AetherV1 / aether /pipelines /aetherv1_pipeline_cogvideox.py
Wenzheng Chang
update bloat16
8040e22
import inspect
import math
from dataclasses import dataclass
from typing import Dict, List, Optional, Tuple, Union
import numpy as np
import PIL
import torch
from diffusers import (
AutoencoderKLCogVideoX,
CogVideoXDPMScheduler,
CogVideoXImageToVideoPipeline,
CogVideoXTransformer3DModel,
)
from diffusers.image_processor import PipelineImageInput
from diffusers.models.embeddings import get_1d_rotary_pos_embed
from diffusers.utils import BaseOutput
from diffusers.utils.torch_utils import randn_tensor
from einops import rearrange
from transformers import AutoTokenizer, T5EncoderModel
from aether.utils.preprocess_utils import imcrop_center
def get_3d_rotary_pos_embed(
embed_dim,
crops_coords,
grid_size,
temporal_size,
theta: int = 10000,
use_real: bool = True,
grid_type: str = "linspace",
max_size: Optional[Tuple[int, int]] = None,
device: Optional[torch.device] = None,
fps_factor: Optional[float] = 1.0,
) -> Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]:
"""
RoPE for video tokens with 3D structure.
Args:
embed_dim: (`int`):
The embedding dimension size, corresponding to hidden_size_head.
crops_coords (`Tuple[int]`):
The top-left and bottom-right coordinates of the crop.
grid_size (`Tuple[int]`):
The grid size of the spatial positional embedding (height, width).
temporal_size (`int`):
The size of the temporal dimension.
theta (`float`):
Scaling factor for frequency computation.
grid_type (`str`):
Whether to use "linspace" or "slice" to compute grids.
fps_factor (`float`):
The relative fps factor of the video, computed by base_fps / fps. Useful for variable fps training.
Returns:
`torch.Tensor`: positional embedding with shape `(temporal_size * grid_size[0] * grid_size[1], embed_dim/2)`.
"""
if use_real is not True:
raise ValueError(
" `use_real = False` is not currently supported for get_3d_rotary_pos_embed"
)
if grid_type == "linspace":
start, stop = crops_coords
grid_size_h, grid_size_w = grid_size
grid_h = torch.linspace(
start[0],
stop[0] * (grid_size_h - 1) / grid_size_h,
grid_size_h,
device=device,
dtype=torch.float32,
)
grid_w = torch.linspace(
start[1],
stop[1] * (grid_size_w - 1) / grid_size_w,
grid_size_w,
device=device,
dtype=torch.float32,
)
grid_t = (
torch.linspace(
0,
temporal_size * (temporal_size - 1) / temporal_size,
temporal_size,
device=device,
dtype=torch.float32,
)
* fps_factor
)
elif grid_type == "slice":
max_h, max_w = max_size
grid_size_h, grid_size_w = grid_size
grid_h = torch.arange(max_h, device=device, dtype=torch.float32)
grid_w = torch.arange(max_w, device=device, dtype=torch.float32)
grid_t = (
torch.arange(temporal_size, device=device, dtype=torch.float32) * fps_factor
)
else:
raise ValueError("Invalid value passed for `grid_type`.")
# Compute dimensions for each axis
dim_t = embed_dim // 4
dim_h = embed_dim // 8 * 3
dim_w = embed_dim // 8 * 3
# Temporal frequencies
freqs_t = get_1d_rotary_pos_embed(dim_t, grid_t, theta=theta, use_real=True)
# Spatial frequencies for height and width
freqs_h = get_1d_rotary_pos_embed(dim_h, grid_h, theta=theta, use_real=True)
freqs_w = get_1d_rotary_pos_embed(dim_w, grid_w, theta=theta, use_real=True)
# BroadCast and concatenate temporal and spaial frequencie (height and width) into a 3d tensor
def combine_time_height_width(freqs_t, freqs_h, freqs_w):
freqs_t = freqs_t[:, None, None, :].expand(
-1, grid_size_h, grid_size_w, -1
) # temporal_size, grid_size_h, grid_size_w, dim_t
freqs_h = freqs_h[None, :, None, :].expand(
temporal_size, -1, grid_size_w, -1
) # temporal_size, grid_size_h, grid_size_2, dim_h
freqs_w = freqs_w[None, None, :, :].expand(
temporal_size, grid_size_h, -1, -1
) # temporal_size, grid_size_h, grid_size_2, dim_w
freqs = torch.cat(
[freqs_t, freqs_h, freqs_w], dim=-1
) # temporal_size, grid_size_h, grid_size_w, (dim_t + dim_h + dim_w)
freqs = freqs.view(
temporal_size * grid_size_h * grid_size_w, -1
) # (temporal_size * grid_size_h * grid_size_w), (dim_t + dim_h + dim_w)
return freqs
t_cos, t_sin = freqs_t # both t_cos and t_sin has shape: temporal_size, dim_t
h_cos, h_sin = freqs_h # both h_cos and h_sin has shape: grid_size_h, dim_h
w_cos, w_sin = freqs_w # both w_cos and w_sin has shape: grid_size_w, dim_w
if grid_type == "slice":
t_cos, t_sin = t_cos[:temporal_size], t_sin[:temporal_size]
h_cos, h_sin = h_cos[:grid_size_h], h_sin[:grid_size_h]
w_cos, w_sin = w_cos[:grid_size_w], w_sin[:grid_size_w]
cos = combine_time_height_width(t_cos, h_cos, w_cos)
sin = combine_time_height_width(t_sin, h_sin, w_sin)
return cos, sin
# Similar to diffusers.pipelines.hunyuandit.pipeline_hunyuandit.get_resize_crop_region_for_grid
def get_resize_crop_region_for_grid(src, tgt_width, tgt_height):
tw = tgt_width
th = tgt_height
h, w = src
r = h / w
if r > (th / tw):
resize_height = th
resize_width = int(round(th / h * w))
else:
resize_width = tw
resize_height = int(round(tw / w * h))
crop_top = int(round((th - resize_height) / 2.0))
crop_left = int(round((tw - resize_width) / 2.0))
return (crop_top, crop_left), (crop_top + resize_height, crop_left + resize_width)
# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps
def retrieve_timesteps(
scheduler,
num_inference_steps: Optional[int] = None,
device: Optional[Union[str, torch.device]] = None,
timesteps: Optional[List[int]] = None,
sigmas: Optional[List[float]] = None,
**kwargs,
):
r"""
Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles
custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`.
Args:
scheduler (`SchedulerMixin`):
The scheduler to get timesteps from.
num_inference_steps (`int`):
The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps`
must be `None`.
device (`str` or `torch.device`, *optional*):
The device to which the timesteps should be moved to. If `None`, the timesteps are not moved.
timesteps (`List[int]`, *optional*):
Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed,
`num_inference_steps` and `sigmas` must be `None`.
sigmas (`List[float]`, *optional*):
Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed,
`num_inference_steps` and `timesteps` must be `None`.
Returns:
`Tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the
second element is the number of inference steps.
"""
if timesteps is not None and sigmas is not None:
raise ValueError(
"Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values"
)
if timesteps is not None:
accepts_timesteps = "timesteps" in set(
inspect.signature(scheduler.set_timesteps).parameters.keys()
)
if not accepts_timesteps:
raise ValueError(
f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom"
f" timestep schedules. Please check whether you are using the correct scheduler."
)
scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs)
timesteps = scheduler.timesteps
num_inference_steps = len(timesteps)
elif sigmas is not None:
accept_sigmas = "sigmas" in set(
inspect.signature(scheduler.set_timesteps).parameters.keys()
)
if not accept_sigmas:
raise ValueError(
f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom"
f" sigmas schedules. Please check whether you are using the correct scheduler."
)
scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs)
timesteps = scheduler.timesteps
num_inference_steps = len(timesteps)
else:
scheduler.set_timesteps(num_inference_steps, device=device, **kwargs)
timesteps = scheduler.timesteps
return timesteps, num_inference_steps
# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.retrieve_latents
def retrieve_latents(
encoder_output: torch.Tensor,
generator: Optional[torch.Generator] = None,
sample_mode: str = "sample",
):
if hasattr(encoder_output, "latent_dist") and sample_mode == "sample":
return encoder_output.latent_dist.sample(generator)
elif hasattr(encoder_output, "latent_dist") and sample_mode == "argmax":
return encoder_output.latent_dist.mode()
elif hasattr(encoder_output, "latents"):
return encoder_output.latents
else:
raise AttributeError("Could not access latents of provided encoder_output")
@dataclass
class AetherV1PipelineOutput(BaseOutput):
rgb: np.ndarray
disparity: np.ndarray
raymap: np.ndarray
class AetherV1PipelineCogVideoX(CogVideoXImageToVideoPipeline):
_supported_tasks = ["reconstruction", "prediction", "planning"]
_default_num_inference_steps = {
"reconstruction": 4,
"prediction": 50,
"planning": 50,
}
_default_guidance_scale = {
"reconstruction": 1.0,
"prediction": 3.0,
"planning": 3.0,
}
_default_use_dynamic_cfg = {
"reconstruction": False,
"prediction": True,
"planning": True,
}
_base_fps = 12
def __init__(
self,
tokenizer: AutoTokenizer,
text_encoder: T5EncoderModel,
vae: AutoencoderKLCogVideoX,
scheduler: CogVideoXDPMScheduler,
transformer: CogVideoXTransformer3DModel,
):
super().__init__(
tokenizer=tokenizer,
text_encoder=text_encoder,
vae=vae,
scheduler=scheduler,
transformer=transformer,
)
self.empty_prompt_embeds, _ = self.encode_prompt(
prompt="",
negative_prompt=None,
do_classifier_free_guidance=False,
num_videos_per_prompt=1,
prompt_embeds=None,
)
self.empty_prompt_embeds = self.empty_prompt_embeds.to(dtype=torch.bfloat16)
def _prepare_rotary_positional_embeddings(
self,
height: int,
width: int,
num_frames: int,
device: torch.device,
fps: Optional[int] = None,
) -> Tuple[torch.Tensor, torch.Tensor]:
grid_height = height // (
self.vae_scale_factor_spatial * self.transformer.config.patch_size
)
grid_width = width // (
self.vae_scale_factor_spatial * self.transformer.config.patch_size
)
p = self.transformer.config.patch_size
p_t = self.transformer.config.patch_size_t
base_size_width = self.transformer.config.sample_width // p
base_size_height = self.transformer.config.sample_height // p
if p_t is None:
# CogVideoX 1.0
grid_crops_coords = get_resize_crop_region_for_grid(
(grid_height, grid_width), base_size_width, base_size_height
)
freqs_cos, freqs_sin = get_3d_rotary_pos_embed(
embed_dim=self.transformer.config.attention_head_dim,
crops_coords=grid_crops_coords,
grid_size=(grid_height, grid_width),
temporal_size=num_frames,
device=device,
fps_factor=self._base_fps / fps,
)
else:
# CogVideoX 1.5
base_num_frames = (num_frames + p_t - 1) // p_t
freqs_cos, freqs_sin = get_3d_rotary_pos_embed(
embed_dim=self.transformer.config.attention_head_dim,
crops_coords=None,
grid_size=(grid_height, grid_width),
temporal_size=base_num_frames,
grid_type="slice",
max_size=(base_size_height, base_size_width),
device=device,
fps_factor=self._base_fps / fps,
)
return freqs_cos, freqs_sin
def check_inputs(
self,
task,
image,
video,
goal,
raymap,
height,
width,
num_frames,
fps,
):
if task not in self._supported_tasks:
raise ValueError(f"`task` has to be one of {self._supported_tasks}.")
if image is None and video is None:
raise ValueError("`image` or `video` has to be provided.")
if image is not None and video is not None:
raise ValueError("`image` and `video` cannot both be provided.")
if image is not None:
if task == "reconstruction":
raise ValueError("`image` is not supported for `reconstruction` task.")
if (
not isinstance(image, torch.Tensor)
and not isinstance(image, np.ndarray)
and not isinstance(image, PIL.Image.Image)
):
raise ValueError(
"`image` has to be of type `torch.Tensor` or `np.ndarray` or `PIL.Image.Image` but is"
f" {type(image)}"
)
if goal is not None:
if task != "planning":
raise ValueError("`goal` is only supported for `planning` task.")
if (
not isinstance(goal, torch.Tensor)
and not isinstance(goal, np.ndarray)
and not isinstance(goal, PIL.Image.Image)
):
raise ValueError(
"`goal` has to be of type `torch.Tensor` or `np.ndarray` or `PIL.Image.Image` but is"
f" {type(goal)}"
)
if video is not None:
if task != "reconstruction":
raise ValueError("`video` is only supported for `reconstruction` task.")
if (
not isinstance(video, torch.Tensor)
and not isinstance(video, np.ndarray)
and not (
isinstance(video, list)
and all(isinstance(v, PIL.Image.Image) for v in video)
)
):
raise ValueError(
"`video` has to be of type `torch.Tensor` or `np.ndarray` or `List[PIL.Image.Image]` but is"
f" {type(video)}"
)
if height % 8 != 0 or width % 8 != 0:
raise ValueError(
f"`height` and `width` have to be divisible by 8 but are {height} and {width}."
)
if num_frames is None:
raise ValueError("`num_frames` is required.")
if num_frames not in [17, 25, 33, 41]:
raise ValueError("`num_frames` has to be one of [17, 25, 33, 41].")
if fps not in [8, 10, 12, 15, 24]:
raise ValueError("`fps` has to be one of [8, 10, 12, 15, 24].")
if (
raymap is not None
and not isinstance(raymap, torch.Tensor)
and not isinstance(raymap, np.ndarray)
):
raise ValueError(
"`raymap` has to be of type `torch.Tensor` or `np.ndarray`."
)
if raymap is not None:
if raymap.shape[-4:] != (
num_frames,
6,
height // self.vae_scale_factor_spatial,
width // self.vae_scale_factor_spatial,
):
raise ValueError(
f"`raymap` shape is not correct. "
f"Expected {num_frames, 6, height // self.vae_scale_factor_spatial, width // self.vae_scale_factor_spatial}, "
f"got {raymap.shape}."
)
def _preprocess_image(self, image, height, width):
if isinstance(image, torch.Tensor):
image = image.cpu().numpy()
if image.dtype == np.uint8:
image = image.astype(np.float32) / 255.0
if image.ndim == 3:
image = [image]
image = imcrop_center(image, height, width)
image = self.video_processor.preprocess(image, height, width)
return image
def preprocess_inputs(
self,
image,
goal,
video,
raymap,
height,
width,
num_frames,
):
if image is not None:
if isinstance(image, PIL.Image.Image):
image = self.video_processor.preprocess(
image, height, width, resize_mode="crop"
).to(device=self._execution_device, dtype=torch.bfloat16)
else:
image = self._preprocess_image(image, height, width).to(
device=self._execution_device, dtype=torch.bfloat16
)
if goal is not None:
if isinstance(goal, PIL.Image.Image):
goal = self.video_processor.preprocess(
goal, height, width, resize_mode="crop"
).to(device=self._execution_device, dtype=torch.bfloat16)
else:
goal = self._preprocess_image(goal, height, width).to(
device=self._execution_device, dtype=torch.bfloat16
)
if video is not None:
if isinstance(video, list) and all(
isinstance(v, PIL.Image.Image) for v in video
):
video = self.video_processor.preprocess(
video, height, width, resize_mode="crop"
).to(device=self._execution_device, dtype=torch.bfloat16)
else:
video = self._preprocess_image(video, height, width).to(
device=self._execution_device, dtype=torch.bfloat16
)
# TODO: check raymap shape
if raymap is not None:
if isinstance(raymap, np.ndarray):
raymap = torch.from_numpy(raymap).to(
self._execution_device, dtype=torch.bfloat16
)
if raymap.ndim == 4:
raymap = raymap.unsqueeze(0).to(
self._execution_device, dtype=torch.bfloat16
)
return image, goal, video, raymap
@torch.no_grad()
def prepare_latents(
self,
image: Optional[torch.Tensor] = None,
goal: Optional[torch.Tensor] = None,
video: Optional[torch.Tensor] = None,
raymap: Optional[torch.Tensor] = None,
batch_size: int = 1,
num_frames: int = 13,
height: int = 60,
width: int = 90,
dtype: Optional[torch.dtype] = None,
device: Optional[torch.device] = None,
generator: Optional[torch.Generator] = None,
):
if isinstance(generator, list) and len(generator) != batch_size:
raise ValueError(
f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
f" size of {batch_size}. Make sure the batch size matches the length of the generators."
)
num_frames = (num_frames - 1) // self.vae_scale_factor_temporal + 1
shape = (
batch_size,
num_frames,
56,
height // self.vae_scale_factor_spatial,
width // self.vae_scale_factor_spatial,
)
# For CogVideoX1.5, the latent should add 1 for padding (Not use)
if self.transformer.config.patch_size_t is not None:
shape = (
shape[:1]
+ (shape[1] + shape[1] % self.transformer.config.patch_size_t,)
+ shape[2:]
)
if image is not None:
image = image.unsqueeze(2)
if isinstance(generator, list):
image_latents = [
retrieve_latents(
self.vae.encode(image[i].unsqueeze(0)), generator[i]
)
for i in range(batch_size)
]
else:
image_latents = [
retrieve_latents(self.vae.encode(img.unsqueeze(0)), generator)
for img in image
]
image_latents = (
torch.cat(image_latents, dim=0).to(dtype).permute(0, 2, 1, 3, 4)
) # [B, F, C, H, W]
if not self.vae.config.invert_scale_latents:
image_latents = self.vae_scaling_factor_image * image_latents
else:
# This is awkward but required because the CogVideoX team forgot to multiply the
# scaling factor during training :)
image_latents = 1 / self.vae_scaling_factor_image * image_latents
if goal is not None:
goal = goal.unsqueeze(2)
if isinstance(generator, list):
goal_latents = [
retrieve_latents(
self.vae.encode(goal[i].unsqueeze(0)), generator[i]
)
for i in range(batch_size)
]
else:
goal_latents = [
retrieve_latents(self.vae.encode(img.unsqueeze(0)), generator)
for img in goal
]
goal_latents = (
torch.cat(goal_latents, dim=0).to(dtype).permute(0, 2, 1, 3, 4)
) # [B, F, C, H, W]
if not self.vae.config.invert_scale_latents:
goal_latents = self.vae_scaling_factor_image * goal_latents
else:
# This is awkward but required because the CogVideoX team forgot to multiply the
# scaling factor during training :)
goal_latents = 1 / self.vae_scaling_factor_image * goal_latents
if video is not None:
if video.ndim == 4:
video = video.unsqueeze(0)
video = video.permute(0, 2, 1, 3, 4)
if isinstance(generator, list):
video_latents = [
retrieve_latents(
self.vae.encode(video[i].unsqueeze(0)), generator[i]
)
for i in range(batch_size)
]
else:
video_latents = [
retrieve_latents(self.vae.encode(img.unsqueeze(0)), generator)
for img in video
]
video_latents = (
torch.cat(video_latents, dim=0).to(dtype).permute(0, 2, 1, 3, 4)
) # [B, F, C, H, W]
if not self.vae.config.invert_scale_latents:
video_latents = self.vae_scaling_factor_image * video_latents
else:
# This is awkward but required because the CogVideoX team forgot to multiply the
# scaling factor during training :)
video_latents = 1 / self.vae_scaling_factor_image * video_latents
if image is not None and goal is None:
padding_shape = (
batch_size,
num_frames - image_latents.shape[1],
*image_latents.shape[2:],
)
padding = torch.zeros(padding_shape, device=device, dtype=dtype)
condition_latents = torch.cat([image_latents, padding], dim=1)
elif goal is not None:
padding_shape = (
batch_size,
num_frames - goal_latents.shape[1] - image_latents.shape[1],
*image_latents.shape[2:],
)
padding = torch.zeros(padding_shape, device=device, dtype=dtype)
condition_latents = torch.cat([image_latents, padding, goal_latents], dim=1)
elif video is not None:
condition_latents = video_latents
if raymap is not None:
if raymap.shape[1] % self.vae_scale_factor_temporal != 0:
# repeat
raymap = torch.cat(
[
raymap[
:,
: self.vae_scale_factor_temporal
- raymap.shape[1] % self.vae_scale_factor_temporal,
],
raymap,
],
dim=1,
)
camera_conditions = rearrange(
raymap,
"b (n t) c h w -> b t (n c) h w",
n=self.vae_scale_factor_temporal,
)
else:
camera_conditions = torch.zeros(
batch_size,
num_frames,
24,
height // self.vae_scale_factor_spatial,
width // self.vae_scale_factor_spatial,
device=device,
dtype=dtype,
)
condition_latents = torch.cat([condition_latents, camera_conditions], dim=2)
latents = randn_tensor(shape, device=device, generator=generator, dtype=dtype)
# scale the initial noise by the standard deviation required by the scheduler
latents = latents * self.scheduler.init_noise_sigma
return latents, condition_latents
@torch.no_grad()
def __call__(
self,
task: Optional[str] = None,
image: Optional[PipelineImageInput] = None,
video: Optional[PipelineImageInput] = None,
goal: Optional[PipelineImageInput] = None,
raymap: Optional[Union[torch.Tensor, np.ndarray]] = None,
height: Optional[int] = None,
width: Optional[int] = None,
num_frames: Optional[int] = None,
num_inference_steps: Optional[int] = None,
timesteps: Optional[List[int]] = None,
guidance_scale: Optional[float] = None,
use_dynamic_cfg: bool = False,
num_videos_per_prompt: int = 1,
eta: float = 0.0,
generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
return_dict: bool = True,
attention_kwargs: Optional[Dict] = None,
fps: Optional[int] = None,
) -> Union[AetherV1PipelineOutput, Tuple]:
if task is None:
if video is not None:
task = "reconstruction"
elif goal is not None:
task = "planning"
else:
task = "prediction"
height = (
height
or self.transformer.config.sample_height * self.vae_scale_factor_spatial
)
width = (
width
or self.transformer.config.sample_width * self.vae_scale_factor_spatial
)
num_frames = num_frames or self.transformer.config.sample_frames
fps = fps or self._base_fps
num_videos_per_prompt = 1
# 1. Check inputs. Raise error if not correct
self.check_inputs(
task=task,
image=image,
video=video,
goal=goal,
raymap=raymap,
height=height,
width=width,
num_frames=num_frames,
fps=fps,
)
# 2. Preprocess inputs
image, goal, video, raymap = self.preprocess_inputs(
image=image,
goal=goal,
video=video,
raymap=raymap,
height=height,
width=width,
num_frames=num_frames,
)
self._guidance_scale = guidance_scale
self._current_timestep = None
self._attention_kwargs = attention_kwargs
self._interrupt = False
batch_size = 1
device = self._execution_device
# 3. Encode input prompt
prompt_embeds = self.empty_prompt_embeds.to(device)
num_inference_steps = (
num_inference_steps or self._default_num_inference_steps[task]
)
guidance_scale = guidance_scale or self._default_guidance_scale[task]
use_dynamic_cfg = use_dynamic_cfg or self._default_use_dynamic_cfg[task]
# 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 = guidance_scale > 1.0
# 4. Prepare timesteps
timesteps, num_inference_steps = retrieve_timesteps(
self.scheduler, num_inference_steps, device, timesteps
)
self._num_timesteps = len(timesteps)
# 5. Prepare latents
latents, condition_latents = self.prepare_latents(
image,
goal,
video,
raymap,
batch_size * num_videos_per_prompt,
num_frames,
height,
width,
prompt_embeds.dtype,
device,
generator,
)
# 6. 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. Create rotary embeds if required
image_rotary_emb = (
self._prepare_rotary_positional_embeddings(
height, width, latents.size(1), device, fps=fps
)
if self.transformer.config.use_rotary_positional_embeddings
else None
)
# 8. Create ofs embeds if required
ofs_emb = (
None
if self.transformer.config.ofs_embed_dim is None
else latents.new_full((1,), fill_value=2.0)
)
# 8. Denoising loop
num_warmup_steps = max(
len(timesteps) - num_inference_steps * self.scheduler.order, 0
)
with self.progress_bar(total=num_inference_steps) as progress_bar:
# for DPM-solver++
old_pred_original_sample = None
for i, t in enumerate(timesteps):
if self.interrupt:
continue
self._current_timestep = t
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
)
if do_classifier_free_guidance:
if task == "planning":
assert goal is not None
uncond = condition_latents.clone()
uncond[:, :, : self.vae.config.latent_channels] = 0
latent_condition = torch.cat([uncond, condition_latents])
elif task == "prediction":
uncond = condition_latents.clone()
uncond[:, :1, : self.vae.config.latent_channels] = 0
latent_condition = torch.cat([uncond, condition_latents])
else:
raise ValueError(
f"Task {task} not supported for classifier-free guidance."
)
else:
latent_condition = condition_latents
latent_model_input = torch.cat(
[latent_model_input, latent_condition], dim=2
)
# broadcast to batch dimension in a way that's compatible with ONNX/Core ML
timestep = t.expand(latent_model_input.shape[0])
# predict noise model_output
noise_pred = self.transformer(
hidden_states=latent_model_input,
encoder_hidden_states=prompt_embeds.repeat(
latent_model_input.shape[0], 1, 1
),
timestep=timestep,
ofs=ofs_emb,
image_rotary_emb=image_rotary_emb,
attention_kwargs=attention_kwargs,
return_dict=False,
)[0]
noise_pred = noise_pred.float()
# perform guidance
if use_dynamic_cfg:
self._guidance_scale = 1 + guidance_scale * (
(
1
- math.cos(
math.pi
* (
(num_inference_steps - t.item())
/ num_inference_steps
)
** 5.0
)
)
/ 2
)
if do_classifier_free_guidance:
noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
noise_pred = noise_pred_uncond + self.guidance_scale * (
noise_pred_text - noise_pred_uncond
)
# compute the previous noisy sample x_t -> x_t-1
if not isinstance(self.scheduler, CogVideoXDPMScheduler):
latents = self.scheduler.step(
noise_pred, t, latents, **extra_step_kwargs, return_dict=False
)[0]
else:
latents, old_pred_original_sample = self.scheduler.step(
noise_pred,
old_pred_original_sample,
t,
timesteps[i - 1] if i > 0 else None,
latents,
**extra_step_kwargs,
return_dict=False,
)
latents = latents.to(prompt_embeds.dtype)
if i == len(timesteps) - 1 or (
(i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0
):
progress_bar.update()
self._current_timestep = None
rgb_latents = latents[:, :, : self.vae.config.latent_channels]
disparity_latents = latents[
:, :, self.vae.config.latent_channels : self.vae.config.latent_channels * 2
]
camera_latents = latents[:, :, self.vae.config.latent_channels * 2 :]
rgb_video = self.decode_latents(rgb_latents)
rgb_video = self.video_processor.postprocess_video(
video=rgb_video, output_type="np"
)
disparity_video = self.decode_latents(disparity_latents)
disparity_video = disparity_video.mean(dim=1, keepdim=False)
disparity_video = disparity_video * 0.5 + 0.5
disparity_video = torch.square(disparity_video)
disparity_video = disparity_video.float().cpu().numpy()
raymap = (
rearrange(camera_latents, "b t (n c) h w -> b (n t) c h w", n=4)[
:, -rgb_video.shape[1] :, :, :
]
.float()
.cpu()
.numpy()
)
# Offload all models
self.maybe_free_model_hooks()
if not return_dict:
return (
rgb_video,
disparity_video,
raymap,
)
return AetherV1PipelineOutput(
rgb=rgb_video.squeeze(0),
disparity=disparity_video.squeeze(0),
raymap=raymap.squeeze(0),
)