diffusers 0.24.0__py3-none-any.whl → 0.25.0__py3-none-any.whl

diffusers/pipelines/stable_diffusion_xl/pipeline_stable_diffusion_xl_inpaint.py

@@ -33,9 +33,10 @@ from ...loaders import (
     StableDiffusionXLLoraLoaderMixin,
     TextualInversionLoaderMixin,
 )
-from ...models import AutoencoderKL, UNet2DConditionModel
+from ...models import AutoencoderKL, ImageProjection, UNet2DConditionModel
 from ...models.attention_processor import (
     AttnProcessor2_0,
+    FusedAttnProcessor2_0,
     LoRAAttnProcessor2_0,
     LoRAXFormersAttnProcessor,
     XFormersAttnProcessor,
@@ -321,12 +322,12 @@ class StableDiffusionXLInpaintPipeline(
     This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the
     library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.)
 
-    In addition the pipeline inherits the following loading methods:
-        - *LoRA*: [`loaders.StableDiffusionXLLoraLoaderMixin.load_lora_weights`]
-        - *Ckpt*: [`loaders.FromSingleFileMixin.from_single_file`]
-
-    as well as the following saving methods:
-        - *LoRA*: [`loaders.StableDiffusionXLLoraLoaderMixin.save_lora_weights`]
+    The pipeline also inherits the following loading methods:
+        - [`~loaders.TextualInversionLoaderMixin.load_textual_inversion`] for loading textual inversion embeddings
+        - [`~loaders.FromSingleFileMixin.from_single_file`] for loading `.ckpt` files
+        - [`~loaders.StableDiffusionXLLoraLoaderMixin.load_lora_weights`] for loading LoRA weights
+        - [`~loaders.StableDiffusionXLLoraLoaderMixin.save_lora_weights`] for saving LoRA weights
+        - [`~loaders.IPAdapterMixin.load_ip_adapter`] for loading IP Adapters
 
     Args:
         vae ([`AutoencoderKL`]):
@@ -363,7 +364,7 @@ class StableDiffusionXLInpaintPipeline(
             watermarker will be used.
     """
 
-    model_cpu_offload_seq = "text_encoder->text_encoder_2->unet->vae"
+    model_cpu_offload_seq = "text_encoder->text_encoder_2->image_encoder->unet->vae"
 
     _optional_components = [
         "tokenizer",
@@ -462,18 +463,29 @@ class StableDiffusionXLInpaintPipeline(
         self.vae.disable_tiling()
 
     # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_image
-    def encode_image(self, image, device, num_images_per_prompt):
+    def encode_image(self, image, device, num_images_per_prompt, output_hidden_states=None):
         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)
+        if output_hidden_states:
+            image_enc_hidden_states = self.image_encoder(image, output_hidden_states=True).hidden_states[-2]
+            image_enc_hidden_states = image_enc_hidden_states.repeat_interleave(num_images_per_prompt, dim=0)
+            uncond_image_enc_hidden_states = self.image_encoder(
+                torch.zeros_like(image), output_hidden_states=True
+            ).hidden_states[-2]
+            uncond_image_enc_hidden_states = uncond_image_enc_hidden_states.repeat_interleave(
+                num_images_per_prompt, dim=0
+            )
+            return image_enc_hidden_states, uncond_image_enc_hidden_states
+        else:
+            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)
 
-        uncond_image_embeds = torch.zeros_like(image_embeds)
-        return image_embeds, uncond_image_embeds
+            return image_embeds, uncond_image_embeds
 
     # Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl.StableDiffusionXLPipeline.encode_prompt
     def encode_prompt(
@@ -1073,6 +1085,67 @@ class StableDiffusionXLInpaintPipeline(
         """Disables the FreeU mechanism if enabled."""
         self.unet.disable_freeu()
 
+    # Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl.StableDiffusionXLPipeline.fuse_qkv_projections
+    def fuse_qkv_projections(self, unet: bool = True, vae: bool = True):
+        """
+        Enables fused QKV projections. For self-attention modules, all projection matrices (i.e., query,
+        key, value) are fused. For cross-attention modules, key and value projection matrices are fused.
+
+        <Tip warning={true}>
+
+        This API is 🧪 experimental.
+
+        </Tip>
+
+        Args:
+            unet (`bool`, defaults to `True`): To apply fusion on the UNet.
+            vae (`bool`, defaults to `True`): To apply fusion on the VAE.
+        """
+        self.fusing_unet = False
+        self.fusing_vae = False
+
+        if unet:
+            self.fusing_unet = True
+            self.unet.fuse_qkv_projections()
+            self.unet.set_attn_processor(FusedAttnProcessor2_0())
+
+        if vae:
+            if not isinstance(self.vae, AutoencoderKL):
+                raise ValueError("`fuse_qkv_projections()` is only supported for the VAE of type `AutoencoderKL`.")
+
+            self.fusing_vae = True
+            self.vae.fuse_qkv_projections()
+            self.vae.set_attn_processor(FusedAttnProcessor2_0())
+
+    # Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl.StableDiffusionXLPipeline.unfuse_qkv_projections
+    def unfuse_qkv_projections(self, unet: bool = True, vae: bool = True):
+        """Disable QKV projection fusion if enabled.
+
+        <Tip warning={true}>
+
+        This API is 🧪 experimental.
+
+        </Tip>
+
+        Args:
+            unet (`bool`, defaults to `True`): To apply fusion on the UNet.
+            vae (`bool`, defaults to `True`): To apply fusion on the VAE.
+
+        """
+        if unet:
+            if not self.fusing_unet:
+                logger.warning("The UNet was not initially fused for QKV projections. Doing nothing.")
+            else:
+                self.unet.unfuse_qkv_projections()
+                self.fusing_unet = False
+
+        if vae:
+            if not self.fusing_vae:
+                logger.warning("The VAE was not initially fused for QKV projections. Doing nothing.")
+            else:
+                self.vae.unfuse_qkv_projections()
+                self.fusing_vae = False
+
     # Copied from diffusers.pipelines.latent_consistency_models.pipeline_latent_consistency_text2img.LatentConsistencyModelPipeline.get_guidance_scale_embedding
     def get_guidance_scale_embedding(self, w, embedding_dim=512, dtype=torch.float32):
         """
@@ -1137,6 +1210,10 @@ class StableDiffusionXLInpaintPipeline(
     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__(
@@ -1389,6 +1466,7 @@ class StableDiffusionXLInpaintPipeline(
         self._cross_attention_kwargs = cross_attention_kwargs
         self._denoising_end = denoising_end
         self._denoising_start = denoising_start
+        self._interrupt = False
 
         # 2. Define call parameters
         if prompt is not None and isinstance(prompt, str):
@@ -1568,7 +1646,10 @@ class StableDiffusionXLInpaintPipeline(
         add_time_ids = add_time_ids.to(device)
 
         if ip_adapter_image is not None:
-            image_embeds, negative_image_embeds = self.encode_image(ip_adapter_image, device, num_images_per_prompt)
+            output_hidden_state = False if isinstance(self.unet.encoder_hid_proj, ImageProjection) else True
+            image_embeds, negative_image_embeds = self.encode_image(
+                ip_adapter_image, device, num_images_per_prompt, output_hidden_state
+            )
             if self.do_classifier_free_guidance:
                 image_embeds = torch.cat([negative_image_embeds, image_embeds])
                 image_embeds = image_embeds.to(device)
@@ -1608,6 +1689,8 @@ class StableDiffusionXLInpaintPipeline(
         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
                 # expand the latents if we are doing classifier free guidance
                 latent_model_input = torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents
 

diffusers/pipelines/stable_diffusion_xl/pipeline_stable_diffusion_xl_instruct_pix2pix.py

@@ -24,6 +24,7 @@ from ...loaders import FromSingleFileMixin, StableDiffusionXLLoraLoaderMixin, Te
 from ...models import AutoencoderKL, UNet2DConditionModel
 from ...models.attention_processor import (
     AttnProcessor2_0,
+    FusedAttnProcessor2_0,
     LoRAAttnProcessor2_0,
     LoRAXFormersAttnProcessor,
     XFormersAttnProcessor,
@@ -125,11 +126,11 @@ class StableDiffusionXLInstructPix2PixPipeline(
     This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the
     library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.)
 
-    In addition the pipeline inherits the following loading methods:
-        - *LoRA*: [`loaders.StableDiffusionXLLoraLoaderMixin.load_lora_weights`]
-
-    as well as the following saving methods:
-        - *LoRA*: [`loaders.StableDiffusionXLLoraLoaderMixin.save_lora_weights`]
+    The pipeline also inherits the following loading methods:
+        - [`~loaders.TextualInversionLoaderMixin.load_textual_inversion`] for loading textual inversion embeddings
+        - [`~loaders.FromSingleFileMixin.from_single_file`] for loading `.ckpt` files
+        - [`~loaders.StableDiffusionXLLoraLoaderMixin.load_lora_weights`] for loading LoRA weights
+        - [`~loaders.StableDiffusionXLLoraLoaderMixin.save_lora_weights`] for saving LoRA weights
 
     Args:
         vae ([`AutoencoderKL`]):
@@ -610,6 +611,7 @@ class StableDiffusionXLInstructPix2PixPipeline(
                 XFormersAttnProcessor,
                 LoRAXFormersAttnProcessor,
                 LoRAAttnProcessor2_0,
+                FusedAttnProcessor2_0,
             ),
         )
         # if xformers or torch_2_0 is used attention block does not need

diffusers/pipelines/stable_video_diffusion/pipeline_stable_video_diffusion.py

@@ -25,7 +25,7 @@ from ...image_processor import VaeImageProcessor
 from ...models import AutoencoderKLTemporalDecoder, UNetSpatioTemporalConditionModel
 from ...schedulers import EulerDiscreteScheduler
 from ...utils import BaseOutput, logging
-from ...utils.torch_utils import randn_tensor
+from ...utils.torch_utils import is_compiled_module, randn_tensor
 from ..pipeline_utils import DiffusionPipeline
 
 
@@ -211,7 +211,8 @@ class StableVideoDiffusionPipeline(DiffusionPipeline):
 
         latents = 1 / self.vae.config.scaling_factor * latents
 
-        accepts_num_frames = "num_frames" in set(inspect.signature(self.vae.forward).parameters.keys())
+        forward_vae_fn = self.vae._orig_mod.forward if is_compiled_module(self.vae) else self.vae.forward
+        accepts_num_frames = "num_frames" in set(inspect.signature(forward_vae_fn).parameters.keys())
 
         # decode decode_chunk_size frames at a time to avoid OOM
         frames = []
@@ -290,7 +291,9 @@ class StableVideoDiffusionPipeline(DiffusionPipeline):
     # corresponds to doing no classifier free guidance.
     @property
     def do_classifier_free_guidance(self):
-        return self._guidance_scale > 1 and self.unet.config.time_cond_proj_dim is None
+        if isinstance(self.guidance_scale, (int, float)):
+            return self.guidance_scale
+        return self.guidance_scale.max() > 1
 
     @property
     def num_timesteps(self):
@@ -415,10 +418,10 @@ class StableVideoDiffusionPipeline(DiffusionPipeline):
         # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
         # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
         # corresponds to doing no classifier free guidance.
-        do_classifier_free_guidance = max_guidance_scale > 1.0
+        self._guidance_scale = max_guidance_scale
 
         # 3. Encode input image
-        image_embeddings = self._encode_image(image, device, num_videos_per_prompt, do_classifier_free_guidance)
+        image_embeddings = self._encode_image(image, device, num_videos_per_prompt, self.do_classifier_free_guidance)
 
         # NOTE: Stable Diffusion Video was conditioned on fps - 1, which
         # is why it is reduced here.
@@ -434,7 +437,7 @@ class StableVideoDiffusionPipeline(DiffusionPipeline):
         if needs_upcasting:
             self.vae.to(dtype=torch.float32)
 
-        image_latents = self._encode_vae_image(image, device, num_videos_per_prompt, do_classifier_free_guidance)
+        image_latents = self._encode_vae_image(image, device, num_videos_per_prompt, self.do_classifier_free_guidance)
         image_latents = image_latents.to(image_embeddings.dtype)
 
         # cast back to fp16 if needed
@@ -453,7 +456,7 @@ class StableVideoDiffusionPipeline(DiffusionPipeline):
             image_embeddings.dtype,
             batch_size,
             num_videos_per_prompt,
-            do_classifier_free_guidance,
+            self.do_classifier_free_guidance,
         )
         added_time_ids = added_time_ids.to(device)
 
@@ -489,7 +492,7 @@ class StableVideoDiffusionPipeline(DiffusionPipeline):
         with self.progress_bar(total=num_inference_steps) as progress_bar:
             for i, t in enumerate(timesteps):
                 # expand the latents if we are doing classifier free guidance
-                latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents
+                latent_model_input = torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents
                 latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
 
                 # Concatenate image_latents over channels dimention
@@ -505,7 +508,7 @@ class StableVideoDiffusionPipeline(DiffusionPipeline):
                 )[0]
 
                 # perform guidance
-                if do_classifier_free_guidance:
+                if self.do_classifier_free_guidance:
                     noise_pred_uncond, noise_pred_cond = noise_pred.chunk(2)
                     noise_pred = noise_pred_uncond + self.guidance_scale * (noise_pred_cond - noise_pred_uncond)
 

diffusers/pipelines/t2i_adapter/pipeline_stable_diffusion_xl_adapter.py

@@ -178,6 +178,12 @@ class StableDiffusionXLAdapterPipeline(
     This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the
     library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.)
 
+    The pipeline also inherits the following loading methods:
+        - [`~loaders.TextualInversionLoaderMixin.load_textual_inversion`] for loading textual inversion embeddings
+        - [`~loaders.FromSingleFileMixin.from_single_file`] for loading `.ckpt` files
+        - [`~loaders.StableDiffusionXLLoraLoaderMixin.load_lora_weights`] for loading LoRA weights
+        - [`~loaders.StableDiffusionXLLoraLoaderMixin.save_lora_weights`] for saving LoRA weights
+
     Args:
         adapter ([`T2IAdapter`] or [`MultiAdapter`] or `List[T2IAdapter]`):
             Provides additional conditioning to the unet during the denoising process. If you set multiple Adapter as a

diffusers/pipelines/text_to_video_synthesis/pipeline_text_to_video_synth.py

@@ -83,6 +83,11 @@ class TextToVideoSDPipeline(DiffusionPipeline, TextualInversionLoaderMixin, Lora
     This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods
     implemented for all pipelines (downloading, saving, running on a particular device, etc.).
 
+    The pipeline also inherits the following loading methods:
+        - [`~loaders.TextualInversionLoaderMixin.load_textual_inversion`] for loading textual inversion embeddings
+        - [`~loaders.LoraLoaderMixin.load_lora_weights`] for loading LoRA weights
+        - [`~loaders.LoraLoaderMixin.save_lora_weights`] for saving LoRA weights
+
     Args:
         vae ([`AutoencoderKL`]):
             Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations.

diffusers/pipelines/text_to_video_synthesis/pipeline_text_to_video_synth_img2img.py

@@ -159,6 +159,11 @@ class VideoToVideoSDPipeline(DiffusionPipeline, TextualInversionLoaderMixin, Lor
     This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods
     implemented for all pipelines (downloading, saving, running on a particular device, etc.).
 
+    The pipeline also inherits the following loading methods:
+        - [`~loaders.TextualInversionLoaderMixin.load_textual_inversion`] for loading textual inversion embeddings
+        - [`~loaders.LoraLoaderMixin.load_lora_weights`] for loading LoRA weights
+        - [`~loaders.LoraLoaderMixin.save_lora_weights`] for saving LoRA weights
+
     Args:
         vae ([`AutoencoderKL`]):
             Variational Auto-Encoder (VAE) Model to encode and decode videos to and from latent representations.