )
+            else:
+                control_context_unified = layer(
+                    control_context_unified, unified, unified_attn_mask, unified_freqs_cis, adaln_input
+                )
+
+        hints = torch.unbind(control_context_unified)[:-1]
+        controlnet_block_samples = {
+            layer_idx: hints[idx] * conditioning_scale for idx, layer_idx in enumerate(self.control_layers_places)
+        }
+        return controlnet_block_samples

src/diffusers/models/transformers/transformer_z_image.py

@@ -13,7 +13,7 @@
 # limitations under the License.
 
 import math
-from typing import List, Optional, Tuple
+from typing import Dict, List, Optional, Tuple
 
 import torch
 import torch.nn as nn
@@ -536,6 +536,7 @@ class ZImageTransformer2DModel(ModelMixin, ConfigMixin, PeftAdapterMixin, FromOr
         x: List[torch.Tensor],
         t,
         cap_feats: List[torch.Tensor],
+        controlnet_block_samples: Optional[Dict[int, torch.Tensor]] = None,
         patch_size=2,
         f_patch_size=1,
         return_dict: bool = True,
@@ -635,13 +636,19 @@ class ZImageTransformer2DModel(ModelMixin, ConfigMixin, PeftAdapterMixin, FromOr
             unified_attn_mask[i, :seq_len] = 1
 
         if torch.is_grad_enabled() and self.gradient_checkpointing:
-            for layer in self.layers:
+            for layer_idx, layer in enumerate(self.layers):
                 unified = self._gradient_checkpointing_func(
                     layer, unified, unified_attn_mask, unified_freqs_cis, adaln_input
                 )
+                if controlnet_block_samples is not None:
+                    if layer_idx in controlnet_block_samples:
+                        unified = unified + controlnet_block_samples[layer_idx]
         else:
-            for layer in self.layers:
+            for layer_idx, layer in enumerate(self.layers):
                 unified = layer(unified, unified_attn_mask, unified_freqs_cis, adaln_input)
+                if controlnet_block_samples is not None:
+                    if layer_idx in controlnet_block_samples:
+                        unified = unified + controlnet_block_samples[layer_idx]
 
         unified = self.all_final_layer[f"{patch_size}-{f_patch_size}"](unified, adaln_input)
         unified = list(unified.unbind(dim=0))

src/diffusers/pipelines/__init__.py

@@ -405,7 +405,12 @@ else:
         "Kandinsky5T2IPipeline",
         "Kandinsky5I2IPipeline",
     ]
-    _import_structure["z_image"] = ["ZImageImg2ImgPipeline", "ZImagePipeline"]
+    _import_structure["z_image"] = [
+        "ZImageImg2ImgPipeline",
+        "ZImagePipeline",
+        "ZImageControlNetPipeline",
+        "ZImageControlNetInpaintPipeline",
+    ]
     _import_structure["skyreels_v2"] = [
         "SkyReelsV2DiffusionForcingPipeline",
         "SkyReelsV2DiffusionForcingImageToVideoPipeline",
@@ -845,7 +850,12 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             WuerstchenDecoderPipeline,
             WuerstchenPriorPipeline,
         )
-        from .z_image import ZImageImg2ImgPipeline, ZImagePipeline
+        from .z_image import (
+            ZImageControlNetInpaintPipeline,
+            ZImageControlNetPipeline,
+            ZImageImg2ImgPipeline,
+            ZImagePipeline,
+        )
 
         try:
             if not is_onnx_available():

src/diffusers/pipelines/z_image/__init__.py

@@ -23,6 +23,8 @@ except OptionalDependencyNotAvailable:
 else:
     _import_structure["pipeline_output"] = ["ZImagePipelineOutput"]
     _import_structure["pipeline_z_image"] = ["ZImagePipeline"]
+    _import_structure["pipeline_z_image_controlnet"] = ["ZImageControlNetPipeline"]
+    _import_structure["pipeline_z_image_controlnet_inpaint"] = ["ZImageControlNetInpaintPipeline"]
     _import_structure["pipeline_z_image_img2img"] = ["ZImageImg2ImgPipeline"]
 
 
@@ -36,6 +38,8 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
     else:
         from .pipeline_output import ZImagePipelineOutput
         from .pipeline_z_image import ZImagePipeline
+        from .pipeline_z_image_controlnet import ZImageControlNetPipeline
+        from .pipeline_z_image_controlnet_inpaint import ZImageControlNetInpaintPipeline
         from .pipeline_z_image_img2img import ZImageImg2ImgPipeline
 
 else:

src/diffusers/pipelines/z_image/pipeline_z_image_controlnet.py

@@ -0,0 +1,725 @@
+# Copyright 2025 Alibaba Z-Image Team and The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import inspect
+from typing import Any, Callable, Dict, List, Optional, Union
+
+import torch
+from transformers import AutoTokenizer, PreTrainedModel
+
+from ...image_processor import PipelineImageInput, VaeImageProcessor
+from ...loaders import FromSingleFileMixin
+from ...models.autoencoders import AutoencoderKL
+from ...models.controlnets import ZImageControlNetModel
+from ...models.transformers import ZImageTransformer2DModel
+from ...pipelines.pipeline_utils import DiffusionPipeline
+from ...schedulers import FlowMatchEulerDiscreteScheduler
+from ...utils import logging, replace_example_docstring
+from ...utils.torch_utils import randn_tensor
+from .pipeline_output import ZImagePipelineOutput
+
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+EXAMPLE_DOC_STRING = """
+    Examples:
+        ```py
+        >>> import torch
+        >>> from diffusers import ZImageControlNetPipeline
+        >>> from diffusers import ZImageControlNetModel
+        >>> from diffusers.utils import load_image
+        >>> from huggingface_hub import hf_hub_download
+
+        >>> controlnet = ZImageControlNetModel.from_single_file(
+        ...     hf_hub_download(
+        ...         "alibaba-pai/Z-Image-Turbo-Fun-Controlnet-Union",
+        ...         filename="Z-Image-Turbo-Fun-Controlnet-Union.safetensors",
+        ...     ),
+        ...     torch_dtype=torch.bfloat16,
+        ... )
+
+        >>> # 2.1
+        >>> # controlnet = ZImageControlNetModel.from_single_file(
+        >>> #     hf_hub_download(
+        >>> #         "alibaba-pai/Z-Image-Turbo-Fun-Controlnet-Union-2.0",
+        >>> #         filename="Z-Image-Turbo-Fun-Controlnet-Union-2.1.safetensors",
+        >>> #     ),
+        >>> #     torch_dtype=torch.bfloat16,
+        >>> # )
+
+        >>> # 2.0 - `config` is required
+        >>> # controlnet = ZImageControlNetModel.from_single_file(
+        >>> #     hf_hub_download(
+        >>> #         "alibaba-pai/Z-Image-Turbo-Fun-Controlnet-Union-2.0",
+        >>> #         filename="Z-Image-Turbo-Fun-Controlnet-Union-2.0.safetensors",
+        >>> #     ),
+        >>> #     torch_dtype=torch.bfloat16,
+        >>> #     config="hlky/Z-Image-Turbo-Fun-Controlnet-Union-2.0",
+        >>> # )
+
+        >>> pipe = ZImageControlNetPipeline.from_pretrained(
+        ...     "Tongyi-MAI/Z-Image-Turbo", controlnet=controlnet, torch_dtype=torch.bfloat16
+        ... )
+        >>> pipe.to("cuda")
+
+        >>> # Optionally, set the attention backend to flash-attn 2 or 3, default is SDPA in PyTorch.
+        >>> # (1) Use flash attention 2
+        >>> # pipe.transformer.set_attention_backend("flash")
+        >>> # (2) Use flash attention 3
+        >>> # pipe.transformer.set_attention_backend("_flash_3")
+
+        >>> control_image = load_image(
+        ...     "https://huggingface.co/alibaba-pai/Z-Image-Turbo-Fun-Controlnet-Union/resolve/main/asset/pose.jpg?download=true"
+        ... )
+        >>> prompt = "一位年轻女子站在阳光明媚的海岸线上，白裙在轻拂的海风中微微飘动。她拥有一头鲜艳的紫色长发，在风中轻盈舞动，发间系着一个精致的黑色蝴蝶结，与身后柔和的蔚蓝天空形成鲜明对比。她面容清秀，眉目精致，透着一股甜美的青春气息；神情柔和，略带羞涩，目光静静地凝望着远方的地平线，双手自然交叠于身前，仿佛沉浸在思绪之中。在她身后，是辽阔无垠、波光粼粼的大海，阳光洒在海面上，映出温暖的金色光晕。"
+        >>> image = pipe(
+        ...     prompt,
+        ...     control_image=control_image,
+        ...     controlnet_conditioning_scale=0.75,
+        ...     height=1728,
+        ...     width=992,
+        ...     num_inference_steps=9,
+        ...     guidance_scale=0.0,
+        ...     generator=torch.Generator("cuda").manual_seed(43),
+        ... ).images[0]
+        >>> image.save("zimage.png")
+        ```
+"""
+
+
+# Copied from diffusers.pipelines.flux.pipeline_flux.calculate_shift
+def calculate_shift(
+    image_seq_len,
+    base_seq_len: int = 256,
+    max_seq_len: int = 4096,
+    base_shift: float = 0.5,
+    max_shift: float = 1.15,
+):
+    m = (max_shift - base_shift) / (max_seq_len - base_seq_len)
+    b = base_shift - m * base_seq_len
+    mu = image_seq_len * m + b
+    return mu
+
+
+# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.retrieve_latents
+def retrieve_latents(
+    encoder_output: torch.Tensor, generator: Optional[torch.Generator] = None, sample_mode: str = "sample"
+):
+    if hasattr(encoder_output, "latent_dist") and sample_mode == "sample":
+        return encoder_output.latent_dist.sample(generator)
+    elif hasattr(encoder_output, "latent_dist") and sample_mode == "argmax":
+        return encoder_output.latent_dist.mode()
+    elif hasattr(encoder_output, "latents"):
+        return encoder_output.latents
+    else:
+        raise AttributeError("Could not access latents of provided encoder_output")
+
+
+# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps
+def retrieve_timesteps(
+    scheduler,
+    num_inference_steps: Optional[int] = None,
+    device: Optional[Union[str, torch.device]] = None,
+    timesteps: Optional[List[int]] = None,
+    sigmas: Optional[List[float]] = None,
+    **kwargs,
+):
+    r"""
+    Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles
+    custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`.
+
+    Args:
+        scheduler (`SchedulerMixin`):
+            The scheduler to get timesteps from.
+        num_inference_steps (`int`):
+            The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps`
+            must be `None`.
+        device (`str` or `torch.device`, *optional*):
+            The device to which the timesteps should be moved to. If `None`, the timesteps are not moved.
+        timesteps (`List[int]`, *optional*):
+            Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed,
+            `num_inference_steps` and `sigmas` must be `None`.
+        sigmas (`List[float]`, *optional*):
+            Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed,
+            `num_inference_steps` and `timesteps` must be `None`.
+
+    Returns:
+        `Tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the
+        second element is the number of inference steps.
+    """
+    if timesteps is not None and sigmas is not None:
+        raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values")
+    if timesteps is not None:
+        accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys())
+        if not accepts_timesteps:
+            raise ValueError(
+                f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom"
+                f" timestep schedules. Please check whether you are using the correct scheduler."
+            )
+        scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs)
+        timesteps = scheduler.timesteps
+        num_inference_steps = len(timesteps)
+    elif sigmas is not None:
+        accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys())
+        if not accept_sigmas:
+            raise ValueError(
+                f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom"
+                f" sigmas schedules. Please check whether you are using the correct scheduler."
+            )
+        scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs)
+        timesteps = scheduler.timesteps
+        num_inference_steps = len(timesteps)
+    else:
+        scheduler.set_timesteps(num_inference_steps, device=device, **kwargs)
+        timesteps = scheduler.timesteps
+    return timesteps, num_inference_steps
+
+
+class ZImageControlNetPipeline(DiffusionPipeline, FromSingleFileMixin):
+    model_cpu_offload_seq = "text_encoder->transformer->vae"
+    _optional_components = []
+    _callback_tensor_inputs = ["latents", "prompt_embeds"]
+
+    def __init__(
+        self,
+        scheduler: FlowMatchEulerDiscreteScheduler,
+        vae: AutoencoderKL,
+        text_encoder: PreTrainedModel,
+        tokenizer: AutoTokenizer,
+        transformer: ZImageTransformer2DModel,
+        controlnet: ZImageControlNetModel,
+    ):
+        super().__init__()
+        controlnet = ZImageControlNetModel.from_transformer(controlnet, transformer)
+
+        self.register_modules(
+            vae=vae,
+            text_encoder=text_encoder,
+            tokenizer=tokenizer,
+            scheduler=scheduler,
+            transformer=transformer,
+            controlnet=controlnet,
+        )
+        self.vae_scale_factor = (
+            2 ** (len(self.vae.config.block_out_channels) - 1) if hasattr(self, "vae") and self.vae is not None else 8
+        )
+        self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor * 2)
+
+    def encode_prompt(
+        self,
+        prompt: Union[str, List[str]],
+        device: Optional[torch.device] = None,
+        do_classifier_free_guidance: bool = True,
+        negative_prompt: Optional[Union[str, List[str]]] = None,
+        prompt_embeds: Optional[List[torch.FloatTensor]] = None,
+        negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+        max_sequence_length: int = 512,
+    ):
+        prompt = [prompt] if isinstance(prompt, str) else prompt
+        prompt_embeds = self._encode_prompt(
+            prompt=prompt,
+            device=device,
+            prompt_embeds=prompt_embeds,
+            max_sequence_length=max_sequence_length,
+        )
+
+        if do_classifier_free_guidance:
+            if negative_prompt is None:
+                negative_prompt = ["" for _ in prompt]
+            else:
+                negative_prompt = [negative_prompt] if isinstance(negative_prompt, str) else negative_prompt
+            assert len(prompt) == len(negative_prompt)
+            negative_prompt_embeds = self._encode_prompt(
+                prompt=negative_prompt,
+                device=device,
+                prompt_embeds=negative_prompt_embeds,
+                max_sequence_length=max_sequence_length,
+            )
+        else:
+            negative_prompt_embeds = []
+        return prompt_embeds, negative_prompt_embeds
+
+    def _encode_prompt(
+        self,
+        prompt: Union[str, List[str]],
+        device: Optional[torch.device] = None,
+        prompt_embeds: Optional[List[torch.FloatTensor]] = None,
+        max_sequence_length: int = 512,
+    ) -> List[torch.FloatTensor]:
+        device = device or self._execution_device
+
+        if prompt_embeds is not None:
+            return prompt_embeds
+
+        if isinstance(prompt, str):
+            prompt = [prompt]
+
+        for i, prompt_item in enumerate(prompt):
+            messages = [
+                {"role": "user", "content": prompt_item},
+            ]
+            prompt_item = self.tokenizer.apply_chat_template(
+                messages,
+                tokenize=False,
+                add_generation_prompt=True,
+                enable_thinking=True,
+            )
+            prompt[i] = prompt_item
+
+        text_inputs = self.tokenizer(
+            prompt,
+            padding="max_length",
+            max_length=max_sequence_length,
+            truncation=True,
+            return_tensors="pt",
+        )
+
+        text_input_ids = text_inputs.input_ids.to(device)
+        prompt_masks = text_inputs.attention_mask.to(device).bool()
+
+        prompt_embeds = self.text_encoder(
+            input_ids=text_input_ids,
+            attention_mask=prompt_masks,
+            output_hidden_states=True,
+        ).hidden_states[-2]
+
+        embeddings_list = []
+
+        for i in range(len(prompt_embeds)):
+            embeddings_list.append(prompt_embeds[i][prompt_masks[i]])
+
+        return embeddings_list
+
+    def prepare_latents(
+        self,
+        batch_size,
+        num_channels_latents,
+        height,
+        width,
+        dtype,
+        device,
+        generator,
+        latents=None,
+    ):
+        height = 2 * (int(height) // (self.vae_scale_factor * 2))
+        width = 2 * (int(width) // (self.vae_scale_factor * 2))
+
+        shape = (batch_size, num_channels_latents, height, width)
+
+        if latents is None:
+            latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+        else:
+            if latents.shape != shape:
+                raise ValueError(f"Unexpected latents shape, got {latents.shape}, expected {shape}")
+            latents = latents.to(device)
+        return latents
+
+    # Copied from diffusers.pipelines.controlnet_sd3.pipeline_stable_diffusion_3_controlnet.StableDiffusion3ControlNetPipeline.prepare_image
+    def prepare_image(
+        self,
+        image,
+        width,
+        height,
+        batch_size,
+        num_images_per_prompt,
+        device,
+        dtype,
+        do_classifier_free_guidance=False,
+        guess_mode=False,
+    ):
+        if isinstance(image, torch.Tensor):
+            pass
+        else:
+            image = self.image_processor.preprocess(image, height=height, width=width)
+
+        image_batch_size = image.shape[0]
+
+        if image_batch_size == 1:
+            repeat_by = batch_size
+        else:
+            # image batch size is the same as prompt batch size
+            repeat_by = num_images_per_prompt
+
+        image = image.repeat_interleave(repeat_by, dim=0)
+
+        image = image.to(device=device, dtype=dtype)
+
+        if do_classifier_free_guidance and not guess_mode:
+            image = torch.cat([image] * 2)
+
+        return image
+
+    @property
+    def guidance_scale(self):
+        return self._guidance_scale
+
+    @property
+    def do_classifier_free_guidance(self):
+        return self._guidance_scale > 1
+
+    @property
+    def joint_attention_kwargs(self):
+        return self._joint_attention_kwargs
+
+    @property
+    def num_timesteps(self):
+        return self._num_timesteps
+
+    @property
+    def interrupt(self):
+        return self._interrupt
+
+    @torch.no_grad()
+    @replace_example_docstring(EXAMPLE_DOC_STRING)
+    def __call__(
+        self,
+        prompt: Union[str, List[str]] = None,
+        height: Optional[int] = None,
+        width: Optional[int] = None,
+        num_inference_steps: int = 50,
+        sigmas: Optional[List[float]] = None,
+        guidance_scale: float = 5.0,
+        control_image: PipelineImageInput = None,
+        controlnet_conditioning_scale: Union[float, List[float]] = 0.75,
+        cfg_normalization: bool = False,
+        cfg_truncation: float = 1.0,
+        negative_prompt: Optional[Union[str, List[str]]] = None,
+        num_images_per_prompt: Optional[int] = 1,
+        generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+        latents: Optional[torch.FloatTensor] = None,
+        prompt_embeds: Optional[List[torch.FloatTensor]] = None,
+        negative_prompt_embeds: Optional[List[torch.FloatTensor]] = None,
+        output_type: Optional[str] = "pil",
+        return_dict: bool = True,
+        joint_attention_kwargs: Optional[Dict[str, Any]] = None,
+        callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None,
+        callback_on_step_end_tensor_inputs: List[str] = ["latents"],
+        max_sequence_length: int = 512,
+    ):
+        r"""
+        Function invoked when calling the pipeline for generation.
+
+        Args:
+            prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`.
+                instead.
+            height (`int`, *optional*, defaults to 1024):
+                The height in pixels of the generated image.
+            width (`int`, *optional*, defaults to 1024):
+                The width in pixels of the generated image.
+            num_inference_steps (`int`, *optional*, defaults to 50):
+                The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+                expense of slower inference.
+            sigmas (`List[float]`, *optional*):
+                Custom sigmas to use for the denoising process with schedulers which support a `sigmas` argument in
+                their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is passed
+                will be used.
+            guidance_scale (`float`, *optional*, defaults to 5.0):
+                Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
+                `guidance_scale` is defined as `w` of equation 2. of [Imagen
+                Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
+                1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
+                usually at the expense of lower image quality.
+            cfg_normalization (`bool`, *optional*, defaults to False):
+                Whether to apply configuration normalization.
+            cfg_truncation (`float`, *optional*, defaults to 1.0):
+                The truncation value for configuration.
+            negative_prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts not to guide the image generation. If not defined, one has to pass
+                `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
+                less than `1`).
+            num_images_per_prompt (`int`, *optional*, defaults to 1):
+                The number of images to generate per prompt.
+            generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+                One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html)
+                to make generation deterministic.
+            latents (`torch.FloatTensor`, *optional*):
+                Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image
+                generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+                tensor will be generated by sampling using the supplied random `generator`.
+            prompt_embeds (`List[torch.FloatTensor]`, *optional*):
+                Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+                provided, text embeddings will be generated from `prompt` input argument.
+            negative_prompt_embeds (`List[torch.FloatTensor]`, *optional*):
+                Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+                weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+                argument.
+            output_type (`str`, *optional*, defaults to `"pil"`):
+                The output format of the generate image. Choose between
+                [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`~pipelines.stable_diffusion.ZImagePipelineOutput`] instead of a plain
+                tuple.
+            joint_attention_kwargs (`dict`, *optional*):
+                A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under
+                `self.processor` in
+                [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
+            callback_on_step_end (`Callable`, *optional*):
+                A function that calls at the end of each denoising steps during the inference. The function is called
+                with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int,
+                callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by
+                `callback_on_step_end_tensor_inputs`.
+            callback_on_step_end_tensor_inputs (`List`, *optional*):
+                The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list
+                will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the
+                `._callback_tensor_inputs` attribute of your pipeline class.
+            max_sequence_length (`int`, *optional*, defaults to 512):
+                Maximum sequence length to use with the `prompt`.
+
+        Examples:
+
+        Returns:
+            [`~pipelines.z_image.ZImagePipelineOutput`] or `tuple`: [`~pipelines.z_image.ZImagePipelineOutput`] if
+            `return_dict` is True, otherwise a `tuple`. When returning a tuple, the first element is a list with the
+            generated images.
+        """
+        height = height or 1024
+        width = width or 1024
+
+        vae_scale = self.vae_scale_factor * 2
+        if height % vae_scale != 0:
+            raise ValueError(
+                f"Height must be divisible by {vae_scale} (got {height}). "
+                f"Please adjust the height to a multiple of {vae_scale}."
+            )
+        if width % vae_scale != 0:
+            raise ValueError(
+                f"Width must be divisible by {vae_scale} (got {width}). "
+                f"Please adjust the width to a multiple of {vae_scale}."
+            )
+
+        device = self._execution_device
+
+        self._guidance_scale = guidance_scale
+        self._joint_attention_kwargs = joint_attention_kwargs
+        self._interrupt = False
+        self._cfg_normalization = cfg_normalization
+        self._cfg_truncation = cfg_truncation
+        # 2. Define call parameters
+        if prompt is not None and isinstance(prompt, str):
+            batch_size = 1
+        elif prompt is not None and isinstance(prompt, list):
+            batch_size = len(prompt)
+        else:
+            batch_size = len(prompt_embeds)
+
+        # If prompt_embeds is provided and prompt is None, skip encoding
+        if prompt_embeds is not None and prompt is None:
+            if self.do_classifier_free_guidance and negative_prompt_embeds is None:
+                raise ValueError(
+                    "When `prompt_embeds` is provided without `prompt`, "
+                    "`negative_prompt_embeds` must also be provided for classifier-free guidance."
+                )
+        else:
+            (
+                prompt_embeds,
+                negative_prompt_embeds,
+            ) = self.encode_prompt(
+                prompt=prompt,
+                negative_prompt=negative_prompt,
+                do_classifier_free_guidance=self.do_classifier_free_guidance,
+                prompt_embeds=prompt_embeds,
+                negative_prompt_embeds=negative_prompt_embeds,
+                device=device,
+                max_sequence_length=max_sequence_length,
+            )
+
+        # 4. Prepare latent variables
+        num_channels_latents = self.transformer.in_channels
+
+        control_image = self.prepare_image(
+            image=control_image,
+            width=width,
+            height=height,
+            batch_size=batch_size * num_images_per_prompt,
+            num_images_per_prompt=num_images_per_prompt,
+            device=device,
+            dtype=self.vae.dtype,
+        )
+        height, width = control_image.shape[-2:]
+        control_image = retrieve_latents(self.vae.encode(control_image), generator=generator, sample_mode="argmax")
+        control_image = (control_image - self.vae.config.shift_factor) * self.vae.config.scaling_factor
+        control_image = control_image.unsqueeze(2)
+
+        if num_channels_latents != self.controlnet.config.control_in_dim:
+            # For model version 2.0
+            control_image = torch.cat(
+                [
+                    control_image,
+                    torch.zeros(
+                        control_image.shape[0],
+                        self.controlnet.config.control_in_dim - num_channels_latents,
+                        *control_image.shape[2:],
+                    ).to(device=control_image.device, dtype=control_image.dtype),
+                ],
+                dim=1,
+            )
+
+        latents = self.prepare_latents(
+            batch_size * num_images_per_prompt,
+            num_channels_latents,
+            height,
+            width,
+            torch.float32,
+            device,
+            generator,
+            latents,
+        )
+
+        # Repeat prompt_embeds for num_images_per_prompt
+        if num_images_per_prompt > 1:
+            prompt_embeds = [pe for pe in prompt_embeds for _ in range(num_images_per_prompt)]
+            if self.do_classifier_free_guidance and negative_prompt_embeds:
+                negative_prompt_embeds = [npe for npe in negative_prompt_embeds for _ in range(num_images_per_prompt)]
+
+        actual_batch_size = batch_size * num_images_per_prompt
+        image_seq_len = (latents.shape[2] // 2) * (latents.shape[3] // 2)
+
+        # 5. Prepare timesteps
+        mu = calculate_shift(
+            image_seq_len,
+            self.scheduler.config.get("base_image_seq_len", 256),
+            self.scheduler.config.get("max_image_seq_len", 4096),
+            self.scheduler.config.get("base_shift", 0.5),
+            self.scheduler.config.get("max_shift", 1.15),
+        )
+        self.scheduler.sigma_min = 0.0
+        scheduler_kwargs = {"mu": mu}
+        timesteps, num_inference_steps = retrieve_timesteps(
+            self.scheduler,
+            num_inference_steps,
+            device,
+            sigmas=sigmas,
+            **scheduler_kwargs,
+        )
+        num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0)
+        self._num_timesteps = len(timesteps)
+
+        # 6. Denoising loop
+        with self.progress_bar(total=num_inference_steps) as progress_bar:
+            for i, t in enumerate(timesteps):
+                if self.interrupt:
+                    continue
+
+                # broadcast to batch dimension in a way that's compatible with ONNX/Core ML
+                timestep = t.expand(latents.shape[0])
+                timestep = (1000 - timestep) / 1000
+                # Normalized time for time-aware config (0 at start, 1 at end)
+                t_norm = timestep[0].item()
+
+                # Handle cfg truncation
+                current_guidance_scale = self.guidance_scale
+                if (
+                    self.do_classifier_free_guidance
+                    and self._cfg_truncation is not None
+                    and float(self._cfg_truncation) <= 1
+                ):
+                    if t_norm > self._cfg_truncation:
+                        current_guidance_scale = 0.0
+
+                # Run CFG only if configured AND scale is non-zero
+                apply_cfg = self.do_classifier_free_guidance and current_guidance_scale > 0
+
+                if apply_cfg:
+                    latents_typed = latents.to(self.transformer.dtype)
+                    latent_model_input = latents_typed.repeat(2, 1, 1, 1)
+                    prompt_embeds_model_input = prompt_embeds + negative_prompt_embeds
+                    timestep_model_input = timestep.repeat(2)
+                else:
+                    latent_model_input = latents.to(self.transformer.dtype)
+                    prompt_embeds_model_input = prompt_embeds
+                    timestep_model_input = timestep
+
+                latent_model_input = latent_model_input.unsqueeze(2)
+                latent_model_input_list = list(latent_model_input.unbind(dim=0))
+
+                controlnet_block_samples = self.controlnet(
+                    latent_model_input_list,
+                    timestep_model_input,
+                    prompt_embeds_model_input,
+                    control_image,
+                    conditioning_scale=controlnet_conditioning_scale,
+                )
+
+                model_out_list = self.transformer(
+                    latent_model_input_list,
+                    timestep_model_input,
+                    prompt_embeds_model_input,
+                    controlnet_block_samples=controlnet_block_samples,
+                )[0]
+
+                if apply_cfg:
+                    # Perform CFG
+                    pos_out = model_out_list[:actual_batch_size]
+                    neg_out = model_out_list[actual_batch_size:]
+
+                    noise_pred = []
+                    for j in range(actual_batch_size):
+                        pos = pos_out[j].float()
+                        neg = neg_out[j].float()
+
+                        pred = pos + current_guidance_scale * (pos - neg)
+
+                        # Renormalization
+                        if self._cfg_normalization and float(self._cfg_normalization) > 0.0:
+                            ori_pos_norm = torch.linalg.vector_norm(pos)
+                            new_pos_norm = torch.linalg.vector_norm(pred)
+                            max_new_norm = ori_pos_norm * float(self._cfg_normalization)
+                            if new_pos_norm > max_new_norm:
+                                pred = pred * (max_new_norm / new_pos_norm)
+
+                        noise_pred.append(pred)
+
+                    noise_pred = torch.stack(noise_pred, dim=0)
+                else:
+                    noise_pred = torch.stack([t.float() for t in model_out_list], dim=0)
+
+                noise_pred = noise_pred.squeeze(2)
+                noise_pred = -noise_pred
+
+                # compute the previous noisy sample x_t -> x_t-1
+                latents = self.scheduler.step(noise_pred.to(torch.float32), t, latents, return_dict=False)[0]
+                assert latents.dtype == torch.float32
+
+                if callback_on_step_end is not None:
+                    callback_kwargs = {}
+                    for k in callback_on_step_end_tensor_inputs:
+                        callback_kwargs[k] = locals()[k]
+                    callback_outputs = callback_on_step_end(self, i, t, callback_kwargs)
+
+                    latents = callback_outputs.pop("latents", latents)
+                    prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds)
+                    negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds)
+
+                # call the callback, if provided
+                if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
+                    progress_bar.update()
+
+        if output_type == "latent":
+            image = latents
+
+        else:
+            latents = latents.to(self.vae.dtype)
+            latents = (latents / self.vae.config.scaling_factor) + self.vae.config.shift_factor
+
+            image = self.vae.decode(latents, return_dict=False)[0]
+            image = self.image_processor.postprocess(image, output_type=output_type)
+
+        # Offload all models
+        self.maybe_free_model_hooks()
+
+        if not return_dict:
+            return (image,)
+
+        return ZImagePipelineOutput(images=image)

src/diffusers/pipelines/z_image/pipeline_z_image_controlnet_inpaint.py

@@ -0,0 +1,747 @@
+# Copyright 2025 Alibaba Z-Image Team and The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import inspect
+from typing import Any, Callable, Dict, List, Optional, Union
+
+import torch
+import torch.nn.functional as F
+from transformers import AutoTokenizer, PreTrainedModel
+
+from ...image_processor import PipelineImageInput, VaeImageProcessor
+from ...loaders import FromSingleFileMixin
+from ...models.autoencoders import AutoencoderKL
+from ...models.controlnets import ZImageControlNetModel
+from ...models.transformers import ZImageTransformer2DModel
+from ...pipelines.pipeline_utils import DiffusionPipeline
+from ...schedulers import FlowMatchEulerDiscreteScheduler
+from ...utils import logging, replace_example_docstring
+from ...utils.torch_utils import randn_tensor
+from .pipeline_output import ZImagePipelineOutput
+
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+EXAMPLE_DOC_STRING = """
+    Examples:
+        ```py
+        >>> import torch
+        >>> from diffusers import ZImageControlNetInpaintPipeline
+        >>> from diffusers import ZImageControlNetModel
+        >>> from diffusers.utils import load_image
+        >>> from huggingface_hub import hf_hub_download
+
+        >>> controlnet = ZImageControlNetModel.from_single_file(
+        ...     hf_hub_download(
+        ...         "alibaba-pai/Z-Image-Turbo-Fun-Controlnet-Union-2.0",
+        ...         filename="Z-Image-Turbo-Fun-Controlnet-Union-2.1.safetensors",
+        ...     ),
+        ...     torch_dtype=torch.bfloat16,
+        ... )
+
+        >>> # 2.0 - `config` is required
+        >>> # controlnet = ZImageControlNetModel.from_single_file(
+        >>> #     hf_hub_download(
+        >>> #         "alibaba-pai/Z-Image-Turbo-Fun-Controlnet-Union-2.0",
+        >>> #         filename="Z-Image-Turbo-Fun-Controlnet-Union-2.0.safetensors",
+        >>> #     ),
+        >>> #     torch_dtype=torch.bfloat16,
+        >>> #     config="hlky/Z-Image-Turbo-Fun-Controlnet-Union-2.0",
+        >>> # )
+
+        >>> pipe = ZImageControlNetInpaintPipeline.from_pretrained(
+        ...     "Tongyi-MAI/Z-Image-Turbo", controlnet=controlnet, torch_dtype=torch.bfloat16
+        ... )
+        >>> pipe.to("cuda")
+
+        >>> # Optionally, set the attention backend to flash-attn 2 or 3, default is SDPA in PyTorch.
+        >>> # (1) Use flash attention 2
+        >>> # pipe.transformer.set_attention_backend("flash")
+        >>> # (2) Use flash attention 3
+        >>> # pipe.transformer.set_attention_backend("_flash_3")
+
+        >>> image = load_image(
+        ...     "https://huggingface.co/alibaba-pai/Z-Image-Turbo-Fun-Controlnet-Union-2.0/resolve/main/asset/inpaint.jpg?download=true"
+        ... )
+        >>> mask_image = load_image(
+        ...     "https://huggingface.co/alibaba-pai/Z-Image-Turbo-Fun-Controlnet-Union-2.0/resolve/main/asset/mask.jpg?download=true"
+        ... )
+        >>> control_image = load_image(
+        ...     "https://huggingface.co/alibaba-pai/Z-Image-Turbo-Fun-Controlnet-Union-2.0/resolve/main/asset/pose.jpg?download=true"
+        ... )
+        >>> prompt = "一位年轻女子站在阳光明媚的海岸线上，画面为全身竖构图，身体微微侧向右侧，左手自然下垂，右臂弯曲扶在腰间，她的手指清晰可见，站姿放松而略带羞涩。她身穿轻盈的白色连衣裙，裙摆在海风中轻轻飘动，布料半透、质感柔软。女子拥有一头鲜艳的及腰紫色长发，被海风吹起，在身侧轻盈飞舞，发间系着一个精致的黑色蝴蝶结，与发色形成对比。她面容清秀，眉目精致，肤色白皙细腻，表情温柔略显羞涩，微微低头，眼神静静望向远处的海平线，流露出甜美的青春气息与若有所思的神情。背景是辽阔无垠的海洋与蔚蓝天空，阳光从侧前方洒下，海面波光粼粼，泛着温暖的金色光晕，天空清澈明亮，云朵稀薄，整体色调清新唯美。"
+        >>> image = pipe(
+        ...     prompt,
+        ...     image=image,
+        ...     mask_image=mask_image,
+        ...     control_image=control_image,
+        ...     controlnet_conditioning_scale=0.75,
+        ...     height=1728,
+        ...     width=992,
+        ...     num_inference_steps=25,
+        ...     guidance_scale=0.0,
+        ...     generator=torch.Generator("cuda").manual_seed(43),
+        ... ).images[0]
+        >>> image.save("zimage-inpaint.png")
+        ```
+"""
+
+
+# Copied from diffusers.pipelines.flux.pipeline_flux.calculate_shift
+def calculate_shift(
+    image_seq_len,
+    base_seq_len: int = 256,
+    max_seq_len: int = 4096,
+    base_shift: float = 0.5,
+    max_shift: float = 1.15,
+):
+    m = (max_shift - base_shift) / (max_seq_len - base_seq_len)
+    b = base_shift - m * base_seq_len
+    mu = image_seq_len * m + b
+    return mu
+
+
+# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.retrieve_latents
+def retrieve_latents(
+    encoder_output: torch.Tensor, generator: Optional[torch.Generator] = None, sample_mode: str = "sample"
+):
+    if hasattr(encoder_output, "latent_dist") and sample_mode == "sample":
+        return encoder_output.latent_dist.sample(generator)
+    elif hasattr(encoder_output, "latent_dist") and sample_mode == "argmax":
+        return encoder_output.latent_dist.mode()
+    elif hasattr(encoder_output, "latents"):
+        return encoder_output.latents
+    else:
+        raise AttributeError("Could not access latents of provided encoder_output")
+
+
+# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps
+def retrieve_timesteps(
+    scheduler,
+    num_inference_steps: Optional[int] = None,
+    device: Optional[Union[str, torch.device]] = None,
+    timesteps: Optional[List[int]] = None,
+    sigmas: Optional[List[float]] = None,
+    **kwargs,
+):
+    r"""
+    Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles
+    custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`.
+
+    Args:
+        scheduler (`SchedulerMixin`):
+            The scheduler to get timesteps from.
+        num_inference_steps (`int`):
+            The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps`
+            must be `None`.
+        device (`str` or `torch.device`, *optional*):
+            The device to which the timesteps should be moved to. If `None`, the timesteps are not moved.
+        timesteps (`List[int]`, *optional*):
+            Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed,
+            `num_inference_steps` and `sigmas` must be `None`.
+        sigmas (`List[float]`, *optional*):
+            Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed,
+            `num_inference_steps` and `timesteps` must be `None`.
+
+    Returns:
+        `Tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the
+        second element is the number of inference steps.
+    """
+    if timesteps is not None and sigmas is not None:
+        raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values")
+    if timesteps is not None:
+        accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys())
+        if not accepts_timesteps:
+            raise ValueError(
+                f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom"
+                f" timestep schedules. Please check whether you are using the correct scheduler."
+            )
+        scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs)
+        timesteps = scheduler.timesteps
+        num_inference_steps = len(timesteps)
+    elif sigmas is not None:
+        accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys())
+        if not accept_sigmas:
+            raise ValueError(
+                f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom"
+                f" sigmas schedules. Please check whether you are using the correct scheduler."
+            )
+        scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs)
+        timesteps = scheduler.timesteps
+        num_inference_steps = len(timesteps)
+    else:
+        scheduler.set_timesteps(num_inference_steps, device=device, **kwargs)
+        timesteps = scheduler.timesteps
+    return timesteps, num_inference_steps
+
+
+class ZImageControlNetInpaintPipeline(DiffusionPipeline, FromSingleFileMixin):
+    model_cpu_offload_seq = "text_encoder->transformer->vae"
+    _optional_components = []
+    _callback_tensor_inputs = ["latents", "prompt_embeds"]
+
+    def __init__(
+        self,
+        scheduler: FlowMatchEulerDiscreteScheduler,
+        vae: AutoencoderKL,
+        text_encoder: PreTrainedModel,
+        tokenizer: AutoTokenizer,
+        transformer: ZImageTransformer2DModel,
+        controlnet: ZImageControlNetModel,
+    ):
+        super().__init__()
+        if transformer.in_channels == controlnet.config.control_in_dim:
+            raise ValueError(
+                "ZImageControlNetInpaintPipeline is not compatible with `alibaba-pai/Z-Image-Turbo-Fun-Controlnet-Union`, use `alibaba-pai/Z-Image-Turbo-Fun-Controlnet-Union-2.0`."
+            )
+        controlnet = ZImageControlNetModel.from_transformer(controlnet, transformer)
+
+        self.register_modules(
+            vae=vae,
+            text_encoder=text_encoder,
+            tokenizer=tokenizer,
+            scheduler=scheduler,
+            transformer=transformer,
+            controlnet=controlnet,
+        )
+        self.vae_scale_factor = (
+            2 ** (len(self.vae.config.block_out_channels) - 1) if hasattr(self, "vae") and self.vae is not None else 8
+        )
+        self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor * 2)
+        self.mask_processor = VaeImageProcessor(
+            vae_scale_factor=self.vae_scale_factor, do_normalize=False, do_binarize=True, do_convert_grayscale=True
+        )
+
+    def encode_prompt(
+        self,
+        prompt: Union[str, List[str]],
+        device: Optional[torch.device] = None,
+        do_classifier_free_guidance: bool = True,
+        negative_prompt: Optional[Union[str, List[str]]] = None,
+        prompt_embeds: Optional[List[torch.FloatTensor]] = None,
+        negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+        max_sequence_length: int = 512,
+    ):
+        prompt = [prompt] if isinstance(prompt, str) else prompt
+        prompt_embeds = self._encode_prompt(
+            prompt=prompt,
+            device=device,
+            prompt_embeds=prompt_embeds,
+            max_sequence_length=max_sequence_length,
+        )
+
+        if do_classifier_free_guidance:
+            if negative_prompt is None:
+                negative_prompt = ["" for _ in prompt]
+            else:
+                negative_prompt = [negative_prompt] if isinstance(negative_prompt, str) else negative_prompt
+            assert len(prompt) == len(negative_prompt)
+            negative_prompt_embeds = self._encode_prompt(
+                prompt=negative_prompt,
+                device=device,
+                prompt_embeds=negative_prompt_embeds,
+                max_sequence_length=max_sequence_length,
+            )
+        else:
+            negative_prompt_embeds = []
+        return prompt_embeds, negative_prompt_embeds
+
+    def _encode_prompt(
+        self,
+        prompt: Union[str, List[str]],
+        device: Optional[torch.device] = None,
+        prompt_embeds: Optional[List[torch.FloatTensor]] = None,
+        max_sequence_length: int = 512,
+    ) -> List[torch.FloatTensor]:
+        device = device or self._execution_device
+
+        if prompt_embeds is not None:
+            return prompt_embeds
+
+        if isinstance(prompt, str):
+            prompt = [prompt]
+
+        for i, prompt_item in enumerate(prompt):
+            messages = [
+                {"role": "user", "content": prompt_item},
+            ]
+            prompt_item = self.tokenizer.apply_chat_template(
+                messages,
+                tokenize=False,
+                add_generation_prompt=True,
+                enable_thinking=True,
+            )
+            prompt[i] = prompt_item
+
+        text_inputs = self.tokenizer(
+            prompt,
+            padding="max_length",
+            max_length=max_sequence_length,
+            truncation=True,
+            return_tensors="pt",
+        )
+
+        text_input_ids = text_inputs.input_ids.to(device)
+        prompt_masks = text_inputs.attention_mask.to(device).bool()
+
+        prompt_embeds = self.text_encoder(
+            input_ids=text_input_ids,
+            attention_mask=prompt_masks,
+            output_hidden_states=True,
+        ).hidden_states[-2]
+
+        embeddings_list = []
+
+        for i in range(len(prompt_embeds)):
+            embeddings_list.append(prompt_embeds[i][prompt_masks[i]])
+
+        return embeddings_list
+
+    def prepare_latents(
+        self,
+        batch_size,
+        num_channels_latents,
+        height,
+        width,
+        dtype,
+        device,
+        generator,
+        latents=None,
+    ):
+        height = 2 * (int(height) // (self.vae_scale_factor * 2))
+        width = 2 * (int(width) // (self.vae_scale_factor * 2))
+
+        shape = (batch_size, num_channels_latents, height, width)
+
+        if latents is None:
+            latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+        else:
+            if latents.shape != shape:
+                raise ValueError(f"Unexpected latents shape, got {latents.shape}, expected {shape}")
+            latents = latents.to(device)
+        return latents
+
+    # Copied from diffusers.pipelines.controlnet_sd3.pipeline_stable_diffusion_3_controlnet.StableDiffusion3ControlNetPipeline.prepare_image
+    def prepare_image(
+        self,
+        image,
+        width,
+        height,
+        batch_size,
+        num_images_per_prompt,
+        device,
+        dtype,
+        do_classifier_free_guidance=False,
+        guess_mode=False,
+    ):
+        if isinstance(image, torch.Tensor):
+            pass
+        else:
+            image = self.image_processor.preprocess(image, height=height, width=width)
+
+        image_batch_size = image.shape[0]
+
+        if image_batch_size == 1:
+            repeat_by = batch_size
+        else:
+            # image batch size is the same as prompt batch size
+            repeat_by = num_images_per_prompt
+
+        image = image.repeat_interleave(repeat_by, dim=0)
+
+        image = image.to(device=device, dtype=dtype)
+
+        if do_classifier_free_guidance and not guess_mode:
+            image = torch.cat([image] * 2)
+
+        return image
+
+    @property
+    def guidance_scale(self):
+        return self._guidance_scale
+
+    @property
+    def do_classifier_free_guidance(self):
+        return self._guidance_scale > 1
+
+    @property
+    def joint_attention_kwargs(self):
+        return self._joint_attention_kwargs
+
+    @property
+    def num_timesteps(self):
+        return self._num_timesteps
+
+    @property
+    def interrupt(self):
+        return self._interrupt
+
+    @torch.no_grad()
+    @replace_example_docstring(EXAMPLE_DOC_STRING)
+    def __call__(
+        self,
+        prompt: Union[str, List[str]] = None,
+        height: Optional[int] = None,
+        width: Optional[int] = None,
+        num_inference_steps: int = 50,
+        sigmas: Optional[List[float]] = None,
+        guidance_scale: float = 5.0,
+        image: PipelineImageInput = None,
+        mask_image: PipelineImageInput = None,
+        control_image: PipelineImageInput = None,
+        controlnet_conditioning_scale: Union[float, List[float]] = 0.75,
+        cfg_normalization: bool = False,
+        cfg_truncation: float = 1.0,
+        negative_prompt: Optional[Union[str, List[str]]] = None,
+        num_images_per_prompt: Optional[int] = 1,
+        generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+        latents: Optional[torch.FloatTensor] = None,
+        prompt_embeds: Optional[List[torch.FloatTensor]] = None,
+        negative_prompt_embeds: Optional[List[torch.FloatTensor]] = None,
+        output_type: Optional[str] = "pil",
+        return_dict: bool = True,
+        joint_attention_kwargs: Optional[Dict[str, Any]] = None,
+        callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None,
+        callback_on_step_end_tensor_inputs: List[str] = ["latents"],
+        max_sequence_length: int = 512,
+    ):
+        r"""
+        Function invoked when calling the pipeline for generation.
+
+        Args:
+            prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`.
+                instead.
+            height (`int`, *optional*, defaults to 1024):
+                The height in pixels of the generated image.
+            width (`int`, *optional*, defaults to 1024):
+                The width in pixels of the generated image.
+            num_inference_steps (`int`, *optional*, defaults to 50):
+                The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+                expense of slower inference.
+            sigmas (`List[float]`, *optional*):
+                Custom sigmas to use for the denoising process with schedulers which support a `sigmas` argument in
+                their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is passed
+                will be used.
+            guidance_scale (`float`, *optional*, defaults to 5.0):
+                Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
+                `guidance_scale` is defined as `w` of equation 2. of [Imagen
+                Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
+                1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
+                usually at the expense of lower image quality.
+            cfg_normalization (`bool`, *optional*, defaults to False):
+                Whether to apply configuration normalization.
+            cfg_truncation (`float`, *optional*, defaults to 1.0):
+                The truncation value for configuration.
+            negative_prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts not to guide the image generation. If not defined, one has to pass
+                `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
+                less than `1`).
+            num_images_per_prompt (`int`, *optional*, defaults to 1):
+                The number of images to generate per prompt.
+            generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+                One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html)
+                to make generation deterministic.
+            latents (`torch.FloatTensor`, *optional*):
+                Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image
+                generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+                tensor will be generated by sampling using the supplied random `generator`.
+            prompt_embeds (`List[torch.FloatTensor]`, *optional*):
+                Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+                provided, text embeddings will be generated from `prompt` input argument.
+            negative_prompt_embeds (`List[torch.FloatTensor]`, *optional*):
+                Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+                weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+                argument.
+            output_type (`str`, *optional*, defaults to `"pil"`):
+                The output format of the generate image. Choose between
+                [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`~pipelines.stable_diffusion.ZImagePipelineOutput`] instead of a plain
+                tuple.
+            joint_attention_kwargs (`dict`, *optional*):
+                A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under
+                `self.processor` in
+                [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
+            callback_on_step_end (`Callable`, *optional*):
+                A function that calls at the end of each denoising steps during the inference. The function is called
+                with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int,
+                callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by
+                `callback_on_step_end_tensor_inputs`.
+            callback_on_step_end_tensor_inputs (`List`, *optional*):
+                The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list
+                will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the
+                `._callback_tensor_inputs` attribute of your pipeline class.
+            max_sequence_length (`int`, *optional*, defaults to 512):
+                Maximum sequence length to use with the `prompt`.
+
+        Examples:
+
+        Returns:
+            [`~pipelines.z_image.ZImagePipelineOutput`] or `tuple`: [`~pipelines.z_image.ZImagePipelineOutput`] if
+            `return_dict` is True, otherwise a `tuple`. When returning a tuple, the first element is a list with the
+            generated images.
+        """
+        height = height or 1024
+        width = width or 1024
+
+        vae_scale = self.vae_scale_factor * 2
+        if height % vae_scale != 0:
+            raise ValueError(
+                f"Height must be divisible by {vae_scale} (got {height}). "
+                f"Please adjust the height to a multiple of {vae_scale}."
+            )
+        if width % vae_scale != 0:
+            raise ValueError(
+                f"Width must be divisible by {vae_scale} (got {width}). "
+                f"Please adjust the width to a multiple of {vae_scale}."
+            )
+
+        device = self._execution_device
+
+        self._guidance_scale = guidance_scale
+        self._joint_attention_kwargs = joint_attention_kwargs
+        self._interrupt = False
+        self._cfg_normalization = cfg_normalization
+        self._cfg_truncation = cfg_truncation
+        # 2. Define call parameters
+        if prompt is not None and isinstance(prompt, str):
+            batch_size = 1
+        elif prompt is not None and isinstance(prompt, list):
+            batch_size = len(prompt)
+        else:
+            batch_size = len(prompt_embeds)
+
+        # If prompt_embeds is provided and prompt is None, skip encoding
+        if prompt_embeds is not None and prompt is None:
+            if self.do_classifier_free_guidance and negative_prompt_embeds is None:
+                raise ValueError(
+                    "When `prompt_embeds` is provided without `prompt`, "
+                    "`negative_prompt_embeds` must also be provided for classifier-free guidance."
+                )
+        else:
+            (
+                prompt_embeds,
+                negative_prompt_embeds,
+            ) = self.encode_prompt(
+                prompt=prompt,
+                negative_prompt=negative_prompt,
+                do_classifier_free_guidance=self.do_classifier_free_guidance,
+                prompt_embeds=prompt_embeds,
+                negative_prompt_embeds=negative_prompt_embeds,
+                device=device,
+                max_sequence_length=max_sequence_length,
+            )
+
+        # 4. Prepare latent variables
+        num_channels_latents = self.transformer.in_channels
+
+        control_image = self.prepare_image(
+            image=control_image,
+            width=width,
+            height=height,
+            batch_size=batch_size * num_images_per_prompt,
+            num_images_per_prompt=num_images_per_prompt,
+            device=device,
+            dtype=self.vae.dtype,
+        )
+        height, width = control_image.shape[-2:]
+        control_image = retrieve_latents(self.vae.encode(control_image), generator=generator, sample_mode="argmax")
+        control_image = (control_image - self.vae.config.shift_factor) * self.vae.config.scaling_factor
+        control_image = control_image.unsqueeze(2)
+
+        mask_condition = self.mask_processor.preprocess(mask_image, height=height, width=width)
+        mask_condition = torch.tile(mask_condition, [1, 3, 1, 1]).to(
+            device=control_image.device, dtype=control_image.dtype
+        )
+
+        init_image = self.prepare_image(
+            image=image,
+            width=width,
+            height=height,
+            batch_size=batch_size * num_images_per_prompt,
+            num_images_per_prompt=num_images_per_prompt,
+            device=device,
+            dtype=self.vae.dtype,
+        )
+        height, width = init_image.shape[-2:]
+        init_image = init_image * (mask_condition < 0.5)
+        init_image = retrieve_latents(self.vae.encode(init_image), generator=generator, sample_mode="argmax")
+        init_image = (init_image - self.vae.config.shift_factor) * self.vae.config.scaling_factor
+        init_image = init_image.unsqueeze(2)
+
+        mask_condition = F.interpolate(1 - mask_condition[:, :1], size=init_image.size()[-2:], mode="nearest").to(
+            device=control_image.device, dtype=control_image.dtype
+        )
+        mask_condition = mask_condition.unsqueeze(2)
+
+        control_image = torch.cat([control_image, mask_condition, init_image], dim=1)
+
+        latents = self.prepare_latents(
+            batch_size * num_images_per_prompt,
+            num_channels_latents,
+            height,
+            width,
+            torch.float32,
+            device,
+            generator,
+            latents,
+        )
+
+        # Repeat prompt_embeds for num_images_per_prompt
+        if num_images_per_prompt > 1:
+            prompt_embeds = [pe for pe in prompt_embeds for _ in range(num_images_per_prompt)]
+            if self.do_classifier_free_guidance and negative_prompt_embeds:
+                negative_prompt_embeds = [npe for npe in negative_prompt_embeds for _ in range(num_images_per_prompt)]
+
+        actual_batch_size = batch_size * num_images_per_prompt
+        image_seq_len = (latents.shape[2] // 2) * (latents.shape[3] // 2)
+
+        # 5. Prepare timesteps
+        mu = calculate_shift(
+            image_seq_len,
+            self.scheduler.config.get("base_image_seq_len", 256),
+            self.scheduler.config.get("max_image_seq_len", 4096),
+            self.scheduler.config.get("base_shift", 0.5),
+            self.scheduler.config.get("max_shift", 1.15),
+        )
+        self.scheduler.sigma_min = 0.0
+        scheduler_kwargs = {"mu": mu}
+        timesteps, num_inference_steps = retrieve_timesteps(
+            self.scheduler,
+            num_inference_steps,
+            device,
+            sigmas=sigmas,
+            **scheduler_kwargs,
+        )
+        num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0)
+        self._num_timesteps = len(timesteps)
+
+        # 6. Denoising loop
+        with self.progress_bar(total=num_inference_steps) as progress_bar:
+            for i, t in enumerate(timesteps):
+                if self.interrupt:
+                    continue
+
+                # broadcast to batch dimension in a way that's compatible with ONNX/Core ML
+                timestep = t.expand(latents.shape[0])
+                timestep = (1000 - timestep) / 1000
+                # Normalized time for time-aware config (0 at start, 1 at end)
+                t_norm = timestep[0].item()
+
+                # Handle cfg truncation
+                current_guidance_scale = self.guidance_scale
+                if (
+                    self.do_classifier_free_guidance
+                    and self._cfg_truncation is not None
+                    and float(self._cfg_truncation) <= 1
+                ):
+                    if t_norm > self._cfg_truncation:
+                        current_guidance_scale = 0.0
+
+                # Run CFG only if configured AND scale is non-zero
+                apply_cfg = self.do_classifier_free_guidance and current_guidance_scale > 0
+
+                if apply_cfg:
+                    latents_typed = latents.to(self.transformer.dtype)
+                    latent_model_input = latents_typed.repeat(2, 1, 1, 1)
+                    prompt_embeds_model_input = prompt_embeds + negative_prompt_embeds
+                    timestep_model_input = timestep.repeat(2)
+                else:
+                    latent_model_input = latents.to(self.transformer.dtype)
+                    prompt_embeds_model_input = prompt_embeds
+                    timestep_model_input = timestep
+
+                latent_model_input = latent_model_input.unsqueeze(2)
+                latent_model_input_list = list(latent_model_input.unbind(dim=0))
+
+                controlnet_block_samples = self.controlnet(
+                    latent_model_input_list,
+                    timestep_model_input,
+                    prompt_embeds_model_input,
+                    control_image,
+                    conditioning_scale=controlnet_conditioning_scale,
+                )
+
+                model_out_list = self.transformer(
+                    latent_model_input_list,
+                    timestep_model_input,
+                    prompt_embeds_model_input,
+                    controlnet_block_samples=controlnet_block_samples,
+                )[0]
+
+                if apply_cfg:
+                    # Perform CFG
+                    pos_out = model_out_list[:actual_batch_size]
+                    neg_out = model_out_list[actual_batch_size:]
+
+                    noise_pred = []
+                    for j in range(actual_batch_size):
+                        pos = pos_out[j].float()
+                        neg = neg_out[j].float()
+
+                        pred = pos + current_guidance_scale * (pos - neg)
+
+                        # Renormalization
+                        if self._cfg_normalization and float(self._cfg_normalization) > 0.0:
+                            ori_pos_norm = torch.linalg.vector_norm(pos)
+                            new_pos_norm = torch.linalg.vector_norm(pred)
+                            max_new_norm = ori_pos_norm * float(self._cfg_normalization)
+                            if new_pos_norm > max_new_norm:
+                                pred = pred * (max_new_norm / new_pos_norm)
+
+                        noise_pred.append(pred)
+
+                    noise_pred = torch.stack(noise_pred, dim=0)
+                else:
+                    noise_pred = torch.stack([t.float() for t in model_out_list], dim=0)
+
+                noise_pred = noise_pred.squeeze(2)
+                noise_pred = -noise_pred
+
+                # compute the previous noisy sample x_t -> x_t-1
+                latents = self.scheduler.step(noise_pred.to(torch.float32), t, latents, return_dict=False)[0]
+                assert latents.dtype == torch.float32
+
+                if callback_on_step_end is not None:
+                    callback_kwargs = {}
+                    for k in callback_on_step_end_tensor_inputs:
+                        callback_kwargs[k] = locals()[k]
+                    callback_outputs = callback_on_step_end(self, i, t, callback_kwargs)
+
+                    latents = callback_outputs.pop("latents", latents)
+                    prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds)
+                    negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds)
+
+                # call the callback, if provided
+                if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
+                    progress_bar.update()
+
+        if output_type == "latent":
+            image = latents
+
+        else:
+            latents = latents.to(self.vae.dtype)
+            latents = (latents / self.vae.config.scaling_factor) + self.vae.config.shift_factor
+
+            image = self.vae.decode(latents, return_dict=False)[0]
+            image = self.image_processor.postprocess(image, output_type=output_type)
+
+        # Offload all models
+        self.maybe_free_model_hooks()
+
+        if not return_dict:
+            return (image,)
+
+        return ZImagePipelineOutput(images=image)

src/diffusers/utils/dummy_pt_objects.py

@@ -1777,6 +1777,21 @@ class WanVACETransformer3DModel(metaclass=DummyObject):
         requires_backends(cls, ["torch"])
 
 
+class ZImageControlNetModel(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+    @classmethod
+    def from_config(cls, *args, **kwargs):
+        requires_backends(cls, ["torch"])
+
+    @classmethod
+    def from_pretrained(cls, *args, **kwargs):
+        requires_backends(cls, ["torch"])
+
+
 class ZImageTransformer2DModel(metaclass=DummyObject):
     _backends = ["torch"]
 

src/diffusers/utils/dummy_torch_and_transformers_objects.py

@@ -3857,6 +3857,36 @@ class WuerstchenPriorPipeline(metaclass=DummyObject):
         requires_backends(cls, ["torch", "transformers"])
 
 
+class ZImageControlNetInpaintPipeline(metaclass=DummyObject):
+    _backends = ["torch", "transformers"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch", "transformers"])
+
+    @classmethod
+    def from_config(cls, *args, **kwargs):
+        requires_backends(cls, ["torch", "transformers"])
+
+    @classmethod
+    def from_pretrained(cls, *args, **kwargs):
+        requires_backends(cls, ["torch", "transformers"])
+
+
+class ZImageControlNetPipeline(metaclass=DummyObject):
+    _backends = ["torch", "transformers"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch", "transformers"])
+
+    @classmethod
+    def from_config(cls, *args, **kwargs):
+        requires_backends(cls, ["torch", "transformers"])
+
+    @classmethod
+    def from_pretrained(cls, *args, **kwargs):
+        requires_backends(cls, ["torch", "transformers"])
+
+
 class ZImageImg2ImgPipeline(metaclass=DummyObject):
     _backends = ["torch", "transformers"]