Merge branch 'main' into lora-load-adapter

Author: sayakpaul

.github/workflows/push_tests.yml

@@ -81,7 +81,7 @@ jobs:
       - name: Environment
         run: |
           python utils/print_env.py
-      - name: Slow PyTorch CUDA checkpoint tests on Ubuntu
+      - name: PyTorch CUDA checkpoint tests on Ubuntu
         env:
           HF_TOKEN: ${{ secrets.HF_TOKEN }}
           # https://pytorch.org/docs/stable/notes/randomness.html#avoiding-nondeterministic-algorithms
@@ -184,7 +184,7 @@ jobs:
       run: |
         python utils/print_env.py
 
-    - name: Run slow Flax TPU tests
+    - name: Run Flax TPU tests
       env:
         HF_TOKEN: ${{ secrets.HF_TOKEN }}
       run: |
@@ -232,7 +232,7 @@ jobs:
       run: |
         python utils/print_env.py
 
-    - name: Run slow ONNXRuntime CUDA tests
+    - name: Run ONNXRuntime CUDA tests
       env:
         HF_TOKEN: ${{ secrets.HF_TOKEN }}
       run: |

examples/dreambooth/train_dreambooth_lora_sd3.py

@@ -86,6 +86,15 @@ def save_model_card(
     validation_prompt=None,
     repo_folder=None,
 ):
+    if "large" in base_model:
+        model_variant = "SD3.5-Large"
+        license_url = "https://huggingface.co/stabilityai/stable-diffusion-3.5-large/blob/main/LICENSE.md"
+        variant_tags = ["sd3.5-large", "sd3.5", "sd3.5-diffusers"]
+    else:
+        model_variant = "SD3"
+        license_url = "https://huggingface.co/stabilityai/stable-diffusion-3-medium/blob/main/LICENSE.md"
+        variant_tags = ["sd3", "sd3-diffusers"]
+
     widget_dict = []
     if images is not None:
         for i, image in enumerate(images):
@@ -95,7 +104,7 @@ def save_model_card(
             )
 
     model_description = f"""
-# SD3 DreamBooth LoRA - {repo_id}
+# {model_variant} DreamBooth LoRA - {repo_id}
 
 <Gallery />
 
@@ -120,7 +129,7 @@ def save_model_card(
 ```py
 from diffusers import AutoPipelineForText2Image
 import torch
-pipeline = AutoPipelineForText2Image.from_pretrained('stabilityai/stable-diffusion-3-medium-diffusers', torch_dtype=torch.float16).to('cuda')
+pipeline = AutoPipelineForText2Image.from_pretrained({base_model}, torch_dtype=torch.float16).to('cuda')
 pipeline.load_lora_weights('{repo_id}', weight_name='pytorch_lora_weights.safetensors')
 image = pipeline('{validation_prompt if validation_prompt else instance_prompt}').images[0]
 ```
@@ -135,7 +144,7 @@ def save_model_card(
 
 ## License
 
-Please adhere to the licensing terms as described [here](https://huggingface.co/stabilityai/stable-diffusion-3-medium/blob/main/LICENSE).
+Please adhere to the licensing terms as described [here]({license_url}).
 """
     model_card = load_or_create_model_card(
         repo_id_or_path=repo_id,
@@ -151,11 +160,11 @@ def save_model_card(
         "diffusers-training",
         "diffusers",
         "lora",
-        "sd3",
-        "sd3-diffusers",
         "template:sd-lora",
     ]
 
+    tags += variant_tags
+
     model_card = populate_model_card(model_card, tags=tags)
     model_card.save(os.path.join(repo_folder, "README.md"))
 

examples/dreambooth/train_dreambooth_sd3.py

@@ -77,6 +77,15 @@ def save_model_card(
     validation_prompt=None,
     repo_folder=None,
 ):
+    if "large" in base_model:
+        model_variant = "SD3.5-Large"
+        license_url = "https://huggingface.co/stabilityai/stable-diffusion-3.5-large/blob/main/LICENSE.md"
+        variant_tags = ["sd3.5-large", "sd3.5", "sd3.5-diffusers"]
+    else:
+        model_variant = "SD3"
+        license_url = "https://huggingface.co/stabilityai/stable-diffusion-3-medium/blob/main/LICENSE.md"
+        variant_tags = ["sd3", "sd3-diffusers"]
+
     widget_dict = []
     if images is not None:
         for i, image in enumerate(images):
@@ -86,7 +95,7 @@ def save_model_card(
             )
 
     model_description = f"""
-# SD3 DreamBooth - {repo_id}
+# {model_variant} DreamBooth - {repo_id}
 
 <Gallery />
 
@@ -113,7 +122,7 @@ def save_model_card(
 
 ## License
 
-Please adhere to the licensing terms as described `[here](https://huggingface.co/stabilityai/stable-diffusion-3-medium/blob/main/LICENSE)`.
+Please adhere to the licensing terms as described `[here]({license_url})`.
 """
     model_card = load_or_create_model_card(
         repo_id_or_path=repo_id,
@@ -128,10 +137,9 @@ def save_model_card(
         "text-to-image",
         "diffusers-training",
         "diffusers",
-        "sd3",
-        "sd3-diffusers",
         "template:sd-lora",
     ]
+    tags += variant_tags
 
     model_card = populate_model_card(model_card, tags=tags)
     model_card.save(os.path.join(repo_folder, "README.md"))

src/diffusers/callbacks.py

@@ -97,13 +97,17 @@ def callback_fn(self, pipeline, step_index, timestep, callback_kwargs) -> Dict[s
 
 class SDXLCFGCutoffCallback(PipelineCallback):
     """
-    Callback function for Stable Diffusion XL Pipelines. After certain number of steps (set by `cutoff_step_ratio` or
-    `cutoff_step_index`), this callback will disable the CFG.
+    Callback function for the base Stable Diffusion XL Pipelines. After certain number of steps (set by
+    `cutoff_step_ratio` or `cutoff_step_index`), this callback will disable the CFG.
 
     Note: This callback mutates the pipeline by changing the `_guidance_scale` attribute to 0.0 after the cutoff step.
     """
 
-    tensor_inputs = ["prompt_embeds", "add_text_embeds", "add_time_ids"]
+    tensor_inputs = [
+        "prompt_embeds",
+        "add_text_embeds",
+        "add_time_ids",
+    ]
 
     def callback_fn(self, pipeline, step_index, timestep, callback_kwargs) -> Dict[str, Any]:
         cutoff_step_ratio = self.config.cutoff_step_ratio
@@ -129,6 +133,55 @@ def callback_fn(self, pipeline, step_index, timestep, callback_kwargs) -> Dict[s
             callback_kwargs[self.tensor_inputs[0]] = prompt_embeds
             callback_kwargs[self.tensor_inputs[1]] = add_text_embeds
             callback_kwargs[self.tensor_inputs[2]] = add_time_ids
+
+        return callback_kwargs
+
+
+class SDXLControlnetCFGCutoffCallback(PipelineCallback):
+    """
+    Callback function for the Controlnet Stable Diffusion XL Pipelines. After certain number of steps (set by
+    `cutoff_step_ratio` or `cutoff_step_index`), this callback will disable the CFG.
+
+    Note: This callback mutates the pipeline by changing the `_guidance_scale` attribute to 0.0 after the cutoff step.
+    """
+
+    tensor_inputs = [
+        "prompt_embeds",
+        "add_text_embeds",
+        "add_time_ids",
+        "image",
+    ]
+
+    def callback_fn(self, pipeline, step_index, timestep, callback_kwargs) -> Dict[str, Any]:
+        cutoff_step_ratio = self.config.cutoff_step_ratio
+        cutoff_step_index = self.config.cutoff_step_index
+
+        # Use cutoff_step_index if it's not None, otherwise use cutoff_step_ratio
+        cutoff_step = (
+            cutoff_step_index if cutoff_step_index is not None else int(pipeline.num_timesteps * cutoff_step_ratio)
+        )
+
+        if step_index == cutoff_step:
+            prompt_embeds = callback_kwargs[self.tensor_inputs[0]]
+            prompt_embeds = prompt_embeds[-1:]  # "-1" denotes the embeddings for conditional text tokens.
+
+            add_text_embeds = callback_kwargs[self.tensor_inputs[1]]
+            add_text_embeds = add_text_embeds[-1:]  # "-1" denotes the embeddings for conditional pooled text tokens
+
+            add_time_ids = callback_kwargs[self.tensor_inputs[2]]
+            add_time_ids = add_time_ids[-1:]  # "-1" denotes the embeddings for conditional added time vector
+
+            # For Controlnet
+            image = callback_kwargs[self.tensor_inputs[3]]
+            image = image[-1:]
+
+            pipeline._guidance_scale = 0.0
+
+            callback_kwargs[self.tensor_inputs[0]] = prompt_embeds
+            callback_kwargs[self.tensor_inputs[1]] = add_text_embeds
+            callback_kwargs[self.tensor_inputs[2]] = add_time_ids
+            callback_kwargs[self.tensor_inputs[3]] = image
+
         return callback_kwargs
 
 

src/diffusers/pipelines/controlnet/pipeline_controlnet_sd_xl.py

@@ -242,6 +242,7 @@ class StableDiffusionXLControlNetPipeline(
         "add_time_ids",
         "negative_pooled_prompt_embeds",
         "negative_add_time_ids",
+        "image",
     ]
 
     def __init__(
@@ -1540,6 +1541,7 @@ def __call__(
                     )
                     add_time_ids = callback_outputs.pop("add_time_ids", add_time_ids)
                     negative_add_time_ids = callback_outputs.pop("negative_add_time_ids", negative_add_time_ids)
+                    image = callback_outputs.pop("image", image)
 
                 # call the callback, if provided
                 if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):

src/diffusers/pipelines/flux/pipeline_flux_controlnet_image_to_image.py

@@ -903,9 +903,12 @@ def __call__(
 
                 timestep = t.expand(latents.shape[0]).to(latents.dtype)
 
-                guidance = (
-                    torch.tensor([guidance_scale], device=device) if self.controlnet.config.guidance_embeds else None
-                )
+                if isinstance(self.controlnet, FluxMultiControlNetModel):
+                    use_guidance = self.controlnet.nets[0].config.guidance_embeds
+                else:
+                    use_guidance = self.controlnet.config.guidance_embeds
+
+                guidance = torch.tensor([guidance_scale], device=device) if use_guidance else None
                 guidance = guidance.expand(latents.shape[0]) if guidance is not None else None
 
                 if isinstance(controlnet_keep[i], list):

tests/lora/test_lora_layers_flux.py

@@ -27,6 +27,7 @@
 from diffusers.utils.testing_utils import (
     floats_tensor,
     is_peft_available,
+    nightly,
     numpy_cosine_similarity_distance,
     require_peft_backend,
     require_torch_gpu,
@@ -165,9 +166,10 @@ def test_modify_padding_mode(self):
 
 
 @slow
+@nightly
 @require_torch_gpu
 @require_peft_backend
-# @unittest.skip("We cannot run inference on this model with the current CI hardware")
+@unittest.skip("We cannot run inference on this model with the current CI hardware")
 # TODO (DN6, sayakpaul): move these tests to a beefier GPU
 class FluxLoRAIntegrationTests(unittest.TestCase):
     """internal note: The integration slices were obtained on audace.

tests/lora/test_lora_layers_sd.py

@@ -34,6 +34,7 @@
 from diffusers.utils.import_utils import is_accelerate_available
 from diffusers.utils.testing_utils import (
     load_image,
+    nightly,
     numpy_cosine_similarity_distance,
     require_peft_backend,
     require_torch_gpu,
@@ -207,6 +208,7 @@ def test_integration_move_lora_dora_cpu(self):
 
 
 @slow
+@nightly
 @require_torch_gpu
 @require_peft_backend
 class LoraIntegrationTests(unittest.TestCase):

tests/lora/test_lora_layers_sdxl.py

@@ -113,6 +113,7 @@ def tearDown(self):
 
 
 @slow
+@nightly
 @require_torch_gpu
 @require_peft_backend
 class LoraSDXLIntegrationTests(unittest.TestCase):