AI/diffusers
1
0
Fork 0
mirror of https://github.com/huggingface/diffusers.git synced 2026-01-29 07:22:12 +03:00
Code Activity

first commit

Browse Source
This commit is contained in:
yiyixuxu
2025-03-17 02:14:55 +01:00
parent 6b9a3334db
commit c952370cb4
8 changed files with 1364 additions and 9 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
src/diffusers/__init__.py
View File

@@ -276,6 +276,7 @@ else:
"UnCLIPScheduler",
"UniPCMultistepScheduler",
"VQDiffusionScheduler",
"SCMScheduler",
]
)
_import_structure["training_utils"] = ["EMAModel"]
@@ -421,6 +422,7 @@ else:
"ReduxImageEncoder",
"SanaPAGPipeline",
"SanaPipeline",
"SanaSCMPipeline",
"SemanticStableDiffusionPipeline",
"ShapEImg2ImgPipeline",
"ShapEPipeline",
@@ -839,6 +841,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
UnCLIPScheduler,
UniPCMultistepScheduler,
VQDiffusionScheduler,
SCMScheduler,
)
from .training_utils import EMAModel
@@ -965,6 +968,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
ReduxImageEncoder,
SanaPAGPipeline,
SanaPipeline,
SanaSCMPipeline,
SemanticStableDiffusionPipeline,
ShapEImg2ImgPipeline,
ShapEPipeline,

5
src/diffusers/models/attention_processor.py
View File

@@ -6020,6 +6020,11 @@ class SanaLinearAttnProcessor2_0:
key = attn.to_k(encoder_hidden_states)
value = attn.to_v(encoder_hidden_states)
if attn.norm_q is not None:
query = attn.norm_q(query)
if attn.norm_k is not None:
key = attn.norm_k(key)
query = query.transpose(1, 2).unflatten(1, (attn.heads, -1))
key = key.transpose(1, 2).unflatten(1, (attn.heads, -1)).transpose(2, 3)
value = value.transpose(1, 2).unflatten(1, (attn.heads, -1))

127
src/diffusers/models/transformers/sana_transformer.py
View File

@@ -30,7 +30,9 @@ from ..embeddings import PatchEmbed, PixArtAlphaTextProjection
from ..modeling_outputs import Transformer2DModelOutput
from ..modeling_utils import ModelMixin
from ..normalization import AdaLayerNormSingle, RMSNorm
from ..embeddings import TimestepEmbedding, Timesteps
import torch.nn.functional as F
logger = logging.get_logger(__name__) # pylint: disable=invalid-name
@@ -96,6 +98,102 @@ class SanaModulatedNorm(nn.Module):
return hidden_states
class SanaCombinedTimestepGuidanceEmbeddings(nn.Module):
"""
For Sana.
Reference:
"""
def __init__(self, embedding_dim):
super().__init__()
self.time_proj = Timesteps(num_channels=256, flip_sin_to_cos=True, downscale_freq_shift=0)
self.timestep_embedder = TimestepEmbedding(in_channels=256, time_embed_dim=embedding_dim)
self.guidance_condition_proj = Timesteps(num_channels=256, flip_sin_to_cos=True, downscale_freq_shift=0)
self.guidance_embedder = TimestepEmbedding(in_channels=256, time_embed_dim=embedding_dim)
self.silu = nn.SiLU()
self.linear = nn.Linear(embedding_dim, 6 * embedding_dim, bias=True)
def forward(self, timestep: torch.Tensor, guidance: torch.Tensor = None, hidden_dtype: torch.dtype = None):
timesteps_proj = self.time_proj(timestep)
timesteps_emb = self.timestep_embedder(timesteps_proj.to(dtype=hidden_dtype)) # (N, D)
guidance_proj = self.guidance_condition_proj(guidance)
guidance_emb = self.guidance_embedder(guidance_proj.to(dtype=hidden_dtype))
conditioning = timesteps_emb + guidance_emb
return self.linear(self.silu(conditioning)), conditioning
class SanaAttnProcessor2_0:
r"""
Processor for implementing scaled dot-product attention (enabled by default if you're using PyTorch 2.0).
"""
def __init__(self):
if not hasattr(F, "scaled_dot_product_attention"):
raise ImportError("SanaAttnProcessor2_0 requires PyTorch 2.0, to use it, please upgrade PyTorch to 2.0.")
def __call__(
self,
attn: Attention,
hidden_states: torch.Tensor,
encoder_hidden_states: Optional[torch.Tensor] = None,
attention_mask: Optional[torch.Tensor] = None,
) -> torch.Tensor:
batch_size, sequence_length, _ = (
hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
)
if attention_mask is not None:
attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
# scaled_dot_product_attention expects attention_mask shape to be
# (batch, heads, source_length, target_length)
attention_mask = attention_mask.view(batch_size, attn.heads, -1, attention_mask.shape[-1])
query = attn.to_q(hidden_states)
if encoder_hidden_states is None:
encoder_hidden_states = hidden_states
key = attn.to_k(encoder_hidden_states)
value = attn.to_v(encoder_hidden_states)
if attn.norm_q is not None:
query = attn.norm_q(query)
if attn.norm_k is not None:
key = attn.norm_k(key)
inner_dim = key.shape[-1]
head_dim = inner_dim // attn.heads
query = query.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
key = key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
value = value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
# the output of sdp = (batch, num_heads, seq_len, head_dim)
# TODO: add support for attn.scale when we move to Torch 2.1
hidden_states = F.scaled_dot_product_attention(
query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False
)
hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
hidden_states = hidden_states.to(query.dtype)
# linear proj
hidden_states = attn.to_out[0](hidden_states)
# dropout
hidden_states = attn.to_out[1](hidden_states)
hidden_states = hidden_states / attn.rescale_output_factor
return hidden_states
class SanaTransformerBlock(nn.Module):
r"""
Transformer block introduced in [Sana](https://huggingface.co/papers/2410.10629).
@@ -115,6 +213,7 @@ class SanaTransformerBlock(nn.Module):
norm_eps: float = 1e-6,
attention_out_bias: bool = True,
mlp_ratio: float = 2.5,
qk_norm: Optional[str] = None,
) -> None:
super().__init__()
@@ -124,6 +223,8 @@ class SanaTransformerBlock(nn.Module):
query_dim=dim,
heads=num_attention_heads,
dim_head=attention_head_dim,
kv_heads=num_attention_heads if qk_norm is not None else None,
qk_norm=qk_norm,
dropout=dropout,
bias=attention_bias,
cross_attention_dim=None,
@@ -135,13 +236,15 @@ class SanaTransformerBlock(nn.Module):
self.norm2 = nn.LayerNorm(dim, elementwise_affine=norm_elementwise_affine, eps=norm_eps)
self.attn2 = Attention(
query_dim=dim,
qk_norm=qk_norm,
kv_heads=num_cross_attention_heads if qk_norm is not None else None,
cross_attention_dim=cross_attention_dim,
heads=num_cross_attention_heads,
dim_head=cross_attention_head_dim,
dropout=dropout,
bias=True,
out_bias=attention_out_bias,
processor=AttnProcessor2_0(),
processor=SanaAttnProcessor2_0(),
)
# 3. Feed-forward
@@ -258,6 +361,8 @@ class SanaTransformer2DModel(ModelMixin, ConfigMixin, PeftAdapterMixin, FromOrig
norm_elementwise_affine: bool = False,
norm_eps: float = 1e-6,
interpolation_scale: Optional[int] = None,
guidance_embeds: bool = False,
qk_norm: Optional[str] = None,
) -> None:
super().__init__()
@@ -276,7 +381,10 @@ class SanaTransformer2DModel(ModelMixin, ConfigMixin, PeftAdapterMixin, FromOrig
)
# 2. Additional condition embeddings
self.time_embed = AdaLayerNormSingle(inner_dim)
if guidance_embeds:
self.time_embed = SanaCombinedTimestepGuidanceEmbeddings(inner_dim)
else:
self.time_embed = AdaLayerNormSingle(inner_dim)
self.caption_projection = PixArtAlphaTextProjection(in_features=caption_channels, hidden_size=inner_dim)
self.caption_norm = RMSNorm(inner_dim, eps=1e-5, elementwise_affine=True)
@@ -296,6 +404,7 @@ class SanaTransformer2DModel(ModelMixin, ConfigMixin, PeftAdapterMixin, FromOrig
norm_elementwise_affine=norm_elementwise_affine,
norm_eps=norm_eps,
mlp_ratio=mlp_ratio,
qk_norm=qk_norm,
)
for _ in range(num_layers)
]
@@ -372,7 +481,8 @@ class SanaTransformer2DModel(ModelMixin, ConfigMixin, PeftAdapterMixin, FromOrig
self,
hidden_states: torch.Tensor,
encoder_hidden_states: torch.Tensor,
timestep: torch.LongTensor,
timestep: torch.Tensor,
guidance: Optional[torch.Tensor] = None,
encoder_attention_mask: Optional[torch.Tensor] = None,
attention_mask: Optional[torch.Tensor] = None,
attention_kwargs: Optional[Dict[str, Any]] = None,
@@ -423,9 +533,14 @@ class SanaTransformer2DModel(ModelMixin, ConfigMixin, PeftAdapterMixin, FromOrig
hidden_states = self.patch_embed(hidden_states)
timestep, embedded_timestep = self.time_embed(
timestep, batch_size=batch_size, hidden_dtype=hidden_states.dtype
)
if guidance is not None:
timestep, embedded_timestep = self.time_embed(
timestep, guidance=guidance, hidden_dtype=hidden_states.dtype
)
else:
timestep, embedded_timestep = self.time_embed(
timestep, batch_size=batch_size, hidden_dtype=hidden_states.dtype
)
encoder_hidden_states = self.caption_projection(encoder_hidden_states)
encoder_hidden_states = encoder_hidden_states.view(batch_size, -1, hidden_states.shape[-1])

4
src/diffusers/pipelines/__init__.py
View File

@@ -280,7 +280,7 @@ else:
_import_structure["paint_by_example"] = ["PaintByExamplePipeline"]
_import_structure["pia"] = ["PIAPipeline"]
_import_structure["pixart_alpha"] = ["PixArtAlphaPipeline", "PixArtSigmaPipeline"]
_import_structure["sana"] = ["SanaPipeline"]
_import_structure["sana"] = ["SanaPipeline", "SanaSCMPipeline"]
_import_structure["semantic_stable_diffusion"] = ["SemanticStableDiffusionPipeline"]
_import_structure["shap_e"] = ["ShapEImg2ImgPipeline", "ShapEPipeline"]
_import_structure["stable_audio"] = [
@@ -651,7 +651,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
from .paint_by_example import PaintByExamplePipeline
from .pia import PIAPipeline
from .pixart_alpha import PixArtAlphaPipeline, PixArtSigmaPipeline
from .sana import SanaPipeline
from .sana import SanaPipeline, SanaSCMPipeline
from .semantic_stable_diffusion import SemanticStableDiffusionPipeline
from .shap_e import ShapEImg2ImgPipeline, ShapEPipeline
from .stable_audio import StableAudioPipeline, StableAudioProjectionModel

2
src/diffusers/pipelines/sana/__init__.py
View File

@@ -23,6 +23,7 @@ except OptionalDependencyNotAvailable:
_dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_objects))
else:
_import_structure["pipeline_sana"] = ["SanaPipeline"]
_import_structure["pipeline_sana_scm"] = ["SanaSCMPipeline"]
if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
try:
@@ -33,6 +34,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
from ...utils.dummy_torch_and_transformers_objects import *
else:
from .pipeline_sana import SanaPipeline
from .pipeline_sana_scm import SanaSCMPipeline
else:
import sys

991
src/diffusers/pipelines/sana/pipeline_sana_scm.py Normal file
View File

@@ -0,0 +1,991 @@
# Copyright 2024 PixArt-Sigma Authors 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 html
import inspect
import re
import urllib.parse as ul
import warnings
from typing import Any, Callable, Dict, List, Optional, Tuple, Union
import torch
from transformers import Gemma2PreTrainedModel, GemmaTokenizer, GemmaTokenizerFast
from ...callbacks import MultiPipelineCallbacks, PipelineCallback
from ...image_processor import PixArtImageProcessor
from ...loaders import SanaLoraLoaderMixin
from ...models import AutoencoderDC, SanaTransformer2DModel
from ...schedulers import DPMSolverMultistepScheduler
from ...utils import (
BACKENDS_MAPPING,
USE_PEFT_BACKEND,
is_bs4_available,
is_ftfy_available,
is_torch_xla_available,
logging,
replace_example_docstring,
scale_lora_layers,
unscale_lora_layers,
)
from ...utils.torch_utils import randn_tensor
from ..pipeline_utils import DiffusionPipeline
from ..pixart_alpha.pipeline_pixart_alpha import (
ASPECT_RATIO_512_BIN,
ASPECT_RATIO_1024_BIN,
)
from ..pixart_alpha.pipeline_pixart_sigma import ASPECT_RATIO_2048_BIN
from .pipeline_output import SanaPipelineOutput
if is_torch_xla_available():
import torch_xla.core.xla_model as xm
XLA_AVAILABLE = True
else:
XLA_AVAILABLE = False
logger = logging.get_logger(__name__) # pylint: disable=invalid-name
if is_bs4_available():
from bs4 import BeautifulSoup
if is_ftfy_available():
import ftfy
ASPECT_RATIO_4096_BIN = {
"0.25": [2048.0, 8192.0],
"0.26": [2048.0, 7936.0],
"0.27": [2048.0, 7680.0],
"0.28": [2048.0, 7424.0],
"0.32": [2304.0, 7168.0],
"0.33": [2304.0, 6912.0],
"0.35": [2304.0, 6656.0],
"0.4": [2560.0, 6400.0],
"0.42": [2560.0, 6144.0],
"0.48": [2816.0, 5888.0],
"0.5": [2816.0, 5632.0],
"0.52": [2816.0, 5376.0],
"0.57": [3072.0, 5376.0],
"0.6": [3072.0, 5120.0],
"0.68": [3328.0, 4864.0],
"0.72": [3328.0, 4608.0],
"0.78": [3584.0, 4608.0],
"0.82": [3584.0, 4352.0],
"0.88": [3840.0, 4352.0],
"0.94": [3840.0, 4096.0],
"1.0": [4096.0, 4096.0],
"1.07": [4096.0, 3840.0],
"1.13": [4352.0, 3840.0],
"1.21": [4352.0, 3584.0],
"1.29": [4608.0, 3584.0],
"1.38": [4608.0, 3328.0],
"1.46": [4864.0, 3328.0],
"1.67": [5120.0, 3072.0],
"1.75": [5376.0, 3072.0],
"2.0": [5632.0, 2816.0],
"2.09": [5888.0, 2816.0],
"2.4": [6144.0, 2560.0],
"2.5": [6400.0, 2560.0],
"2.89": [6656.0, 2304.0],
"3.0": [6912.0, 2304.0],
"3.11": [7168.0, 2304.0],
"3.62": [7424.0, 2048.0],
"3.75": [7680.0, 2048.0],
"3.88": [7936.0, 2048.0],
"4.0": [8192.0, 2048.0],
}
EXAMPLE_DOC_STRING = """
Examples:
```py
>>> import torch
>>> from diffusers import SanaPipeline
>>> pipe = SanaPipeline.from_pretrained(
... "Efficient-Large-Model/Sana_1600M_1024px_BF16_diffusers", torch_dtype=torch.float32
... )
>>> pipe.to("cuda")
>>> pipe.text_encoder.to(torch.bfloat16)
>>> pipe.transformer = pipe.transformer.to(torch.bfloat16)
>>> image = pipe(prompt='a cyberpunk cat with a neon sign that says "Sana"')[0]
>>> image[0].save("output.png")
```
"""
# 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 SanaSCMPipeline(DiffusionPipeline, SanaLoraLoaderMixin):
r"""
Pipeline for text-to-image generation using [Sana](https://huggingface.co/papers/2410.10629).
"""
# fmt: off
bad_punct_regex = re.compile(r"[" + "#®•©™&@·º½¾¿¡§~" + r"\)" + r"\(" + r"\]" + r"\[" + r"\}" + r"\{" + r"\|" + "\\" + r"\/" + r"\*" + r"]{1,}")
# fmt: on
model_cpu_offload_seq = "text_encoder->transformer->vae"
_callback_tensor_inputs = ["latents", "prompt_embeds", "negative_prompt_embeds"]
def __init__(
self,
tokenizer: Union[GemmaTokenizer, GemmaTokenizerFast],
text_encoder: Gemma2PreTrainedModel,
vae: AutoencoderDC,
transformer: SanaTransformer2DModel,
scheduler: DPMSolverMultistepScheduler,
):
super().__init__()
self.register_modules(
tokenizer=tokenizer, text_encoder=text_encoder, vae=vae, transformer=transformer, scheduler=scheduler
)
self.vae_scale_factor = (
2 ** (len(self.vae.config.encoder_block_out_channels) - 1)
if hasattr(self, "vae") and self.vae is not None
else 32
)
self.image_processor = PixArtImageProcessor(vae_scale_factor=self.vae_scale_factor)
def enable_vae_slicing(self):
r"""
Enable sliced VAE decoding. When this option is enabled, the VAE will split the input tensor in slices to
compute decoding in several steps. This is useful to save some memory and allow larger batch sizes.
"""
self.vae.enable_slicing()
def disable_vae_slicing(self):
r"""
Disable sliced VAE decoding. If `enable_vae_slicing` was previously enabled, this method will go back to
computing decoding in one step.
"""
self.vae.disable_slicing()
def enable_vae_tiling(self):
r"""
Enable tiled VAE decoding. When this option is enabled, the VAE will split the input tensor into tiles to
compute decoding and encoding in several steps. This is useful for saving a large amount of memory and to allow
processing larger images.
"""
self.vae.enable_tiling()
def disable_vae_tiling(self):
r"""
Disable tiled VAE decoding. If `enable_vae_tiling` was previously enabled, this method will go back to
computing decoding in one step.
"""
self.vae.disable_tiling()
def encode_prompt(
self,
prompt: Union[str, List[str]],
do_classifier_free_guidance: bool = True,
negative_prompt: str = "",
num_images_per_prompt: int = 1,
device: Optional[torch.device] = None,
prompt_embeds: Optional[torch.Tensor] = None,
negative_prompt_embeds: Optional[torch.Tensor] = None,
prompt_attention_mask: Optional[torch.Tensor] = None,
negative_prompt_attention_mask: Optional[torch.Tensor] = None,
clean_caption: bool = False,
max_sequence_length: int = 300,
complex_human_instruction: Optional[List[str]] = None,
lora_scale: Optional[float] = None,
):
r"""
Encodes the prompt into text encoder hidden states.
Args:
prompt (`str` or `List[str]`, *optional*):
prompt to be encoded
negative_prompt (`str` or `List[str]`, *optional*):
The prompt 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`). For
PixArt-Alpha, this should be "".
do_classifier_free_guidance (`bool`, *optional*, defaults to `True`):
whether to use classifier free guidance or not
num_images_per_prompt (`int`, *optional*, defaults to 1):
number of images that should be generated per prompt
device: (`torch.device`, *optional*):
torch device to place the resulting embeddings on
prompt_embeds (`torch.Tensor`, *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 (`torch.Tensor`, *optional*):
Pre-generated negative text embeddings. For Sana, it's should be the embeddings of the "" string.
clean_caption (`bool`, defaults to `False`):
If `True`, the function will preprocess and clean the provided caption before encoding.
max_sequence_length (`int`, defaults to 300): Maximum sequence length to use for the prompt.
complex_human_instruction (`list[str]`, defaults to `complex_human_instruction`):
If `complex_human_instruction` is not empty, the function will use the complex Human instruction for
the prompt.
"""
if device is None:
device = self._execution_device
# set lora scale so that monkey patched LoRA
# function of text encoder can correctly access it
if lora_scale is not None and isinstance(self, SanaLoraLoaderMixin):
self._lora_scale = lora_scale
# dynamically adjust the LoRA scale
if self.text_encoder is not None and USE_PEFT_BACKEND:
scale_lora_layers(self.text_encoder, lora_scale)
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 = prompt_embeds.shape[0]
if getattr(self, "tokenizer", None) is not None:
self.tokenizer.padding_side = "right"
# See Section 3.1. of the paper.
max_length = max_sequence_length
select_index = [0] + list(range(-max_length + 1, 0))
if prompt_embeds is None:
prompt = self._text_preprocessing(prompt, clean_caption=clean_caption)
# prepare complex human instruction
if not complex_human_instruction:
max_length_all = max_length
else:
chi_prompt = "\n".join(complex_human_instruction)
prompt = [chi_prompt + p for p in prompt]
num_chi_prompt_tokens = len(self.tokenizer.encode(chi_prompt))
max_length_all = num_chi_prompt_tokens + max_length - 2
text_inputs = self.tokenizer(
prompt,
padding="max_length",
max_length=max_length_all,
truncation=True,
add_special_tokens=True,
return_tensors="pt",
)
text_input_ids = text_inputs.input_ids
prompt_attention_mask = text_inputs.attention_mask
prompt_attention_mask = prompt_attention_mask.to(device)
prompt_embeds = self.text_encoder(text_input_ids.to(device), attention_mask=prompt_attention_mask)
prompt_embeds = prompt_embeds[0][:, select_index]
prompt_attention_mask = prompt_attention_mask[:, select_index]
if self.transformer is not None:
dtype = self.transformer.dtype
elif self.text_encoder is not None:
dtype = self.text_encoder.dtype
else:
dtype = None
prompt_embeds = prompt_embeds.to(dtype=dtype, device=device)
bs_embed, seq_len, _ = prompt_embeds.shape
# duplicate text embeddings and attention mask for each generation per prompt, using mps friendly method
prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1)
prompt_attention_mask = prompt_attention_mask.view(bs_embed, -1)
prompt_attention_mask = prompt_attention_mask.repeat(num_images_per_prompt, 1)
# get unconditional embeddings for classifier free guidance
if do_classifier_free_guidance and negative_prompt_embeds is None:
uncond_tokens = [negative_prompt] * batch_size if isinstance(negative_prompt, str) else negative_prompt
uncond_tokens = self._text_preprocessing(uncond_tokens, clean_caption=clean_caption)
max_length = prompt_embeds.shape[1]
uncond_input = self.tokenizer(
uncond_tokens,
padding="max_length",
max_length=max_length,
truncation=True,
return_attention_mask=True,
add_special_tokens=True,
return_tensors="pt",
)
negative_prompt_attention_mask = uncond_input.attention_mask
negative_prompt_attention_mask = negative_prompt_attention_mask.to(device)
negative_prompt_embeds = self.text_encoder(
uncond_input.input_ids.to(device), attention_mask=negative_prompt_attention_mask
)
negative_prompt_embeds = negative_prompt_embeds[0]
if do_classifier_free_guidance:
# duplicate unconditional embeddings for each generation per prompt, using mps friendly method
seq_len = negative_prompt_embeds.shape[1]
negative_prompt_embeds = negative_prompt_embeds.to(dtype=dtype, device=device)
negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1)
negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1)
negative_prompt_attention_mask = negative_prompt_attention_mask.view(bs_embed, -1)
negative_prompt_attention_mask = negative_prompt_attention_mask.repeat(num_images_per_prompt, 1)
else:
negative_prompt_embeds = None
negative_prompt_attention_mask = None
if self.text_encoder is not None:
if isinstance(self, SanaLoraLoaderMixin) and USE_PEFT_BACKEND:
# Retrieve the original scale by scaling back the LoRA layers
unscale_lora_layers(self.text_encoder, lora_scale)
return prompt_embeds, prompt_attention_mask, negative_prompt_embeds, negative_prompt_attention_mask
# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs
def prepare_extra_step_kwargs(self, generator, eta):
# prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
# eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
# eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
# and should be between [0, 1]
accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
extra_step_kwargs = {}
if accepts_eta:
extra_step_kwargs["eta"] = eta
# check if the scheduler accepts generator
accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys())
if accepts_generator:
extra_step_kwargs["generator"] = generator
return extra_step_kwargs
def check_inputs(
self,
prompt,
height,
width,
callback_on_step_end_tensor_inputs=None,
negative_prompt=None,
prompt_embeds=None,
negative_prompt_embeds=None,
prompt_attention_mask=None,
negative_prompt_attention_mask=None,
):
if height % 32 != 0 or width % 32 != 0:
raise ValueError(f"`height` and `width` have to be divisible by 32 but are {height} and {width}.")
if callback_on_step_end_tensor_inputs is not None and not all(
k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs
):
raise ValueError(
f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}"
)
if prompt is not None and prompt_embeds is not None:
raise ValueError(
f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
" only forward one of the two."
)
elif prompt is None and prompt_embeds is None:
raise ValueError(
"Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
)
elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)):
raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
if prompt is not None and negative_prompt_embeds is not None:
raise ValueError(
f"Cannot forward both `prompt`: {prompt} and `negative_prompt_embeds`:"
f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
)
if negative_prompt is not None and negative_prompt_embeds is not None:
raise ValueError(
f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:"
f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
)
if prompt_embeds is not None and prompt_attention_mask is None:
raise ValueError("Must provide `prompt_attention_mask` when specifying `prompt_embeds`.")
if negative_prompt_embeds is not None and negative_prompt_attention_mask is None:
raise ValueError("Must provide `negative_prompt_attention_mask` when specifying `negative_prompt_embeds`.")
if prompt_embeds is not None and negative_prompt_embeds is not None:
if prompt_embeds.shape != negative_prompt_embeds.shape:
raise ValueError(
"`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but"
f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`"
f" {negative_prompt_embeds.shape}."
)
if prompt_attention_mask.shape != negative_prompt_attention_mask.shape:
raise ValueError(
"`prompt_attention_mask` and `negative_prompt_attention_mask` must have the same shape when passed directly, but"
f" got: `prompt_attention_mask` {prompt_attention_mask.shape} != `negative_prompt_attention_mask`"
f" {negative_prompt_attention_mask.shape}."
)
# Copied from diffusers.pipelines.deepfloyd_if.pipeline_if.IFPipeline._text_preprocessing
def _text_preprocessing(self, text, clean_caption=False):
if clean_caption and not is_bs4_available():
logger.warning(BACKENDS_MAPPING["bs4"][-1].format("Setting `clean_caption=True`"))
logger.warning("Setting `clean_caption` to False...")
clean_caption = False
if clean_caption and not is_ftfy_available():
logger.warning(BACKENDS_MAPPING["ftfy"][-1].format("Setting `clean_caption=True`"))
logger.warning("Setting `clean_caption` to False...")
clean_caption = False
if not isinstance(text, (tuple, list)):
text = [text]
def process(text: str):
if clean_caption:
text = self._clean_caption(text)
text = self._clean_caption(text)
else:
text = text.lower().strip()
return text
return [process(t) for t in text]
# Copied from diffusers.pipelines.deepfloyd_if.pipeline_if.IFPipeline._clean_caption
def _clean_caption(self, caption):
caption = str(caption)
caption = ul.unquote_plus(caption)
caption = caption.strip().lower()
caption = re.sub("<person>", "person", caption)
# urls:
caption = re.sub(
r"\b((?:https?:(?:\/{1,3}|[a-zA-Z0-9%])|[a-zA-Z0-9.\-]+[.](?:com|co|ru|net|org|edu|gov|it)[\w/-]*\b\/?(?!@)))", # noqa
"",
caption,
) # regex for urls
caption = re.sub(
r"\b((?:www:(?:\/{1,3}|[a-zA-Z0-9%])|[a-zA-Z0-9.\-]+[.](?:com|co|ru|net|org|edu|gov|it)[\w/-]*\b\/?(?!@)))", # noqa
"",
caption,
) # regex for urls
# html:
caption = BeautifulSoup(caption, features="html.parser").text
# @<nickname>
caption = re.sub(r"@[\w\d]+\b", "", caption)
# 31C0—31EF CJK Strokes
# 31F0—31FF Katakana Phonetic Extensions
# 3200—32FF Enclosed CJK Letters and Months
# 3300—33FF CJK Compatibility
# 3400—4DBF CJK Unified Ideographs Extension A
# 4DC0—4DFF Yijing Hexagram Symbols
# 4E00—9FFF CJK Unified Ideographs
caption = re.sub(r"[\u31c0-\u31ef]+", "", caption)
caption = re.sub(r"[\u31f0-\u31ff]+", "", caption)
caption = re.sub(r"[\u3200-\u32ff]+", "", caption)
caption = re.sub(r"[\u3300-\u33ff]+", "", caption)
caption = re.sub(r"[\u3400-\u4dbf]+", "", caption)
caption = re.sub(r"[\u4dc0-\u4dff]+", "", caption)
caption = re.sub(r"[\u4e00-\u9fff]+", "", caption)
#######################################################
# все виды тире / all types of dash --> "-"
caption = re.sub(
r"[\u002D\u058A\u05BE\u1400\u1806\u2010-\u2015\u2E17\u2E1A\u2E3A\u2E3B\u2E40\u301C\u3030\u30A0\uFE31\uFE32\uFE58\uFE63\uFF0D]+", # noqa
"-",
caption,
)
# кавычки к одному стандарту
caption = re.sub(r"[`´«»“”¨]", '"', caption)
caption = re.sub(r"[‘’]", "'", caption)
# &quot;
caption = re.sub(r"&quot;?", "", caption)
# &amp
caption = re.sub(r"&amp", "", caption)
# ip adresses:
caption = re.sub(r"\d{1,3}\.\d{1,3}\.\d{1,3}\.\d{1,3}", " ", caption)
# article ids:
caption = re.sub(r"\d:\d\d\s+$", "", caption)
# \n
caption = re.sub(r"\\n", " ", caption)
# "#123"
caption = re.sub(r"#\d{1,3}\b", "", caption)
# "#12345.."
caption = re.sub(r"#\d{5,}\b", "", caption)
# "123456.."
caption = re.sub(r"\b\d{6,}\b", "", caption)
# filenames:
caption = re.sub(r"[\S]+\.(?:png|jpg|jpeg|bmp|webp|eps|pdf|apk|mp4)", "", caption)
#
caption = re.sub(r"[\"\']{2,}", r'"', caption) # """AUSVERKAUFT"""
caption = re.sub(r"[\.]{2,}", r" ", caption) # """AUSVERKAUFT"""
caption = re.sub(self.bad_punct_regex, r" ", caption) # ***AUSVERKAUFT***, #AUSVERKAUFT
caption = re.sub(r"\s+\.\s+", r" ", caption) # " . "
# this-is-my-cute-cat / this_is_my_cute_cat
regex2 = re.compile(r"(?:\-|\_)")
if len(re.findall(regex2, caption)) > 3:
caption = re.sub(regex2, " ", caption)
caption = ftfy.fix_text(caption)
caption = html.unescape(html.unescape(caption))
caption = re.sub(r"\b[a-zA-Z]{1,3}\d{3,15}\b", "", caption) # jc6640
caption = re.sub(r"\b[a-zA-Z]+\d+[a-zA-Z]+\b", "", caption) # jc6640vc
caption = re.sub(r"\b\d+[a-zA-Z]+\d+\b", "", caption) # 6640vc231
caption = re.sub(r"(worldwide\s+)?(free\s+)?shipping", "", caption)
caption = re.sub(r"(free\s)?download(\sfree)?", "", caption)
caption = re.sub(r"\bclick\b\s(?:for|on)\s\w+", "", caption)
caption = re.sub(r"\b(?:png|jpg|jpeg|bmp|webp|eps|pdf|apk|mp4)(\simage[s]?)?", "", caption)
caption = re.sub(r"\bpage\s+\d+\b", "", caption)
caption = re.sub(r"\b\d*[a-zA-Z]+\d+[a-zA-Z]+\d+[a-zA-Z\d]*\b", r" ", caption) # j2d1a2a...
caption = re.sub(r"\b\d+\.?\d*[xх×]\d+\.?\d*\b", "", caption)
caption = re.sub(r"\b\s+\:\s+", r": ", caption)
caption = re.sub(r"(\D[,\./])\b", r"\1 ", caption)
caption = re.sub(r"\s+", " ", caption)
caption.strip()
caption = re.sub(r"^[\"\']([\w\W]+)[\"\']$", r"\1", caption)
caption = re.sub(r"^[\'\_,\-\:;]", r"", caption)
caption = re.sub(r"[\'\_,\-\:\-\+]$", r"", caption)
caption = re.sub(r"^\.\S+$", "", caption)
return caption.strip()
def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None):
if latents is not None:
return latents.to(device=device, dtype=dtype)
shape = (
batch_size,
num_channels_latents,
int(height) // self.vae_scale_factor,
int(width) // self.vae_scale_factor,
)
if isinstance(generator, list) and len(generator) != batch_size:
raise ValueError(
f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
f" size of {batch_size}. Make sure the batch size matches the length of the generators."
)
latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
return latents
@property
def guidance_scale(self):
return self._guidance_scale
@property
def attention_kwargs(self):
return self._attention_kwargs
@property
def do_classifier_free_guidance(self):
return self._guidance_scale > 1.0
@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,
negative_prompt: str = "",
num_inference_steps: int = 20,
timesteps: List[int] = None,
sigmas: List[float] = None,
guidance_scale: float = 4.5,
num_images_per_prompt: Optional[int] = 1,
height: int = 1024,
width: int = 1024,
eta: float = 0.0,
generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
latents: Optional[torch.Tensor] = None,
prompt_embeds: Optional[torch.Tensor] = None,
prompt_attention_mask: Optional[torch.Tensor] = None,
negative_prompt_embeds: Optional[torch.Tensor] = None,
negative_prompt_attention_mask: Optional[torch.Tensor] = None,
output_type: Optional[str] = "pil",
return_dict: bool = True,
clean_caption: bool = False,
use_resolution_binning: bool = True,
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 = 300,
complex_human_instruction: List[str] = [
"Given a user prompt, generate an 'Enhanced prompt' that provides detailed visual descriptions suitable for image generation. Evaluate the level of detail in the user prompt:",
"- If the prompt is simple, focus on adding specifics about colors, shapes, sizes, textures, and spatial relationships to create vivid and concrete scenes.",
"- If the prompt is already detailed, refine and enhance the existing details slightly without overcomplicating.",
"Here are examples of how to transform or refine prompts:",
"- User Prompt: A cat sleeping -> Enhanced: A small, fluffy white cat curled up in a round shape, sleeping peacefully on a warm sunny windowsill, surrounded by pots of blooming red flowers.",
"- User Prompt: A busy city street -> Enhanced: A bustling city street scene at dusk, featuring glowing street lamps, a diverse crowd of people in colorful clothing, and a double-decker bus passing by towering glass skyscrapers.",
"Please generate only the enhanced description for the prompt below and avoid including any additional commentary or evaluations:",
"User Prompt: ",
],
) -> Union[SanaPipelineOutput, Tuple]:
"""
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.
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_inference_steps (`int`, *optional*, defaults to 20):
The number of denoising steps. More denoising steps usually lead to a higher quality image at the
expense of slower inference.
timesteps (`List[int]`, *optional*):
Custom timesteps to use for the denoising process with schedulers which support a `timesteps` argument
in their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is
passed will be used. Must be in descending order.
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 4.5):
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.
num_images_per_prompt (`int`, *optional*, defaults to 1):
The number of images to generate per prompt.
height (`int`, *optional*, defaults to self.unet.config.sample_size):
The height in pixels of the generated image.
width (`int`, *optional*, defaults to self.unet.config.sample_size):
The width in pixels of the generated image.
eta (`float`, *optional*, defaults to 0.0):
Corresponds to parameter eta (η) in the DDIM paper: https://arxiv.org/abs/2010.02502. Only applies to
[`schedulers.DDIMScheduler`], will be ignored for others.
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.Tensor`, *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 ge generated by sampling using the supplied random `generator`.
prompt_embeds (`torch.Tensor`, *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.
prompt_attention_mask (`torch.Tensor`, *optional*): Pre-generated attention mask for text embeddings.
negative_prompt_embeds (`torch.Tensor`, *optional*):
Pre-generated negative text embeddings. For PixArt-Sigma this negative prompt should be "". If not
provided, negative_prompt_embeds will be generated from `negative_prompt` input argument.
negative_prompt_attention_mask (`torch.Tensor`, *optional*):
Pre-generated attention mask for negative text embeddings.
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.IFPipelineOutput`] instead of a plain tuple.
attention_kwargs:
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).
clean_caption (`bool`, *optional*, defaults to `True`):
Whether or not to clean the caption before creating embeddings. Requires `beautifulsoup4` and `ftfy` to
be installed. If the dependencies are not installed, the embeddings will be created from the raw
prompt.
use_resolution_binning (`bool` defaults to `True`):
If set to `True`, the requested height and width are first mapped to the closest resolutions using
`ASPECT_RATIO_1024_BIN`. After the produced latents are decoded into images, they are resized back to
the requested resolution. Useful for generating non-square images.
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` defaults to `300`):
Maximum sequence length to use with the `prompt`.
complex_human_instruction (`List[str]`, *optional*):
Instructions for complex human attention:
https://github.com/NVlabs/Sana/blob/main/configs/sana_app_config/Sana_1600M_app.yaml#L55.
Examples:
Returns:
[`~pipelines.sana.pipeline_output.SanaPipelineOutput`] or `tuple`:
If `return_dict` is `True`, [`~pipelines.sana.pipeline_output.SanaPipelineOutput`] is returned,
otherwise a `tuple` is returned where the first element is a list with the generated images
"""
if isinstance(callback_on_step_end, (PipelineCallback, MultiPipelineCallbacks)):
callback_on_step_end_tensor_inputs = callback_on_step_end.tensor_inputs
# 1. Check inputs. Raise error if not correct
if use_resolution_binning:
if self.transformer.config.sample_size == 128:
aspect_ratio_bin = ASPECT_RATIO_4096_BIN
elif self.transformer.config.sample_size == 64:
aspect_ratio_bin = ASPECT_RATIO_2048_BIN
elif self.transformer.config.sample_size == 32:
aspect_ratio_bin = ASPECT_RATIO_1024_BIN
elif self.transformer.config.sample_size == 16:
aspect_ratio_bin = ASPECT_RATIO_512_BIN
else:
raise ValueError("Invalid sample size")
orig_height, orig_width = height, width
height, width = self.image_processor.classify_height_width_bin(height, width, ratios=aspect_ratio_bin)
self.check_inputs(
prompt,
height,
width,
callback_on_step_end_tensor_inputs,
negative_prompt,
prompt_embeds,
negative_prompt_embeds,
prompt_attention_mask,
negative_prompt_attention_mask,
)
self._guidance_scale = guidance_scale
self._attention_kwargs = attention_kwargs
self._interrupt = False
# 2. Default height and width to transformer
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 = prompt_embeds.shape[0]
device = self._execution_device
lora_scale = self.attention_kwargs.get("scale", None) if self.attention_kwargs is not None else None
# 3. Encode input prompt
(
prompt_embeds,
prompt_attention_mask,
_,
_,
) = self.encode_prompt(
prompt,
False,
negative_prompt=negative_prompt,
num_images_per_prompt=num_images_per_prompt,
device=device,
prompt_embeds=prompt_embeds,
negative_prompt_embeds=negative_prompt_embeds,
prompt_attention_mask=prompt_attention_mask,
negative_prompt_attention_mask=negative_prompt_attention_mask,
clean_caption=clean_caption,
max_sequence_length=max_sequence_length,
complex_human_instruction=complex_human_instruction,
lora_scale=lora_scale,
)
# prompt_embeds = torch.load("/raid/yiyi/Sana-Sprint-diffusers/y.pt").to(device, dtype=prompt_embeds.dtype)
# 4. Prepare timesteps
timesteps, num_inference_steps = retrieve_timesteps(
self.scheduler, num_inference_steps, device, timesteps, sigmas
)
# 5. Prepare latents.
latent_channels = self.transformer.config.in_channels
latents = self.prepare_latents(
batch_size * num_images_per_prompt,
latent_channels,
height,
width,
torch.float32,
device,
generator,
latents,
)
# latents = torch.load("/raid/yiyi/Sana-Sprint-diffusers/latents.pt").to(device, dtype=latents.dtype)
latents = latents * self.scheduler.config.sigma_data
guidance = torch.full([1], guidance_scale, device=device, dtype=torch.float32)
guidance = guidance.expand(latents.shape[0]).to(prompt_embeds.dtype)
# YiYi TODO: cfg_embed_scale = 0.1 (refactor this out)
guidance = guidance * 0.1
# 6. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
# YiYi TODO: refactor this
timesteps = timesteps[:-1]
# 7. Denoising loop
num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0)
self._num_timesteps = len(timesteps)
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]).to(prompt_embeds.dtype)
# YiYi TODO: self.scheduler.scale_model_input?
latents_model_input = latents / self.scheduler.config.sigma_data
# YiYi TODO: refator this out
scm_timestep = torch.sin(timestep) / (torch.cos(timestep) + torch.sin(timestep))
latent_model_input = latents_model_input * torch.sqrt(scm_timestep**2 + (1 - scm_timestep) ** 2)
latent_model_input = latent_model_input.to(prompt_embeds.dtype)
# predict noise model_output
noise_pred = self.transformer(
latent_model_input,
encoder_hidden_states=prompt_embeds,
encoder_attention_mask=prompt_attention_mask,
guidance=guidance,
timestep=scm_timestep,
return_dict=False,
attention_kwargs=self.attention_kwargs,
)[0]
# YiYi TODO: refator this out
noise_pred = ((1 - 2 * scm_timestep) * latent_model_input + (1 - 2 * scm_timestep + 2 * scm_timestep**2) * noise_pred) / torch.sqrt(
scm_timestep**2 + (1 - scm_timestep) ** 2
)
# YiYi TODO: check if this can be refatored into scheduler
noise_pred = noise_pred.float() * self.scheduler.config.sigma_data
# compute previous image: x_t -> x_t-1
latents, denoised = self.scheduler.step(noise_pred, i, timestep, latents, **extra_step_kwargs, return_dict=False)
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 XLA_AVAILABLE:
xm.mark_step()
# YiYi TODO: refator this out
latents = denoised / self.scheduler.config.sigma_data
if output_type == "latent":
image = latents
else:
latents = latents.to(self.vae.dtype)
try:
image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0]
except torch.cuda.OutOfMemoryError as e:
warnings.warn(
f"{e}. \n"
f"Try to use VAE tiling for large images. For example: \n"
f"pipe.vae.enable_tiling(tile_sample_min_width=512, tile_sample_min_height=512)"
)
if use_resolution_binning:
image = self.image_processor.resize_and_crop_tensor(image, orig_width, orig_height)
if not output_type == "latent":
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 SanaPipelineOutput(images=image)

3
src/diffusers/schedulers/__init__.py
View File

@@ -74,6 +74,7 @@ else:
_import_structure["scheduling_unipc_multistep"] = ["UniPCMultistepScheduler"]
_import_structure["scheduling_utils"] = ["AysSchedules", "KarrasDiffusionSchedulers", "SchedulerMixin"]
_import_structure["scheduling_vq_diffusion"] = ["VQDiffusionScheduler"]
_import_structure["scheduling_scm"] = ["SCMScheduler"]
try:
if not is_flax_available():
@@ -174,7 +175,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
from .scheduling_unipc_multistep import UniPCMultistepScheduler
from .scheduling_utils import AysSchedules, KarrasDiffusionSchedulers, SchedulerMixin
from .scheduling_vq_diffusion import VQDiffusionScheduler
from .scheduling_scm import SCMScheduler
try:
if not is_flax_available():
raise OptionalDependencyNotAvailable()

237
src/diffusers/schedulers/scheduling_scm.py Normal file
View File

@@ -0,0 +1,237 @@
# # Copyright 2024 Sana-Sprint Authors 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.
# DISCLAIMER: This code is strongly influenced by https://github.com/pesser/pytorch_diffusion
# and https://github.com/hojonathanho/diffusion
import math
from dataclasses import dataclass
from typing import List, Optional, Tuple, Union
import numpy as np
import torch
from ..configuration_utils import ConfigMixin, register_to_config
from ..schedulers.scheduling_utils import SchedulerMixin
from ..utils import BaseOutput, logging
from ..utils.torch_utils import randn_tensor
logger = logging.get_logger(__name__) # pylint: disable=invalid-name
@dataclass
# Copied from diffusers.schedulers.scheduling_ddpm.DDPMSchedulerOutput with DDPM->DDIM
class SCMSchedulerOutput(BaseOutput):
"""
Output class for the scheduler's `step` function output.
Args:
prev_sample (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)` for images):
Computed sample `(x_{t-1})` of previous timestep. `prev_sample` should be used as next model input in the
denoising loop.
pred_original_sample (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)` for images):
The predicted denoised sample `(x_{0})` based on the model output from the current timestep.
`pred_original_sample` can be used to preview progress or for guidance.
"""
prev_sample: torch.FloatTensor
denoised: Optional[torch.FloatTensor] = None
class SCMScheduler(SchedulerMixin, ConfigMixin):
"""
`SCMScheduler` extends the denoising procedure introduced in denoising diffusion probabilistic models (DDPMs) with
non-Markovian guidance.
This model inherits from [`SchedulerMixin`] and [`ConfigMixin`]. Check the superclass documentation for the generic
methods the library implements for all schedulers such as loading and saving.
Args:
num_train_timesteps (`int`, defaults to 1000):
The number of diffusion steps to train the model.
beta_start (`float`, defaults to 0.0001):
The starting `beta` value of inference.
beta_end (`float`, defaults to 0.02):
The final `beta` value.
beta_schedule (`str`, defaults to `"linear"`):
The beta schedule, a mapping from a beta range to a sequence of betas for stepping the model. Choose from
`linear`, `scaled_linear`, or `squaredcos_cap_v2`.
trained_betas (`np.ndarray`, *optional*):
Pass an array of betas directly to the constructor to bypass `beta_start` and `beta_end`.
clip_sample (`bool`, defaults to `True`):
Clip the predicted sample for numerical stability.
clip_sample_range (`float`, defaults to 1.0):
The maximum magnitude for sample clipping. Valid only when `clip_sample=True`.
set_alpha_to_one (`bool`, defaults to `True`):
Each diffusion step uses the alphas product value at that step and at the previous one. For the final step
there is no previous alpha. When this option is `True` the previous alpha product is fixed to `1`,
otherwise it uses the alpha value at step 0.
steps_offset (`int`, defaults to 0):
An offset added to the inference steps. You can use a combination of `offset=1` and
`set_alpha_to_one=False` to make the last step use step 0 for the previous alpha product like in Stable
Diffusion.
prediction_type (`str`, defaults to `epsilon`, *optional*):
Prediction type of the scheduler function; can be `epsilon` (predicts the noise of the diffusion process),
`sample` (directly predicts the noisy sample`) or `v_prediction` (see section 2.4 of [Imagen
Video](https://imagen.research.google/video/paper.pdf) paper).
thresholding (`bool`, defaults to `False`):
Whether to use the "dynamic thresholding" method. This is unsuitable for latent-space diffusion models such
as Stable Diffusion.
dynamic_thresholding_ratio (`float`, defaults to 0.995):
The ratio for the dynamic thresholding method. Valid only when `thresholding=True`.
sample_max_value (`float`, defaults to 1.0):
The threshold value for dynamic thresholding. Valid only when `thresholding=True`.
timestep_spacing (`str`, defaults to `"leading"`):
The way the timesteps should be scaled. Refer to Table 2 of the [Common Diffusion Noise Schedules and
Sample Steps are Flawed](https://huggingface.co/papers/2305.08891) for more information.
rescale_betas_zero_snr (`bool`, defaults to `False`):
Whether to rescale the betas to have zero terminal SNR. This enables the model to generate very bright and
dark samples instead of limiting it to samples with medium brightness. Loosely related to
[`--offset_noise`](https://github.com/huggingface/diffusers/blob/74fd735eb073eb1d774b1ab4154a0876eb82f055/examples/dreambooth/train_dreambooth.py#L506).
"""
# _compatibles = [e.name for e in KarrasDiffusionSchedulers]
order = 1
@register_to_config
def __init__(
self,
num_train_timesteps: int = 1000,
beta_start: float = 0.0001,
beta_end: float = 0.02,
beta_schedule: str = "linear",
trained_betas: Optional[Union[np.ndarray, List[float]]] = None,
clip_sample: bool = True,
set_alpha_to_one: bool = True,
steps_offset: int = 0,
prediction_type: str = "trigflow",
thresholding: bool = False,
dynamic_thresholding_ratio: float = 0.995,
clip_sample_range: float = 1.0,
sample_max_value: float = 1.0,
timestep_spacing: str = "leading",
rescale_betas_zero_snr: bool = False,
max_timesteps: float = 1.57080,
intermediate_timesteps: Optional[int] = 1.3,
sigma_data: float = 0.5,
):
# standard deviation of the initial noise distribution
self.init_noise_sigma = 1.0
# setable values
self.num_inference_steps = None
self.timesteps = torch.from_numpy(np.arange(0, num_train_timesteps)[::-1].copy().astype(np.int64))
def set_timesteps(
self,
num_inference_steps: int,
timesteps: torch.Tensor = None,
device: Union[str, torch.device] = None,
):
"""
Sets the discrete timesteps used for the diffusion chain (to be run before inference).
Args:
num_inference_steps (`int`):
The number of diffusion steps used when generating samples with a pre-trained model.
"""
if num_inference_steps > self.config.num_train_timesteps:
raise ValueError(
f"`num_inference_steps`: {num_inference_steps} cannot be larger than `self.config.train_timesteps`:"
f" {self.config.num_train_timesteps} as the unet model trained with this scheduler can only handle"
f" maximal {self.config.num_train_timesteps} timesteps."
)
self.num_inference_steps = num_inference_steps
if timesteps is not None and len(timesteps) == num_inference_steps + 1:
if isinstance(timesteps, list):
self.timesteps = torch.tensor(timesteps, device=device).float()
elif isinstance(timesteps, torch.Tensor):
self.timesteps = timesteps.to(device).float()
else:
raise ValueError(f"Unsupported timesteps type: {type(timesteps)}")
elif self.config.intermediate_timesteps and num_inference_steps == 2:
self.timesteps = torch.tensor([self.config.max_timesteps, self.config.intermediate_timesteps, 0], device=device).float()
elif self.config.intermediate_timesteps:
self.timesteps = torch.linspace(self.config.max_timesteps, 0, num_inference_steps + 1, device=device).float()
warnings.warn(
f"Intermediate timesteps for SCM is not supported when num_inference_steps != 2. "
f"Reset timesteps to {self.timesteps} default max_timesteps"
)
else:
# max_timesteps=arctan(80/0.5)=1.56454 is the default from sCM paper, we choose a different value here
self.timesteps = torch.linspace(self.config.max_timesteps, 0, num_inference_steps + 1, device=device).float()
print(f"Set timesteps: {self.timesteps}")
def step(
self,
model_output: torch.FloatTensor,
timeindex: int,
timestep: float,
sample: torch.FloatTensor,
generator: torch.Generator = None,
return_dict: bool = True,
) -> Union[SCMSchedulerOutput, Tuple]:
"""
Predict the sample from the previous timestep by reversing the SDE. This function propagates the diffusion
process from the learned model outputs (most often the predicted noise).
Args:
model_output (`torch.FloatTensor`):
The direct output from learned diffusion model.
timestep (`float`):
The current discrete timestep in the diffusion chain.
sample (`torch.FloatTensor`):
A current instance of a sample created by the diffusion process.
return_dict (`bool`, *optional*, defaults to `True`):
itself. Useful for methods such as [`CycleDiffusion`].
return_dict (`bool`, *optional*, defaults to `True`):
Whether or not to return a [`~schedulers.scheduling_lcm.LCMSchedulerOutput`] or `tuple`.
Returns:
[`~schedulers.scheduling_utils.SCMSchedulerOutput`] or `tuple`:
If return_dict is `True`, [`~schedulers.scheduling_scm.SCMSchedulerOutput`] is returned, otherwise a
tuple is returned where the first element is the sample tensor.
"""
if self.num_inference_steps is None:
raise ValueError(
"Number of inference steps is 'None', you need to run 'set_timesteps' after creating the scheduler"
)
# 2. compute alphas, betas
t = self.timesteps[timeindex + 1]
s = self.timesteps[timeindex]
# 4. Different Parameterization:
parameterization = self.config.prediction_type
if parameterization == "trigflow":
pred_x0 = torch.cos(s) * sample - torch.sin(s) * model_output
else:
raise ValueError(f"Unsupported parameterization: {parameterization}")
# 5. Sample z ~ N(0, I), For MultiStep Inference
# Noise is not used for one-step sampling.
if len(self.timesteps) > 1:
noise = torch.randn(model_output.shape, device=model_output.device, generator=generator) * self.config.sigma_data
prev_sample = torch.cos(t) * pred_x0 + torch.sin(t) * noise
else:
prev_sample = pred_x0
if not return_dict:
return (prev_sample, pred_x0)
return SCMSchedulerOutput(prev_sample=prev_sample, denoised=pred_x0)
def __len__(self):
return self.config.num_train_timesteps