Support of ip-adapter to the StableDiffusionControlNetInpaintPipeline

src/diffusers/pipelines/controlnet/pipeline_controlnet_inpaint.py

@@ -21,10 +21,10 @@
 import PIL.Image
 import torch
 import torch.nn.functional as F
-from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer
+from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer, CLIPVisionModelWithProjection
 
 from ...image_processor import PipelineImageInput, VaeImageProcessor
-from ...loaders import FromSingleFileMixin, LoraLoaderMixin, TextualInversionLoaderMixin
+from ...loaders import FromSingleFileMixin, IPAdapterMixin, LoraLoaderMixin, TextualInversionLoaderMixin
 from ...models import AutoencoderKL, ControlNetModel, UNet2DConditionModel
 from ...models.lora import adjust_lora_scale_text_encoder
 from ...schedulers import KarrasDiffusionSchedulers
@@ -237,7 +237,7 @@ def prepare_mask_and_masked_image(image, mask, height, width, return_image=False
 
 
 class StableDiffusionControlNetInpaintPipeline(
-    DiffusionPipeline, TextualInversionLoaderMixin, LoraLoaderMixin, FromSingleFileMixin
+    DiffusionPipeline, TextualInversionLoaderMixin, LoraLoaderMixin, IPAdapterMixin, FromSingleFileMixin
 ):
     r"""
     Pipeline for image inpainting using Stable Diffusion with ControlNet guidance.
@@ -247,6 +247,7 @@ class StableDiffusionControlNetInpaintPipeline(
 
     The pipeline also inherits the following loading methods:
         - [`~loaders.TextualInversionLoaderMixin.load_textual_inversion`] for loading textual inversion embeddings
+        - [`~loaders.IPAdapterMixin.load_ip_adapter`] for loading IP Adapters
 
     <Tip>
 
@@ -284,7 +285,7 @@ class StableDiffusionControlNetInpaintPipeline(
     """
 
     model_cpu_offload_seq = "text_encoder->unet->vae"
-    _optional_components = ["safety_checker", "feature_extractor"]
+    _optional_components = ["safety_checker", "feature_extractor", "image_encoder"]
     _exclude_from_cpu_offload = ["safety_checker"]
     _callback_tensor_inputs = ["latents", "prompt_embeds", "negative_prompt_embeds"]
 
@@ -298,6 +299,7 @@ def __init__(
         scheduler: KarrasDiffusionSchedulers,
         safety_checker: StableDiffusionSafetyChecker,
         feature_extractor: CLIPImageProcessor,
+        image_encoder: CLIPVisionModelWithProjection = None,
         requires_safety_checker: bool = True,
     ):
         super().__init__()
@@ -330,6 +332,7 @@ def __init__(
             scheduler=scheduler,
             safety_checker=safety_checker,
             feature_extractor=feature_extractor,
+            image_encoder=image_encoder,
         )
         self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
         self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor)
@@ -589,6 +592,20 @@ def encode_prompt(
 
         return prompt_embeds, negative_prompt_embeds
 
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_image
+    def encode_image(self, image, device, num_images_per_prompt):
+        dtype = next(self.image_encoder.parameters()).dtype
+
+        if not isinstance(image, torch.Tensor):
+            image = self.feature_extractor(image, return_tensors="pt").pixel_values
+
+        image = image.to(device=device, dtype=dtype)
+        image_embeds = self.image_encoder(image).image_embeds
+        image_embeds = image_embeds.repeat_interleave(num_images_per_prompt, dim=0)
+
+        uncond_image_embeds = torch.zeros_like(image_embeds)
+        return image_embeds, uncond_image_embeds
+
     # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.run_safety_checker
     def run_safety_checker(self, image, device, dtype):
         if self.safety_checker is None:
@@ -1049,6 +1066,7 @@ def __call__(
         latents: Optional[torch.FloatTensor] = None,
         prompt_embeds: Optional[torch.FloatTensor] = None,
         negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+        ip_adapter_image: Optional[PipelineImageInput] = None,
         output_type: Optional[str] = "pil",
         return_dict: bool = True,
         cross_attention_kwargs: Optional[Dict[str, Any]] = None,
@@ -1127,6 +1145,7 @@ def __call__(
             negative_prompt_embeds (`torch.FloatTensor`, *optional*):
                 Pre-generated negative text embeddings. Can be used to easily tweak text inputs (prompt weighting). If
                 not provided, `negative_prompt_embeds` are generated from the `negative_prompt` input argument.
+            ip_adapter_image: (`PipelineImageInput`, *optional*): Optional image input to work with IP Adapters.
             output_type (`str`, *optional*, defaults to `"pil"`):
                 The output format of the generated image. Choose between `PIL.Image` or `np.array`.
             return_dict (`bool`, *optional*, defaults to `True`):
@@ -1260,6 +1279,11 @@ def __call__(
         if self.do_classifier_free_guidance:
             prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds])
 
+        if ip_adapter_image is not None:
+            image_embeds, negative_image_embeds = self.encode_image(ip_adapter_image, device, num_images_per_prompt)
+            if self.do_classifier_free_guidance:
+                image_embeds = torch.cat([negative_image_embeds, image_embeds])
+
         # 4. Prepare image
         if isinstance(controlnet, ControlNetModel):
             control_image = self.prepare_control_image(
@@ -1295,7 +1319,7 @@ def __call__(
         else:
             assert False
 
-        # 4. Preprocess mask and image - resizes image and mask w.r.t height and width
+        # 4.1 Preprocess mask and image - resizes image and mask w.r.t height and width
         init_image = self.image_processor.preprocess(image, height=height, width=width)
         init_image = init_image.to(dtype=torch.float32)
 
@@ -1356,7 +1380,10 @@ def __call__(
         # 7. 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)
 
-        # 7.1 Create tensor stating which controlnets to keep
+        # 7.1 Add image embeds for IP-Adapter
+        added_cond_kwargs = {"image_embeds": image_embeds} if ip_adapter_image is not None else None
+
+        # 7.3 Create tensor stating which controlnets to keep
         controlnet_keep = []
         for i in range(len(timesteps)):
             keeps = [
@@ -1419,6 +1446,7 @@ def __call__(
                     cross_attention_kwargs=self.cross_attention_kwargs,
                     down_block_additional_residuals=down_block_res_samples,
                     mid_block_additional_residual=mid_block_res_sample,
+                    added_cond_kwargs=added_cond_kwargs,
                     return_dict=False,
                 )[0]
 

tests/pipelines/controlnet/test_controlnet_inpaint.py

@@ -132,6 +132,7 @@ def get_dummy_components(self):
             "tokenizer": tokenizer,
             "safety_checker": None,
             "feature_extractor": None,
+            "image_encoder": None,
         }
         return components
 
@@ -248,6 +249,7 @@ def get_dummy_components(self):
             "tokenizer": tokenizer,
             "safety_checker": None,
             "feature_extractor": None,
+            "image_encoder": None,
         }
         return components
 
@@ -342,6 +344,7 @@ def init_weights(m):
             "tokenizer": tokenizer,
             "safety_checker": None,
             "feature_extractor": None,
+            "image_encoder": None,
         }
         return components