[Refactor] FreeInit for AnimateDiff based pipelines (huggingface#6874)

* update

* update

* update

* update

* update

* update

* update

* update

* update

* update

src/diffusers/pipelines/animatediff/pipeline_animatediff_video2video.py

@@ -34,6 +34,7 @@
 )
 from ...utils import USE_PEFT_BACKEND, logging, scale_lora_layers, unscale_lora_layers
 from ...utils.torch_utils import randn_tensor
+from ..free_init_utils import FreeInitMixin
 from ..pipeline_utils import DiffusionPipeline
 from .pipeline_output import AnimateDiffPipelineOutput
 
@@ -163,7 +164,9 @@ def retrieve_timesteps(
     return timesteps, num_inference_steps
 
 
-class AnimateDiffVideoToVideoPipeline(DiffusionPipeline, TextualInversionLoaderMixin, IPAdapterMixin, LoraLoaderMixin):
+class AnimateDiffVideoToVideoPipeline(
+    DiffusionPipeline, TextualInversionLoaderMixin, IPAdapterMixin, LoraLoaderMixin, FreeInitMixin
+):
     r"""
     Pipeline for video-to-video generation.
 
@@ -193,7 +196,7 @@ class AnimateDiffVideoToVideoPipeline(DiffusionPipeline, TextualInversionLoaderM
     """
 
     model_cpu_offload_seq = "text_encoder->image_encoder->unet->vae"
-    _optional_components = ["feature_extractor", "image_encoder"]
+    _optional_components = ["feature_extractor", "image_encoder", "motion_adapter"]
     _callback_tensor_inputs = ["latents", "prompt_embeds", "negative_prompt_embeds"]
 
     def __init__(
@@ -215,7 +218,8 @@ def __init__(
         image_encoder: CLIPVisionModelWithProjection = None,
     ):
         super().__init__()
-        unet = UNetMotionModel.from_unet2d(unet, motion_adapter)
+        if isinstance(unet, UNet2DConditionModel):
+            unet = UNetMotionModel.from_unet2d(unet, motion_adapter)
 
         self.register_modules(
             vae=vae,
@@ -584,12 +588,12 @@ def check_inputs(
         if video is not None and latents is not None:
             raise ValueError("Only one of `video` or `latents` should be provided")
 
-    def get_timesteps(self, num_inference_steps, strength, device):
+    def get_timesteps(self, num_inference_steps, timesteps, strength, device):
         # get the original timestep using init_timestep
         init_timestep = min(int(num_inference_steps * strength), num_inference_steps)
 
         t_start = max(num_inference_steps - init_timestep, 0)
-        timesteps = self.scheduler.timesteps[t_start * self.scheduler.order :]
+        timesteps = timesteps[t_start * self.scheduler.order :]
 
         return timesteps, num_inference_steps - t_start
 
@@ -876,9 +880,8 @@ def __call__(
 
         # 4. Prepare timesteps
         timesteps, num_inference_steps = retrieve_timesteps(self.scheduler, num_inference_steps, device, timesteps)
-        timesteps, num_inference_steps = self.get_timesteps(num_inference_steps, strength, device)
+        timesteps, num_inference_steps = self.get_timesteps(num_inference_steps, timesteps, strength, device)
         latent_timestep = timesteps[:1].repeat(batch_size * num_videos_per_prompt)
-        self._num_timesteps = len(timesteps)
 
         # 5. Prepare latent variables
         num_channels_latents = self.unet.config.in_channels
@@ -901,42 +904,55 @@ def __call__(
         # 7. Add image embeds for IP-Adapter
         added_cond_kwargs = {"image_embeds": image_embeds} if ip_adapter_image is not None else None
 
-        # 8. Denoising loop
-        num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
-        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 self.do_classifier_free_guidance else latents
-                latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
-
-                # predict the noise residual
-                noise_pred = self.unet(
-                    latent_model_input,
-                    t,
-                    encoder_hidden_states=prompt_embeds,
-                    cross_attention_kwargs=self.cross_attention_kwargs,
-                    added_cond_kwargs=added_cond_kwargs,
-                ).sample
-
-                # perform guidance
-                if self.do_classifier_free_guidance:
-                    noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
-                    noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
-
-                # compute the previous noisy sample x_t -> x_t-1
-                latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs).prev_sample
-
-                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)
-
-                progress_bar.update()
+        num_free_init_iters = self._free_init_num_iters if self.free_init_enabled else 1
+        for free_init_iter in range(num_free_init_iters):
+            if self.free_init_enabled:
+                latents, timesteps = self._apply_free_init(
+                    latents, free_init_iter, num_inference_steps, device, latents.dtype, generator
+                )
+                num_inference_steps = len(timesteps)
+                # make sure to readjust timesteps based on strength
+                timesteps, num_inference_steps = self.get_timesteps(num_inference_steps, timesteps, strength, device)
+
+            self._num_timesteps = len(timesteps)
+            num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
+            # 8. Denoising loop
+            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 self.do_classifier_free_guidance else latents
+                    latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
+
+                    # predict the noise residual
+                    noise_pred = self.unet(
+                        latent_model_input,
+                        t,
+                        encoder_hidden_states=prompt_embeds,
+                        cross_attention_kwargs=self.cross_attention_kwargs,
+                        added_cond_kwargs=added_cond_kwargs,
+                    ).sample
+
+                    # perform guidance
+                    if self.do_classifier_free_guidance:
+                        noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
+                        noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
+
+                    # compute the previous noisy sample x_t -> x_t-1
+                    latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs).prev_sample
+
+                    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 output_type == "latent":
             return AnimateDiffPipelineOutput(frames=latents)

src/diffusers/pipelines/free_init_utils.py

@@ -0,0 +1,184 @@
+# Copyright 2024 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 math
+from typing import Tuple, Union
+
+import torch
+import torch.fft as fft
+
+from ..utils.torch_utils import randn_tensor
+
+
+class FreeInitMixin:
+    r"""Mixin class for FreeInit."""
+
+    def enable_free_init(
+        self,
+        num_iters: int = 3,
+        use_fast_sampling: bool = False,
+        method: str = "butterworth",
+        order: int = 4,
+        spatial_stop_frequency: float = 0.25,
+        temporal_stop_frequency: float = 0.25,
+    ):
+        """Enables the FreeInit mechanism as in https://arxiv.org/abs/2312.07537.
+
+        This implementation has been adapted from the [official repository](https://github.com/TianxingWu/FreeInit).
+
+        Args:
+            num_iters (`int`, *optional*, defaults to `3`):
+                Number of FreeInit noise re-initialization iterations.
+            use_fast_sampling (`bool`, *optional*, defaults to `False`):
+                Whether or not to speedup sampling procedure at the cost of probably lower quality results. Enables
+                the "Coarse-to-Fine Sampling" strategy, as mentioned in the paper, if set to `True`.
+            method (`str`, *optional*, defaults to `butterworth`):
+                Must be one of `butterworth`, `ideal` or `gaussian` to use as the filtering method for the
+                FreeInit low pass filter.
+            order (`int`, *optional*, defaults to `4`):
+                Order of the filter used in `butterworth` method. Larger values lead to `ideal` method behaviour
+                whereas lower values lead to `gaussian` method behaviour.
+            spatial_stop_frequency (`float`, *optional*, defaults to `0.25`):
+                Normalized stop frequency for spatial dimensions. Must be between 0 to 1. Referred to as `d_s` in
+                the original implementation.
+            temporal_stop_frequency (`float`, *optional*, defaults to `0.25`):
+                Normalized stop frequency for temporal dimensions. Must be between 0 to 1. Referred to as `d_t` in
+                the original implementation.
+        """
+        self._free_init_num_iters = num_iters
+        self._free_init_use_fast_sampling = use_fast_sampling
+        self._free_init_method = method
+        self._free_init_order = order
+        self._free_init_spatial_stop_frequency = spatial_stop_frequency
+        self._free_init_temporal_stop_frequency = temporal_stop_frequency
+
+    def disable_free_init(self):
+        """Disables the FreeInit mechanism if enabled."""
+        self._free_init_num_iters = None
+
+    @property
+    def free_init_enabled(self):
+        return hasattr(self, "_free_init_num_iters") and self._free_init_num_iters is not None
+
+    def _get_free_init_freq_filter(
+        self,
+        shape: Tuple[int, ...],
+        device: Union[str, torch.dtype],
+        filter_type: str,
+        order: float,
+        spatial_stop_frequency: float,
+        temporal_stop_frequency: float,
+    ) -> torch.Tensor:
+        r"""Returns the FreeInit filter based on filter type and other input conditions."""
+
+        time, height, width = shape[-3], shape[-2], shape[-1]
+        mask = torch.zeros(shape)
+
+        if spatial_stop_frequency == 0 or temporal_stop_frequency == 0:
+            return mask
+
+        if filter_type == "butterworth":
+
+            def retrieve_mask(x):
+                return 1 / (1 + (x / spatial_stop_frequency**2) ** order)
+        elif filter_type == "gaussian":
+
+            def retrieve_mask(x):
+                return math.exp(-1 / (2 * spatial_stop_frequency**2) * x)
+        elif filter_type == "ideal":
+
+            def retrieve_mask(x):
+                return 1 if x <= spatial_stop_frequency * 2 else 0
+        else:
+            raise NotImplementedError("`filter_type` must be one of gaussian, butterworth or ideal")
+
+        for t in range(time):
+            for h in range(height):
+                for w in range(width):
+                    d_square = (
+                        ((spatial_stop_frequency / temporal_stop_frequency) * (2 * t / time - 1)) ** 2
+                        + (2 * h / height - 1) ** 2
+                        + (2 * w / width - 1) ** 2
+                    )
+                    mask[..., t, h, w] = retrieve_mask(d_square)
+
+        return mask.to(device)
+
+    def _apply_freq_filter(self, x: torch.Tensor, noise: torch.Tensor, low_pass_filter: torch.Tensor) -> torch.Tensor:
+        r"""Noise reinitialization."""
+        # FFT
+        x_freq = fft.fftn(x, dim=(-3, -2, -1))
+        x_freq = fft.fftshift(x_freq, dim=(-3, -2, -1))
+        noise_freq = fft.fftn(noise, dim=(-3, -2, -1))
+        noise_freq = fft.fftshift(noise_freq, dim=(-3, -2, -1))
+
+        # frequency mix
+        high_pass_filter = 1 - low_pass_filter
+        x_freq_low = x_freq * low_pass_filter
+        noise_freq_high = noise_freq * high_pass_filter
+        x_freq_mixed = x_freq_low + noise_freq_high  # mix in freq domain
+
+        # IFFT
+        x_freq_mixed = fft.ifftshift(x_freq_mixed, dim=(-3, -2, -1))
+        x_mixed = fft.ifftn(x_freq_mixed, dim=(-3, -2, -1)).real
+
+        return x_mixed
+
+    def _apply_free_init(
+        self,
+        latents: torch.Tensor,
+        free_init_iteration: int,
+        num_inference_steps: int,
+        device: torch.device,
+        dtype: torch.dtype,
+        generator: torch.Generator,
+    ):
+        if free_init_iteration == 0:
+            self._free_init_initial_noise = latents.detach().clone()
+            return latents, self.scheduler.timesteps
+
+        latent_shape = latents.shape
+
+        free_init_filter_shape = (1, *latent_shape[1:])
+        free_init_freq_filter = self._get_free_init_freq_filter(
+            shape=free_init_filter_shape,
+            device=device,
+            filter_type=self._free_init_method,
+            order=self._free_init_order,
+            spatial_stop_frequency=self._free_init_spatial_stop_frequency,
+            temporal_stop_frequency=self._free_init_temporal_stop_frequency,
+        )
+
+        current_diffuse_timestep = self.scheduler.config.num_train_timesteps - 1
+        diffuse_timesteps = torch.full((latent_shape[0],), current_diffuse_timestep).long()
+
+        z_t = self.scheduler.add_noise(
+            original_samples=latents, noise=self._free_init_initial_noise, timesteps=diffuse_timesteps.to(device)
+        ).to(dtype=torch.float32)
+
+        z_rand = randn_tensor(
+            shape=latent_shape,
+            generator=generator,
+            device=device,
+            dtype=torch.float32,
+        )
+        latents = self._apply_freq_filter(z_t, z_rand, low_pass_filter=free_init_freq_filter)
+        latents = latents.to(dtype)
+
+        # Coarse-to-Fine Sampling for faster inference (can lead to lower quality)
+        if self._free_init_use_fast_sampling:
+            num_inference_steps = int(num_inference_steps / self._free_init_num_iters * (free_init_iteration + 1))
+            self.scheduler.set_timesteps(num_inference_steps, device=device)
+
+        return latents, self.scheduler.timesteps