[qwen] Qwen image edit followups (#12166)

* add docs.

* more docs.

* xfail full compilation for Qwen for now.

* tests

* up

* up

* up

* reviewer feedback.

docs/source/en/api/pipelines/qwenimage.md

@@ -16,7 +16,12 @@
 
 Qwen-Image from the Qwen team is an image generation foundation model in the Qwen series that achieves significant advances in complex text rendering and precise image editing. Experiments show strong general capabilities in both image generation and editing, with exceptional performance in text rendering, especially for Chinese.
 
-Check out the model card [here](https://huggingface.co/Qwen/Qwen-Image) to learn more.
+Qwen-Image comes in the following variants:
+
+| model type | model id |
+|:----------:|:--------:|
+| Qwen-Image | [`Qwen/Qwen-Image`](https://huggingface.co/Qwen/Qwen-Image) |
+| Qwen-Image-Edit | [`Qwen/Qwen-Image-Edit`](https://huggingface.co/Qwen/Qwen-Image-Edit) |
 
 <Tip>
 
@@ -87,10 +92,6 @@ image.save("qwen_fewsteps.png")
   - all
   - __call__
 
-## QwenImagePipelineOutput
-
-[[autodoc]] pipelines.qwenimage.pipeline_output.QwenImagePipelineOutput
-
 ## QwenImageImg2ImgPipeline
 
 [[autodoc]] QwenImageImg2ImgPipeline
@@ -102,3 +103,13 @@ image.save("qwen_fewsteps.png")
 [[autodoc]] QwenImageInpaintPipeline
   - all
   - __call__
+
+## QwenImageEditPipeline
+
+[[autodoc]] QwenImageEditPipeline
+  - all
+  - __call__
+
+## QwenImagePipelineOutput
+
+[[autodoc]] pipelines.qwenimage.pipeline_output.QwenImagePipelineOutput

src/diffusers/__init__.py

@@ -489,10 +489,10 @@ else:
             "PixArtAlphaPipeline",
             "PixArtSigmaPAGPipeline",
             "PixArtSigmaPipeline",
+            "QwenImageEditPipeline",
             "QwenImageImg2ImgPipeline",
             "QwenImageInpaintPipeline",
             "QwenImagePipeline",
-            "QwenImageEditPipeline",
             "ReduxImageEncoder",
             "SanaControlNetPipeline",
             "SanaPAGPipeline",

src/diffusers/models/transformers/transformer_qwenimage.py

@@ -219,6 +219,7 @@ class QwenEmbedRope(nn.Module):
                 video_freq = self.rope_cache[rope_key]
             else:
                 video_freq = self._compute_video_freqs(frame, height, width, idx)
+            video_freq = video_freq.to(device)
             vid_freqs.append(video_freq)
 
             if self.scale_rope:

src/diffusers/pipelines/qwenimage/__init__.py

@@ -24,9 +24,9 @@ except OptionalDependencyNotAvailable:
 else:
     _import_structure["modeling_qwenimage"] = ["ReduxImageEncoder"]
     _import_structure["pipeline_qwenimage"] = ["QwenImagePipeline"]
+    _import_structure["pipeline_qwenimage_edit"] = ["QwenImageEditPipeline"]
     _import_structure["pipeline_qwenimage_img2img"] = ["QwenImageImg2ImgPipeline"]
     _import_structure["pipeline_qwenimage_inpaint"] = ["QwenImageInpaintPipeline"]
-    _import_structure["pipeline_qwenimage_edit"] = ["QwenImageEditPipeline"]
 
 if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
     try:

src/diffusers/pipelines/qwenimage/pipeline_qwenimage_edit.py

@@ -46,15 +46,20 @@ EXAMPLE_DOC_STRING = """
         >>> import torch
         >>> from PIL import Image
         >>> from diffusers import QwenImageEditPipeline
+        >>> from diffusers.utils import load_image
 
         >>> pipe = QwenImageEditPipeline.from_pretrained("Qwen/Qwen-Image-Edit", torch_dtype=torch.bfloat16)
         >>> pipe.to("cuda")
-        >>> prompt = "Change the cat to a dog"
-        >>> image = Image.open("cat.png")
+        >>> image = load_image(
+        ...     "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/yarn-art-pikachu.png"
+        ... ).convert("RGB")
+        >>> prompt = (
+        ...     "Make Pikachu hold a sign that says 'Qwen Edit is awesome', yarn art style, detailed, vibrant colors"
+        ... )
         >>> # Depending on the variant being used, the pipeline call will slightly vary.
         >>> # Refer to the pipeline documentation for more details.
         >>> image = pipe(image, prompt, num_inference_steps=50).images[0]
-        >>> image.save("qwenimageedit.png")
+        >>> image.save("qwenimage_edit.png")
         ```
 """
 PREFERRED_QWENIMAGE_RESOLUTIONS = [
@@ -178,7 +183,7 @@ def calculate_dimensions(target_area, ratio):
 
 class QwenImageEditPipeline(DiffusionPipeline, QwenImageLoraLoaderMixin):
     r"""
-    The QwenImage pipeline for text-to-image generation.
+    The Qwen-Image-Edit pipeline for image editing.
 
     Args:
         transformer ([`QwenImageTransformer2DModel`]):
@@ -217,8 +222,8 @@ class QwenImageEditPipeline(DiffusionPipeline, QwenImageLoraLoaderMixin):
             transformer=transformer,
             scheduler=scheduler,
         )
-        self.latent_channels = 16
         self.vae_scale_factor = 2 ** len(self.vae.temperal_downsample) if getattr(self, "vae", None) else 8
+        self.latent_channels = self.vae.config.z_dim if getattr(self, "vae", None) else 16
         # QwenImage latents are turned into 2x2 patches and packed. This means the latent width and height has to be divisible
         # by the patch size. So the vae scale factor is multiplied by the patch size to account for this
         self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor * 2)
@@ -635,7 +640,9 @@ class QwenImageEditPipeline(DiffusionPipeline, QwenImageLoraLoaderMixin):
             [`~pipelines.qwenimage.QwenImagePipelineOutput`] if `return_dict` is True, otherwise a `tuple`. When
             returning a tuple, the first element is a list with the generated images.
         """
-        calculated_width, calculated_height, _ = calculate_dimensions(1024 * 1024, image.width / image.height)
+        image_size = image[0].size if isinstance(image, list) else image.size
+        width, height = image_size
+        calculated_width, calculated_height, _ = calculate_dimensions(1024 * 1024, width / height)
         height = height or calculated_height
         width = width or calculated_width
 

tests/models/transformers/test_models_transformer_qwenimage.py

@@ -15,6 +15,7 @@
 
 import unittest
 
+import pytest
 import torch
 
 from diffusers import QwenImageTransformer2DModel
@@ -99,3 +100,7 @@ class QwenImageTransformerCompileTests(TorchCompileTesterMixin, unittest.TestCas
 
     def prepare_dummy_input(self, height, width):
         return QwenImageTransformerTests().prepare_dummy_input(height=height, width=width)
+
+    @pytest.mark.xfail(condition=True, reason="RoPE needs to be revisited.", strict=True)
+    def test_torch_compile_recompilation_and_graph_break(self):
+        super().test_torch_compile_recompilation_and_graph_break()

tests/pipelines/qwenimage/test_qwenimage_edit.py

@@ -0,0 +1,243 @@
+# Copyright 2025 The HuggingFace Team.
+#
+# 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 unittest
+
+import numpy as np
+import pytest
+import torch
+from PIL import Image
+from transformers import Qwen2_5_VLConfig, Qwen2_5_VLForConditionalGeneration, Qwen2Tokenizer, Qwen2VLProcessor
+
+from diffusers import (
+    AutoencoderKLQwenImage,
+    FlowMatchEulerDiscreteScheduler,
+    QwenImageEditPipeline,
+    QwenImageTransformer2DModel,
+)
+from diffusers.utils.testing_utils import enable_full_determinism, torch_device
+
+from ..pipeline_params import TEXT_TO_IMAGE_PARAMS
+from ..test_pipelines_common import PipelineTesterMixin, to_np
+
+
+enable_full_determinism()
+
+
+class QwenImageEditPipelineFastTests(PipelineTesterMixin, unittest.TestCase):
+    pipeline_class = QwenImageEditPipeline
+    params = TEXT_TO_IMAGE_PARAMS - {"cross_attention_kwargs"}
+    batch_params = frozenset(["prompt", "image"])
+    image_params = frozenset(["image"])
+    image_latents_params = frozenset(["latents"])
+    required_optional_params = frozenset(
+        [
+            "num_inference_steps",
+            "generator",
+            "latents",
+            "return_dict",
+            "callback_on_step_end",
+            "callback_on_step_end_tensor_inputs",
+        ]
+    )
+    supports_dduf = False
+    test_xformers_attention = False
+    test_layerwise_casting = True
+    test_group_offloading = True
+
+    def get_dummy_components(self):
+        tiny_ckpt_id = "hf-internal-testing/tiny-random-Qwen2VLForConditionalGeneration"
+
+        torch.manual_seed(0)
+        transformer = QwenImageTransformer2DModel(
+            patch_size=2,
+            in_channels=16,
+            out_channels=4,
+            num_layers=2,
+            attention_head_dim=16,
+            num_attention_heads=3,
+            joint_attention_dim=16,
+            guidance_embeds=False,
+            axes_dims_rope=(8, 4, 4),
+        )
+
+        torch.manual_seed(0)
+        z_dim = 4
+        vae = AutoencoderKLQwenImage(
+            base_dim=z_dim * 6,
+            z_dim=z_dim,
+            dim_mult=[1, 2, 4],
+            num_res_blocks=1,
+            temperal_downsample=[False, True],
+            latents_mean=[0.0] * z_dim,
+            latents_std=[1.0] * z_dim,
+        )
+
+        torch.manual_seed(0)
+        scheduler = FlowMatchEulerDiscreteScheduler()
+
+        torch.manual_seed(0)
+        config = Qwen2_5_VLConfig(
+            text_config={
+                "hidden_size": 16,
+                "intermediate_size": 16,
+                "num_hidden_layers": 2,
+                "num_attention_heads": 2,
+                "num_key_value_heads": 2,
+                "rope_scaling": {
+                    "mrope_section": [1, 1, 2],
+                    "rope_type": "default",
+                    "type": "default",
+                },
+                "rope_theta": 1000000.0,
+            },
+            vision_config={
+                "depth": 2,
+                "hidden_size": 16,
+                "intermediate_size": 16,
+                "num_heads": 2,
+                "out_hidden_size": 16,
+            },
+            hidden_size=16,
+            vocab_size=152064,
+            vision_end_token_id=151653,
+            vision_start_token_id=151652,
+            vision_token_id=151654,
+        )
+        text_encoder = Qwen2_5_VLForConditionalGeneration(config)
+        tokenizer = Qwen2Tokenizer.from_pretrained(tiny_ckpt_id)
+
+        components = {
+            "transformer": transformer,
+            "vae": vae,
+            "scheduler": scheduler,
+            "text_encoder": text_encoder,
+            "tokenizer": tokenizer,
+            "processor": Qwen2VLProcessor.from_pretrained(tiny_ckpt_id),
+        }
+        return components
+
+    def get_dummy_inputs(self, device, seed=0):
+        if str(device).startswith("mps"):
+            generator = torch.manual_seed(seed)
+        else:
+            generator = torch.Generator(device=device).manual_seed(seed)
+
+        inputs = {
+            "prompt": "dance monkey",
+            "image": Image.new("RGB", (32, 32)),
+            "negative_prompt": "bad quality",
+            "generator": generator,
+            "num_inference_steps": 2,
+            "true_cfg_scale": 1.0,
+            "height": 32,
+            "width": 32,
+            "max_sequence_length": 16,
+            "output_type": "pt",
+        }
+
+        return inputs
+
+    def test_inference(self):
+        device = "cpu"
+
+        components = self.get_dummy_components()
+        pipe = self.pipeline_class(**components)
+        pipe.to(device)
+        pipe.set_progress_bar_config(disable=None)
+
+        inputs = self.get_dummy_inputs(device)
+        image = pipe(**inputs).images
+        generated_image = image[0]
+        self.assertEqual(generated_image.shape, (3, 32, 32))
+
+        # fmt: off
+        expected_slice = torch.tensor([[0.5637, 0.6341, 0.6001, 0.5620, 0.5794, 0.5498, 0.5757, 0.6389, 0.4174, 0.3597, 0.5649, 0.4894, 0.4969, 0.5255, 0.4083, 0.4986]])
+        # fmt: on
+
+        generated_slice = generated_image.flatten()
+        generated_slice = torch.cat([generated_slice[:8], generated_slice[-8:]])
+        self.assertTrue(torch.allclose(generated_slice, expected_slice, atol=1e-3))
+
+    def test_inference_batch_single_identical(self):
+        self._test_inference_batch_single_identical(batch_size=3, expected_max_diff=1e-1)
+
+    def test_attention_slicing_forward_pass(
+        self, test_max_difference=True, test_mean_pixel_difference=True, expected_max_diff=1e-3
+    ):
+        if not self.test_attention_slicing:
+            return
+
+        components = self.get_dummy_components()
+        pipe = self.pipeline_class(**components)
+        for component in pipe.components.values():
+            if hasattr(component, "set_default_attn_processor"):
+                component.set_default_attn_processor()
+        pipe.to(torch_device)
+        pipe.set_progress_bar_config(disable=None)
+
+        generator_device = "cpu"
+        inputs = self.get_dummy_inputs(generator_device)
+        output_without_slicing = pipe(**inputs)[0]
+
+        pipe.enable_attention_slicing(slice_size=1)
+        inputs = self.get_dummy_inputs(generator_device)
+        output_with_slicing1 = pipe(**inputs)[0]
+
+        pipe.enable_attention_slicing(slice_size=2)
+        inputs = self.get_dummy_inputs(generator_device)
+        output_with_slicing2 = pipe(**inputs)[0]
+
+        if test_max_difference:
+            max_diff1 = np.abs(to_np(output_with_slicing1) - to_np(output_without_slicing)).max()
+            max_diff2 = np.abs(to_np(output_with_slicing2) - to_np(output_without_slicing)).max()
+            self.assertLess(
+                max(max_diff1, max_diff2),
+                expected_max_diff,
+                "Attention slicing should not affect the inference results",
+            )
+
+    def test_vae_tiling(self, expected_diff_max: float = 0.2):
+        generator_device = "cpu"
+        components = self.get_dummy_components()
+
+        pipe = self.pipeline_class(**components)
+        pipe.to("cpu")
+        pipe.set_progress_bar_config(disable=None)
+
+        # Without tiling
+        inputs = self.get_dummy_inputs(generator_device)
+        inputs["height"] = inputs["width"] = 128
+        output_without_tiling = pipe(**inputs)[0]
+
+        # With tiling
+        pipe.vae.enable_tiling(
+            tile_sample_min_height=96,
+            tile_sample_min_width=96,
+            tile_sample_stride_height=64,
+            tile_sample_stride_width=64,
+        )
+        inputs = self.get_dummy_inputs(generator_device)
+        inputs["height"] = inputs["width"] = 128
+        output_with_tiling = pipe(**inputs)[0]
+
+        self.assertLess(
+            (to_np(output_without_tiling) - to_np(output_with_tiling)).max(),
+            expected_diff_max,
+            "VAE tiling should not affect the inference results",
+        )
+
+    @pytest.mark.xfail(condition=True, reason="Preconfigured embeddings need to be revisited.", strict=True)
+    def test_encode_prompt_works_in_isolation(self, extra_required_param_value_dict=None, atol=1e-4, rtol=1e-4):
+        super().test_encode_prompt_works_in_isolation(extra_required_param_value_dict, atol, rtol)