ImageConductor / pipelines /pipeline_imagecoductor.py
Yw22's picture
envs
6a75750
# Adapted from https://github.com/showlab/Tune-A-Video/blob/main/tuneavideo/pipelines/pipeline_tuneavideo.py
import inspect
import copy
import spaces
from dataclasses import dataclass
from typing import Any, Callable, Dict, List, Optional, Union
import numpy as np
import torch
from diffusers.configuration_utils import FrozenDict
from diffusers.loaders import IPAdapterMixin
from diffusers.models import AutoencoderKL
from diffusers.pipelines.pipeline_utils import DiffusionPipeline
from diffusers.schedulers import (
DDIMScheduler,
DPMSolverMultistepScheduler,
EulerAncestralDiscreteScheduler,
EulerDiscreteScheduler,
LMSDiscreteScheduler,
PNDMScheduler,
)
from diffusers.utils import BaseOutput, deprecate, is_accelerate_available, logging
from einops import rearrange
from packaging import version
from tqdm import tqdm
from transformers import CLIPTextModel, CLIPTokenizer
from modules.unet import UNet3DConditionFlowModel
from modules.flow_controlnet import FlowControlNetModel
from modules.image_controlnet import ImageControlNetModel
logger = logging.get_logger(__name__) # pylint: disable=invalid-name
@dataclass
class ImageConductorPipelineOutput(BaseOutput):
videos: Union[torch.Tensor, np.ndarray]
class ImageConductorPipeline(DiffusionPipeline):
_optional_components = []
def __init__(
self,
vae: AutoencoderKL,
text_encoder: CLIPTextModel,
tokenizer: CLIPTokenizer,
unet: UNet3DConditionFlowModel,
scheduler: Union[
DDIMScheduler,
PNDMScheduler,
LMSDiscreteScheduler,
EulerDiscreteScheduler,
EulerAncestralDiscreteScheduler,
DPMSolverMultistepScheduler,
],
image_controlnet: Union[ImageControlNetModel, None] = None,
flow_controlnet: Union[FlowControlNetModel, None] = None,
):
super().__init__()
if hasattr(scheduler.config, "steps_offset") and scheduler.config.steps_offset != 1:
deprecation_message = (
f"The configuration file of this scheduler: {scheduler} is outdated. `steps_offset`"
f" should be set to 1 instead of {scheduler.config.steps_offset}. Please make sure "
"to update the config accordingly as leaving `steps_offset` might led to incorrect results"
" in future versions. If you have downloaded this checkpoint from the Hugging Face Hub,"
" it would be very nice if you could open a Pull request for the `scheduler/scheduler_config.json`"
" file"
)
deprecate("steps_offset!=1", "1.0.0", deprecation_message, standard_warn=False)
new_config = dict(scheduler.config)
new_config["steps_offset"] = 1
scheduler._internal_dict = FrozenDict(new_config)
if hasattr(scheduler.config, "clip_sample") and scheduler.config.clip_sample is True:
deprecation_message = (
f"The configuration file of this scheduler: {scheduler} has not set the configuration `clip_sample`."
" `clip_sample` should be set to False in the configuration file. Please make sure to update the"
" config accordingly as not setting `clip_sample` in the config might lead to incorrect results in"
" future versions. If you have downloaded this checkpoint from the Hugging Face Hub, it would be very"
" nice if you could open a Pull request for the `scheduler/scheduler_config.json` file"
)
deprecate("clip_sample not set", "1.0.0", deprecation_message, standard_warn=False)
new_config = dict(scheduler.config)
new_config["clip_sample"] = False
scheduler._internal_dict = FrozenDict(new_config)
is_unet_version_less_0_9_0 = hasattr(unet.config, "_diffusers_version") and version.parse(
version.parse(unet.config._diffusers_version).base_version
) < version.parse("0.9.0.dev0")
is_unet_sample_size_less_64 = hasattr(unet.config, "sample_size") and unet.config.sample_size < 64
if is_unet_version_less_0_9_0 and is_unet_sample_size_less_64:
deprecation_message = (
"The configuration file of the unet has set the default `sample_size` to smaller than"
" 64 which seems highly unlikely. If your checkpoint is a fine-tuned version of any of the"
" following: \n- CompVis/stable-diffusion-v1-4 \n- CompVis/stable-diffusion-v1-3 \n-"
" CompVis/stable-diffusion-v1-2 \n- CompVis/stable-diffusion-v1-1 \n- runwayml/stable-diffusion-v1-5"
" \n- runwayml/stable-diffusion-inpainting \n you should change 'sample_size' to 64 in the"
" configuration file. Please make sure to update the config accordingly as leaving `sample_size=32`"
" in the config might lead to incorrect results in future versions. If you have downloaded this"
" checkpoint from the Hugging Face Hub, it would be very nice if you could open a Pull request for"
" the `unet/config.json` file"
)
deprecate("sample_size<64", "1.0.0", deprecation_message, standard_warn=False)
new_config = dict(unet.config)
new_config["sample_size"] = 64
unet._internal_dict = FrozenDict(new_config)
self.register_modules(
vae=vae,
text_encoder=text_encoder,
tokenizer=tokenizer,
unet=unet,
scheduler=scheduler,
image_controlnet=image_controlnet,
flow_controlnet=flow_controlnet,
)
self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
def enable_vae_slicing(self):
self.vae.enable_slicing()
def disable_vae_slicing(self):
self.vae.disable_slicing()
def enable_sequential_cpu_offload(self, gpu_id=0):
if is_accelerate_available():
from accelerate import cpu_offload
else:
raise ImportError("Please install accelerate via `pip install accelerate`")
device = torch.device(f"cuda:{gpu_id}")
for cpu_offloaded_model in [self.unet, self.text_encoder, self.vae]:
if cpu_offloaded_model is not None:
cpu_offload(cpu_offloaded_model, device)
@property
def _execution_device(self):
if self.device != torch.device("meta") or not hasattr(self.unet, "_hf_hook"):
return self.device
for module in self.unet.modules():
if (
hasattr(module, "_hf_hook")
and hasattr(module._hf_hook, "execution_device")
and module._hf_hook.execution_device is not None
):
return torch.device(module._hf_hook.execution_device)
return self.device
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
text_inputs = self.tokenizer(
prompt,
padding="max_length",
max_length=self.tokenizer.model_max_length,
truncation=True,
return_tensors="pt",
)
text_input_ids = text_inputs.input_ids
untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids
if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(text_input_ids, untruncated_ids):
removed_text = self.tokenizer.batch_decode(untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1])
logger.warning(
"The following part of your input was truncated because CLIP can only handle sequences up to"
f" {self.tokenizer.model_max_length} tokens: {removed_text}"
)
if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask:
attention_mask = text_inputs.attention_mask.to(device)
else:
attention_mask = None
text_embeddings = self.text_encoder(
text_input_ids.to(device),
attention_mask=attention_mask,
)
text_embeddings = text_embeddings[0]
# duplicate text embeddings for each generation per prompt, using mps friendly method
bs_embed, seq_len, _ = text_embeddings.shape
text_embeddings = text_embeddings.repeat(1, num_videos_per_prompt, 1)
text_embeddings = text_embeddings.view(bs_embed * num_videos_per_prompt, seq_len, -1)
# get unconditional embeddings for classifier free guidance
if do_classifier_free_guidance:
uncond_tokens: List[str]
if negative_prompt is None:
uncond_tokens = [""] * batch_size
elif type(prompt) is not type(negative_prompt):
raise TypeError(
f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
f" {type(prompt)}."
)
elif isinstance(negative_prompt, str):
uncond_tokens = [negative_prompt]
elif batch_size != len(negative_prompt):
raise ValueError(
f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
" the batch size of `prompt`."
)
else:
uncond_tokens = negative_prompt
max_length = text_input_ids.shape[-1]
uncond_input = self.tokenizer(
uncond_tokens,
padding="max_length",
max_length=max_length,
truncation=True,
return_tensors="pt",
)
if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask:
attention_mask = uncond_input.attention_mask.to(device)
else:
attention_mask = None
uncond_embeddings = self.text_encoder(
uncond_input.input_ids.to(device),
attention_mask=attention_mask,
)
uncond_embeddings = uncond_embeddings[0]
# duplicate unconditional embeddings for each generation per prompt, using mps friendly method
seq_len = uncond_embeddings.shape[1]
uncond_embeddings = uncond_embeddings.repeat(1, num_videos_per_prompt, 1)
uncond_embeddings = uncond_embeddings.view(batch_size * num_videos_per_prompt, seq_len, -1)
# For classifier free guidance, we need to do two forward passes.
# Here we concatenate the unconditional and text embeddings into a single batch
# to avoid doing two forward passes
text_embeddings = torch.cat([uncond_embeddings, text_embeddings])
return text_embeddings
def decode_latents(self, latents):
video_length = latents.shape[2]
latents = 1 / 0.18215 * latents
latents = rearrange(latents, "b c f h w -> (b f) c h w")
# video = self.vae.decode(latents).sample
video = []
for frame_idx in tqdm(range(latents.shape[0])):
video.append(self.vae.decode(latents[frame_idx:frame_idx+1]).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)
# we always cast to float32 as this does not cause significant overhead and is compatible with bfloa16
video = video.cpu().float().numpy()
return video
def prepare_extra_step_kwargs(self, generator, eta):
# prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
# eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
# eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
# and should be between [0, 1]
accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
extra_step_kwargs = {}
if accepts_eta:
extra_step_kwargs["eta"] = eta
# check if the scheduler accepts generator
accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys())
if accepts_generator:
extra_step_kwargs["generator"] = generator
return extra_step_kwargs
def check_inputs(self, prompt, height, width, callback_steps):
if not isinstance(prompt, str) and not isinstance(prompt, list):
raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
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 (callback_steps is None) or (
callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0)
):
raise ValueError(
f"`callback_steps` has to be a positive integer but is {callback_steps} of type"
f" {type(callback_steps)}."
)
def prepare_latents(self, batch_size, num_channels_latents, video_length, height, width, dtype, device, generator, latents=None):
shape = (batch_size, num_channels_latents, video_length, height // self.vae_scale_factor, width // self.vae_scale_factor)
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."
)
if latents is None:
rand_device = "cpu" if device.type == "mps" else device
if isinstance(generator, list):
shape = shape
# shape = (1,) + shape[1:]
latents = [
torch.randn(shape, generator=generator[i], device=rand_device, dtype=dtype)
for i in range(batch_size)
]
latents = torch.cat(latents, dim=0).to(device)
else:
latents = torch.randn(shape, generator=generator, device=rand_device, dtype=dtype).to(device)
else:
if latents.shape != shape:
raise ValueError(f"Unexpected latents shape, got {latents.shape}, expected {shape}")
latents = latents.to(device)
# scale the initial noise by the standard deviation required by the scheduler
latents = latents * self.scheduler.init_noise_sigma
return latents
@torch.no_grad()
def __call__(
self,
prompt: Union[str, List[str]],
video_length: Optional[int],
height: Optional[int] = None,
width: Optional[int] = None,
num_inference_steps: int = 50,
guidance_scale: float = 7.5,
obj_guidance_scale: float = 1.0,
negative_prompt: Optional[Union[str, List[str]]] = None,
num_videos_per_prompt: Optional[int] = 1,
eta: float = 0.0,
generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
latents: Optional[torch.FloatTensor] = None,
output_type: Optional[str] = "tensor",
return_dict: bool = True,
callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
callback_steps: Optional[int] = 1,
# support controlnet
controlnet_images: torch.FloatTensor = None,
controlnet_image_index: list = [0],
controlnet_conditioning_scale: Union[float, List[float]] = 1.0,
controlnet_flows: torch.FloatTensor = None,
eval_mode: bool = False,
control_mode: Optional[str] = "object",
**kwargs,
):
# Default height and width to unet
height = height or self.unet.config.sample_size * self.vae_scale_factor
width = width or self.unet.config.sample_size * self.vae_scale_factor
# Check inputs. Raise error if not correct
self.check_inputs(prompt, height, width, callback_steps)
# Define call parameters
# batch_size = 1 if isinstance(prompt, str) else len(prompt)
batch_size = 1
if latents is not None:
batch_size = latents.shape[0]
if isinstance(prompt, list):
batch_size = len(prompt)
device = self._execution_device
# 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
# do_camera_free_guidance = obj_guidance_scale >= 1.0 ## todo: camera-free guidance
# Encode input prompt, [uncond_embeddings, cond_embeddings] for CFG
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
text_embeddings = self._encode_prompt(
prompt, device, num_videos_per_prompt, do_classifier_free_guidance, negative_prompt
)
# Prepare timesteps
self.scheduler.set_timesteps(num_inference_steps, device=device)
timesteps = self.scheduler.timesteps
# Prepare latent variables
num_channels_latents = self.unet.in_channels
latents = self.prepare_latents(
batch_size * num_videos_per_prompt,
num_channels_latents,
video_length,
height,
width,
text_embeddings.dtype,
device,
generator,
latents,
)
latents_dtype = latents.dtype
obj_latents = copy.deepcopy(latents)
cam_latents = copy.deepcopy(latents)
# Prepare extra step kwargs.
extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
self.image_controlnet.to(device)
self.flow_controlnet.to(device)
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_obj_model_input = torch.cat([obj_latents] * 2) if do_classifier_free_guidance else obj_latents
latent_obj_model_input = self.scheduler.scale_model_input(latent_obj_model_input, t)
latent_cam_model_input = torch.cat([cam_latents] * 2) if do_classifier_free_guidance else cam_latents
latent_cam_model_input = self.scheduler.scale_model_input(latent_cam_model_input, t)
img_down_block_additional_residuals = img_mid_block_additional_residuals = None
flow_down_block_additional_residuals_obj = flow_mid_block_additional_residual_obj = None
flow_down_block_additional_residuals_cam = flow_mid_block_additional_residual_cam = None
if control_mode == "object":
controlnet_noisy_latents = latent_obj_model_input
elif control_mode == "camera":
controlnet_noisy_latents = latent_cam_model_input
if (getattr(self, "image_controlnet", None) != None) and (controlnet_images != None):
assert controlnet_images.dim() == 5
controlnet_prompt_embeds = text_embeddings
# concat mask and images condition
controlnet_images_ref = controlnet_images.to(latents.device)
controlnet_images_shape = list(controlnet_images.shape)
controlnet_images_shape[2] = video_length
controlnet_images = torch.zeros(controlnet_images_shape).to(latents.device)
controlnet_images_mask_shape = list(controlnet_images.shape)
controlnet_images_mask_shape[1] = 1
controlnet_images_mask = torch.zeros(controlnet_images_mask_shape).to(latents.device)
assert controlnet_images.shape[2] >= len(controlnet_image_index)
controlnet_images[:,:,controlnet_image_index] = controlnet_images_ref[:,:,:len(controlnet_image_index)]
controlnet_images_mask[:,:,controlnet_image_index] = 1
# fp16 accelerate training in deepspeed
if not eval_mode:
controlnet_noisy_latents = controlnet_noisy_latents.half()
controlnet_prompt_embeds = controlnet_prompt_embeds.half()
controlnet_images = controlnet_images.half()
controlnet_images_mask = controlnet_images_mask.half()
controlnet_flows = controlnet_flows.half()
text_embeddings = text_embeddings.half()
img_down_block_additional_residuals, img_mid_block_additional_residuals = self.image_controlnet(
controlnet_noisy_latents, t,
encoder_hidden_states=controlnet_prompt_embeds,
controlnet_cond=controlnet_images,
conditioning_mask=controlnet_images_mask,
conditioning_scale=controlnet_conditioning_scale,
guess_mode=False, return_dict=False,
)
if (getattr(self, "flow_controlnet", None) != None) and (controlnet_flows != None):
# Get the flow controlnet
# self.flow_controlnet.disable_adapters()
if control_mode == "camera":
self.flow_controlnet.enable_adapters()
self.flow_controlnet.set_adapter(["cmcm_weights"])
flow_down_block_additional_residuals_cam, flow_mid_block_additional_residual_cam = self.flow_controlnet(
controlnet_noisy_latents, t,
encoder_hidden_states=controlnet_prompt_embeds,
controlnet_cond=controlnet_flows, # bcthw
conditioning_scale=controlnet_conditioning_scale,
guess_mode=False, return_dict=False,
)
## camera transition latent outputs
noise_pred_cam = self.unet(
controlnet_noisy_latents, t,
encoder_hidden_states=text_embeddings,
down_block_additional_residuals = img_down_block_additional_residuals,
mid_block_additional_residual = img_mid_block_additional_residuals,
flow_down_block_additional_residuals = flow_down_block_additional_residuals_cam,
flow_mid_block_additional_residual = flow_mid_block_additional_residual_cam,
).sample.to(dtype=latents_dtype)
# perform CFG for camera
if do_classifier_free_guidance:
noise_pred_cam_uncond, noise_pred_cam_text = noise_pred_cam.chunk(2)
noise_pred_cam = noise_pred_cam_uncond + guidance_scale * (noise_pred_cam_text - noise_pred_cam_uncond)
# compute the previous noisy sample x_t -> x_t-1
cam_latents = self.scheduler.step(noise_pred_cam, t, cam_latents, **extra_step_kwargs).prev_sample
if control_mode == "object":
self.flow_controlnet.enable_adapters()
self.flow_controlnet.set_adapter(["omcm_weights"])
flow_down_block_additional_residuals_obj, flow_mid_block_additional_residual_obj = self.flow_controlnet(
controlnet_noisy_latents, t,
encoder_hidden_states=controlnet_prompt_embeds,
controlnet_cond=controlnet_flows, # bcthw
conditioning_scale=controlnet_conditioning_scale,
guess_mode=False, return_dict=False,
)
## object movement latent outputs
noise_pred_obj = self.unet(
controlnet_noisy_latents, t,
encoder_hidden_states=text_embeddings,
down_block_additional_residuals = img_down_block_additional_residuals,
mid_block_additional_residual = img_mid_block_additional_residuals,
flow_down_block_additional_residuals = flow_down_block_additional_residuals_obj,
flow_mid_block_additional_residual = flow_mid_block_additional_residual_obj,
).sample.to(dtype=latents_dtype)
# perform CFG for object
if do_classifier_free_guidance:
noise_pred_obj_uncond, noise_pred_obj_text = noise_pred_obj.chunk(2)
noise_pred_obj = noise_pred_obj_uncond + guidance_scale * (noise_pred_obj_text - noise_pred_obj_uncond)
# compute the previous noisy sample x_t -> x_t-1
obj_latents = self.scheduler.step(noise_pred_obj, t, obj_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, cam_latents, obj_latents)
# Post-processning
if control_mode == "object":
video_obj = self.decode_latents(obj_latents)
videos = video_obj
elif control_mode == "camera":
video_cam = self.decode_latents(cam_latents)
videos = video_cam
# Convert to tensor
if output_type == "tensor":
videos = torch.from_numpy(videos)
if not return_dict:
return videos
return ImageConductorPipelineOutput(videos=videos)