Longer videos and textual inversions and fp16 autocast

README.md

@@ -162,6 +162,43 @@ Here we demonstrate several best results we found in our experiments.
 </table>
 <p style="margin-left: 2em; margin-top: -1em">Model：<a href="https://civitai.com/models/33208/filmgirl-film-grain-lora-and-loha">FilmVelvia</a></p>
 
+### Longer generations
+You can also generate longer animations by using overlapping sliding windows.
+```
+python -m scripts.animate --config configs/prompts/{your_config}.yaml --L 64 --context_length 16 
+```
+##### Sliding window related parameters:
+
+```L``` - the length of the generated animation.
+
+```context_length``` - the length of the sliding window (limited by motion modules capacity), default to ```L```.
+
+```context_overlap``` - how much neighbouring contexts overlap. By default ```context_length``` / 2
+
+```context_stride``` - (2^```context_stride```) is a max stride between 2 neighbour frames. By default 0
+
+##### Extended this way gallery examples
+
+<table class="center">
+    <tr>
+    <td><img src="__assets__/animations/model_01_4x/01.gif"></td>
+    <td><img src="__assets__/animations/model_01_4x/02.gif"></td>
+    <td><img src="__assets__/animations/model_01_4x/03.gif"></td>
+    <td><img src="__assets__/animations/model_01_4x/04.gif"></td>
+    </tr>
+</table>
+<p style="margin-left: 2em; margin-top: -1em">Model：<a href="https://civitai.com/models/30240/toonyou">ToonYou</a></p>
+
+<table>
+    <tr>
+    <td><img src="__assets__/animations/model_03_4x/01.gif"></td>
+    <td><img src="__assets__/animations/model_03_4x/02.gif"></td>
+    <td><img src="__assets__/animations/model_03_4x/03.gif"></td>
+    <td><img src="__assets__/animations/model_03_4x/04.gif"></td>
+    </tr>
+</table>
+<p style="margin-left: 2em; margin-top: -1em">Model：<a href="https://civitai.com/models/4201/realistic-vision-v20">Realistic Vision V2.0</a></p>
+
 #### Community Cases
 Here are some samples contributed by the community artists. Create a Pull Request if you would like to show your results here😚.
 

animatediff/pipelines/pipeline_animation.py

@@ -1,11 +1,13 @@
 # Adapted from https://github.com/showlab/Tune-A-Video/blob/main/tuneavideo/pipelines/pipeline_tuneavideo.py
 
 import inspect
+import os
 from typing import Callable, List, Optional, Union
 from dataclasses import dataclass
 
 import numpy as np
 import torch
+from torch import nn
 from tqdm import tqdm
 
 from diffusers.utils import is_accelerate_available
@@ -29,6 +31,8 @@
 
 from ..models.unet import UNet3DConditionModel
 
+from ..utils import overlap_policy
+from ..utils.path import get_absolute_path
 
 logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
 
@@ -55,6 +59,7 @@ def __init__(
             EulerAncestralDiscreteScheduler,
             DPMSolverMultistepScheduler,
         ],
+        scan_inversions: bool = True,
     ):
         super().__init__()
 
@@ -114,6 +119,36 @@ def __init__(
             scheduler=scheduler,
         )
         self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
+        self.embeddings_dir = get_absolute_path('models', 'embeddings')
+        self.embeddings_dict = {}
+        self.default_tokens = len(self.tokenizer)
+        self.scan_inversions = scan_inversions
+
+    def update_embeddings(self):
+        if not self.scan_inversions:
+            return
+        names = [p for p in os.listdir(self.embeddings_dir) if p.endswith('.pt')]
+        weight = self.text_encoder.text_model.embeddings.token_embedding.weight
+        added_embeddings = []
+        for name in names:
+            embedding_path = os.path.join(self.embeddings_dir, name)
+            embedding = torch.load(embedding_path)
+            key = os.path.splitext(name)[0]
+            if key in self.tokenizer.encoder:
+                idx = self.tokenizer.encoder[key]
+            else:
+                idx = len(self.tokenizer)
+                self.tokenizer.add_tokens([key])
+            embedding = embedding['string_to_param']['*']
+            if idx not in self.embeddings_dict:
+                added_embeddings.append(name)
+                self.embeddings_dict[idx] = torch.arange(weight.shape[0], weight.shape[0] + embedding.shape[0])
+                weight = torch.cat([weight, embedding.to(weight.device, weight.dtype)], dim=0)
+                self.tokenizer.add_tokens([key])
+        if added_embeddings:
+            self.text_encoder.text_model.embeddings.token_embedding = nn.Embedding(
+                weight.shape[0], weight.shape[1], _weight=weight)
+            logger.info(f'Added {len(added_embeddings)} embeddings: {added_embeddings}')
 
     def enable_vae_slicing(self):
         self.vae.enable_slicing()
@@ -147,9 +182,32 @@ def _execution_device(self):
                 return torch.device(module._hf_hook.execution_device)
         return self.device
 
+    def insert_inversions(self, ids, attention_mask):
+        larger = ids >= self.default_tokens
+        for idx in reversed(torch.where(larger)[1]):
+            ids = torch.cat([
+                ids[:, :idx],
+                self.embeddings_dict[ids[:, idx].item()].unsqueeze(0),
+                ids[:, idx + 1:],
+            ], 1)
+            if attention_mask is not None:
+                attention_mask = torch.cat([
+                    attention_mask[:, :idx],
+                    torch.ones(1, 1, dtype=attention_mask.dtype, device=attention_mask.device),
+                    attention_mask[:, idx + 1:],
+                ], 1)
+        if ids.shape[1] > self.tokenizer.model_max_length:
+            logger.warning(f"After inserting inversions, the sequence length is larger than the max length. Cutting off"
+                           f" {ids.shape[1] - self.tokenizer.model_max_length} tokens.")
+            ids = torch.cat([ids[:, :self.tokenizer.model_max_length - 1], ids[:, -1:]], 1)
+            if attention_mask is not None:
+                attention_mask = attention_mask[:, :self.tokenizer.model_max_length]
+        return ids, attention_mask
+
     def _encode_prompt(self, prompt, device, num_videos_per_prompt, do_classifier_free_guidance, negative_prompt):
         batch_size = len(prompt) if isinstance(prompt, list) else 1
 
+        self.update_embeddings()
         text_inputs = self.tokenizer(
             prompt,
             padding="max_length",
@@ -172,6 +230,7 @@ def _encode_prompt(self, prompt, device, num_videos_per_prompt, do_classifier_fr
         else:
             attention_mask = None
 
+        text_input_ids, attention_mask = self.insert_inversions(text_input_ids, attention_mask)
         text_embeddings = self.text_encoder(
             text_input_ids.to(device),
             attention_mask=attention_mask,
@@ -218,8 +277,10 @@ def _encode_prompt(self, prompt, device, num_videos_per_prompt, do_classifier_fr
             else:
                 attention_mask = None
 
+            uncond_input_ids = uncond_input.input_ids
+            uncond_input_ids, attention_mask = self.insert_inversions(uncond_input_ids, attention_mask)
             uncond_embeddings = self.text_encoder(
-                uncond_input.input_ids.to(device),
+                uncond_input_ids.to(device),
                 attention_mask=attention_mask,
             )
             uncond_embeddings = uncond_embeddings[0]
@@ -242,8 +303,9 @@ def decode_latents(self, latents):
         latents = rearrange(latents, "b c f h w -> (b f) c h w")
         # video = self.vae.decode(latents).sample
         video = []
+        device = self._execution_device
         for frame_idx in tqdm(range(latents.shape[0])):
-            video.append(self.vae.decode(latents[frame_idx:frame_idx+1]).sample)
+            video.append(self.vae.decode(latents[frame_idx:frame_idx+1].to(device)).sample)
         video = torch.cat(video)
         video = rearrange(video, "(b f) c h w -> b c f h w", f=video_length)
         video = (video / 2 + 0.5).clamp(0, 1)
@@ -317,6 +379,9 @@ def __call__(
         self,
         prompt: Union[str, List[str]],
         video_length: Optional[int],
+        temporal_context: Optional[int] = None,
+        strides: int = 3,
+        overlap: int = 4,
         height: Optional[int] = None,
         width: Optional[int] = None,
         num_inference_steps: int = 50,
@@ -330,6 +395,8 @@ def __call__(
         return_dict: bool = True,
         callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
         callback_steps: Optional[int] = 1,
+        seq_policy=overlap_policy.uniform,
+        fp16=False,
         **kwargs,
     ):
         # Default height and width to unet
@@ -348,6 +415,7 @@ def __call__(
             batch_size = len(prompt)
 
         device = self._execution_device
+        cpu = torch.device('cpu')
         # 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.
@@ -356,7 +424,7 @@ def __call__(
         # Encode input prompt
         prompt = prompt if isinstance(prompt, list) else [prompt] * batch_size
         if negative_prompt is not None:
-            negative_prompt = negative_prompt if isinstance(negative_prompt, list) else [negative_prompt] * batch_size 
+            negative_prompt = negative_prompt if isinstance(negative_prompt, list) else [negative_prompt] * batch_size
         text_embeddings = self._encode_prompt(
             prompt, device, num_videos_per_prompt, do_classifier_free_guidance, negative_prompt
         )
@@ -373,45 +441,49 @@ def __call__(
             video_length,
             height,
             width,
-            text_embeddings.dtype,
-            device,
+            torch.float32,
+            cpu,  # using cpu to store latents allows generated frame amount not to be limited by vram but by ram
             generator,
             latents,
         )
         latents_dtype = latents.dtype
 
         # Prepare extra step kwargs.
         extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
-
+        total = sum(
+            len(list(seq_policy(i, num_inference_steps, latents.shape[2], temporal_context, strides, overlap)))
+            for i in range(len(timesteps))
+        )
         # Denoising loop
         num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
-        with self.progress_bar(total=num_inference_steps) as progress_bar:
+        with self.progress_bar(total=total) 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 = self.scheduler.scale_model_input(latent_model_input, t)
-
-                # predict the noise residual
-                noise_pred = self.unet(latent_model_input, t, encoder_hidden_states=text_embeddings).sample.to(dtype=latents_dtype)
-                # noise_pred = []
-                # import pdb
-                # pdb.set_trace()
-                # for batch_idx in range(latent_model_input.shape[0]):
-                #     noise_pred_single = self.unet(latent_model_input[batch_idx:batch_idx+1], t, encoder_hidden_states=text_embeddings[batch_idx:batch_idx+1]).sample.to(dtype=latents_dtype)
-                #     noise_pred.append(noise_pred_single)
-                # noise_pred = torch.cat(noise_pred)
+                noise_pred = torch.zeros((latents.shape[0] * (2 if do_classifier_free_guidance else 1),
+                                          *latents.shape[1:]), device=latents.device, dtype=latents_dtype)
+                counter = torch.zeros((1, 1, latents.shape[2], 1, 1), device=latents.device, dtype=latents_dtype)
+                for seq in seq_policy(i, num_inference_steps, latents.shape[2], temporal_context, strides, overlap):
+                    # expand the latents if we are doing classifier free guidance
+                    latent_model_input = latents[:, :, seq].to(device)\
+                        .repeat(2 if do_classifier_free_guidance else 1, 1, 1, 1, 1)
+                    latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
+
+                    # predict the noise residual
+                    with torch.autocast('cuda', enabled=fp16, dtype=torch.float16):
+                        pred = self.unet(latent_model_input, t, encoder_hidden_states=text_embeddings)
+                    noise_pred[:, :, seq] += pred.sample.to(dtype=latents_dtype, device=cpu)
+                    counter[:, :, seq] += 1
+                    progress_bar.update()
 
                 # perform guidance
                 if do_classifier_free_guidance:
-                    noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
+                    noise_pred_uncond, noise_pred_text = (noise_pred / counter).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
 
                 # 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 callback is not None and i % callback_steps == 0:
                         callback(i, t, latents)
 

animatediff/utils/overlap_policy.py

@@ -0,0 +1,20 @@
+import numpy as np
+
+
+def ordered_halving(i):
+    return int('{:064b}'.format(i)[::-1], 2) / (1 << 64)
+
+
+def uniform(step, steps, n, context_size, strides, overlap, closed_loop=True):
+    if n <= context_size:
+        yield list(range(n))
+        return
+    strides = min(strides, int(np.ceil(np.log2(n / context_size))) + 1)
+    for stride in 1 << np.arange(strides):
+        pad = int(round(n * ordered_halving(step)))
+        for j in range(
+                int(ordered_halving(step) * stride) + pad,
+                n + pad + (0 if closed_loop else -overlap),
+                (context_size * stride - overlap)
+        ):
+            yield [e % n for e in range(j, j + context_size * stride, stride)]

animatediff/utils/path.py

@@ -0,0 +1,16 @@
+import os
+
+project_path = os.path.dirname(os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
+
+
+def get_absolute_path(*relative):
+    if relative[0].startswith('/'):
+        return os.path.join(*relative)  # absolute path
+    return os.path.join(project_path, *relative)
+
+
+if __name__ == '__main__':
+    print(get_absolute_path('test'))
+    print(get_absolute_path('/test'))
+    print(get_absolute_path('test', 'test'))
+    print(get_absolute_path('/test', 'test'))

scripts/animate.py

@@ -30,7 +30,12 @@
 def main(args):
     *_, func_args = inspect.getargvalues(inspect.currentframe())
     func_args = dict(func_args)
-
+
+    if args.context_length == 0:
+        args.context_length = args.L
+    if args.context_overlap == -1:
+        args.context_overlap = args.context_length // 2
+
     time_str = datetime.datetime.now().strftime("%Y-%m-%dT%H-%M-%S")
     savedir = f"samples/{Path(args.config).stem}-{time_str}"
     os.makedirs(savedir)
@@ -58,6 +63,7 @@ def main(args):
             pipeline = AnimationPipeline(
                 vae=vae, text_encoder=text_encoder, tokenizer=tokenizer, unet=unet,
                 scheduler=DDIMScheduler(**OmegaConf.to_container(inference_config.noise_scheduler_kwargs)),
+                scan_inversions=not args.disable_inversions,
             ).to("cuda")
 
             # 1. unet ckpt
@@ -130,6 +136,10 @@ def main(args):
                     width               = args.W,
                     height              = args.H,
                     video_length        = args.L,
+                    temporal_context    = args.context_length,
+                    strides             = args.context_stride + 1,
+                    overlap             = args.context_overlap,
+                    fp16                = not args.fp32,
                 ).videos
                 samples.append(sample)
 
@@ -150,7 +160,18 @@ def main(args):
     parser.add_argument("--pretrained_model_path", type=str, default="models/StableDiffusion/stable-diffusion-v1-5",)
     parser.add_argument("--inference_config",      type=str, default="configs/inference/inference.yaml")    
     parser.add_argument("--config",                type=str, required=True)
-
+
+    parser.add_argument("--fp32", action="store_true")
+    parser.add_argument("--disable_inversions", action="store_true",
+                        help="do not scan for downloaded textual inversions")
+
+    parser.add_argument("--context_length", type=int, default=0,
+                        help="temporal transformer context length (0 for same as -L)")
+    parser.add_argument("--context_stride", type=int, default=0,
+                        help="max stride of motion is 2^context_stride")
+    parser.add_argument("--context_overlap", type=int, default=-1,
+                        help="overlap between chunks of context (-1 for half of context length)")
+
     parser.add_argument("--L", type=int, default=16 )
     parser.add_argument("--W", type=int, default=512)
     parser.add_argument("--H", type=int, default=512)