File size: 11,553 Bytes
6eca12e |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 |
import inspect
import os
import time
from typing import Any, Callable, Dict, List, Optional, Union, Tuple
import gc
import torch
import numpy as np
from glob import glob
from diffusers import StableDiffusionXLPipeline, UNet2DConditionModel
from diffusers.loaders import TextualInversionLoaderMixin
from diffusers.image_processor import VaeImageProcessor
from diffusers.models import AutoencoderKL
from diffusers.schedulers import (DPMSolverMultistepScheduler,
EulerAncestralDiscreteScheduler,
EulerDiscreteScheduler,
KarrasDiffusionSchedulers)
from diffusers.models.embeddings import TimestepEmbedding, Timesteps
from diffusers.utils.torch_utils import randn_tensor
from diffusers.utils import logging
from PIL import Image
from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer, CLIPTextModelWithProjection
from .lyrasd_vae_model import LyraSdVaeModel
from .module.lyrasd_ip_adapter import LyraIPAdapter
from .lora_util import add_text_lora_layer, add_xltext_lora_layer, add_lora_to_opt_model, load_state_dict
from safetensors.torch import load_file
from .lyrasdxl_pipeline_base import LyraSDXLPipelineBase
def rescale_noise_cfg(noise_cfg, noise_pred_text, guidance_rescale=0.0):
"""
Rescale `noise_cfg` according to `guidance_rescale`. Based on findings of [Common Diffusion Noise Schedules and
Sample Steps are Flawed](https://arxiv.org/pdf/2305.08891.pdf). See Section 3.4
"""
std_text = noise_pred_text.std(
dim=list(range(1, noise_pred_text.ndim)), keepdim=True)
std_cfg = noise_cfg.std(dim=list(range(1, noise_cfg.ndim)), keepdim=True)
# rescale the results from guidance (fixes overexposure)
noise_pred_rescaled = noise_cfg * (std_text / std_cfg)
# mix with the original results from guidance by factor guidance_rescale to avoid "plain looking" images
noise_cfg = guidance_rescale * noise_pred_rescaled + \
(1 - guidance_rescale) * noise_cfg
return noise_cfg
class LyraSdXLTxt2ImgPipeline(LyraSDXLPipelineBase, StableDiffusionXLPipeline):
device = torch.device("cpu")
dtype = torch.float32
def __init__(self, device=torch.device("cuda"), dtype=torch.float16, vae_scale_factor=8, vae_scaling_factor=0.13025) -> None:
self.register_to_config(force_zeros_for_empty_prompt=True)
super().__init__(device, dtype, vae_scale_factor=vae_scale_factor, vae_scaling_factor=vae_scaling_factor)
@torch.no_grad()
def __call__(
self,
prompt: Union[str, List[str]] = None,
prompt_2: Optional[Union[str, List[str]]] = None,
height: Optional[int] = None,
width: Optional[int] = None,
num_inference_steps: int = 50,
denoising_end: Optional[float] = None,
guidance_scale: float = 5.0,
negative_prompt: Optional[Union[str, List[str]]] = None,
negative_prompt_2: Optional[Union[str, List[str]]] = None,
num_images_per_prompt: Optional[int] = 1,
eta: float = 0.0,
generator: Optional[Union[torch.Generator,
List[torch.Generator]]] = None,
latents: Optional[torch.FloatTensor] = None,
prompt_embeds: Optional[torch.FloatTensor] = None,
negative_prompt_embeds: Optional[torch.FloatTensor] = None,
pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
negative_pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
output_type: Optional[str] = "pil",
return_dict: bool = True,
callback: Optional[Callable[[
int, int, torch.FloatTensor], None]] = None,
callback_steps: int = 1,
cross_attention_kwargs: Optional[Dict[str, Any]] = None,
guidance_rescale: float = 0.0,
original_size: Optional[Tuple[int, int]] = None,
crops_coords_top_left: Tuple[int, int] = (0, 0),
target_size: Optional[Tuple[int, int]] = None,
extra_tensor_dict: Optional[Dict[str, torch.FloatTensor]] = {},
param_scale_dict: Optional[Dict[str, int]] = {},
clip_skip: Optional[int] = None
):
# 0. Default height and width to unet
height = height or self.default_sample_size * self.vae_scale_factor
width = width or self.default_sample_size * self.vae_scale_factor
original_size = original_size or (height, width)
target_size = target_size or (height, width)
# 1. Check inputs. Raise error if not correct
self.check_inputs(
prompt,
prompt_2,
height,
width,
callback_steps,
negative_prompt,
negative_prompt_2,
prompt_embeds,
negative_prompt_embeds,
pooled_prompt_embeds,
negative_pooled_prompt_embeds,
)
# 2. Define call parameters
if prompt is not None and isinstance(prompt, str):
batch_size = 1
elif prompt is not None and isinstance(prompt, list):
batch_size = len(prompt)
else:
batch_size = prompt_embeds.shape[0]
device = self._execution_device
do_classifier_free_guidance = guidance_scale > 1.0
# 3. Encode input prompt
text_encoder_lora_scale = (
cross_attention_kwargs.get(
"scale", None) if cross_attention_kwargs is not None else None
)
(
prompt_embeds,
negative_prompt_embeds,
pooled_prompt_embeds,
negative_pooled_prompt_embeds,
) = self.encode_prompt(
prompt=prompt,
prompt_2=prompt_2,
device=device,
num_images_per_prompt=num_images_per_prompt,
do_classifier_free_guidance=do_classifier_free_guidance,
negative_prompt=negative_prompt,
negative_prompt_2=negative_prompt_2,
prompt_embeds=prompt_embeds,
negative_prompt_embeds=negative_prompt_embeds,
pooled_prompt_embeds=pooled_prompt_embeds,
negative_pooled_prompt_embeds=negative_pooled_prompt_embeds,
lora_scale=text_encoder_lora_scale,
clip_skip=clip_skip
)
# 4. Prepare timesteps
self.scheduler.set_timesteps(num_inference_steps, device=device)
timesteps = self.scheduler.timesteps
# 5. Prepare latent variables
num_channels_latents = self.unet_in_channels
latents = self.prepare_latents(
batch_size * num_images_per_prompt,
num_channels_latents,
height,
width,
prompt_embeds.dtype,
device,
generator,
latents,
)
# 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. Prepare added time ids & embeddings
add_text_embeds = pooled_prompt_embeds
add_time_ids = list(
original_size + crops_coords_top_left + target_size)
add_time_ids = torch.tensor([add_time_ids], dtype=prompt_embeds.dtype)
if do_classifier_free_guidance:
prompt_embeds = torch.cat(
[negative_prompt_embeds, prompt_embeds], dim=0)
add_text_embeds = torch.cat(
[negative_pooled_prompt_embeds, add_text_embeds], dim=0)
add_time_ids = torch.cat([add_time_ids, add_time_ids], dim=0)
prompt_embeds = prompt_embeds.to(device)
add_text_embeds = add_text_embeds.to(device)
add_time_ids = add_time_ids.to(device).repeat(
batch_size * num_images_per_prompt, 1)
# 8. Denoising loop
num_warmup_steps = max(
len(timesteps) - num_inference_steps * self.scheduler.order, 0)
# 7.1 Apply denoising_end
if denoising_end is not None and type(denoising_end) == float and denoising_end > 0 and denoising_end < 1:
discrete_timestep_cutoff = int(
round(
self.scheduler.config.num_train_timesteps
- (denoising_end * self.scheduler.config.num_train_timesteps)
)
)
num_inference_steps = len(
list(filter(lambda ts: ts >= discrete_timestep_cutoff, timesteps)))
timesteps = timesteps[:num_inference_steps]
aug_emb = self._get_aug_emb(
add_time_ids, add_text_embeds, prompt_embeds.dtype)
extra_tensor_dict2 = {}
for name in extra_tensor_dict:
if name in ["fp_hidden_states", "ip_hidden_states"]:
v1, v2 = extra_tensor_dict[name][0], extra_tensor_dict[name][1]
extra_tensor_dict2[name] = torch.cat(
[v1.repeat(num_images_per_prompt, 1, 1), v2.repeat(num_images_per_prompt, 1, 1)])
else:
extra_tensor_dict2[name] = extra_tensor_dict[name]
# np.save("/workspace/prompt_embeds.npy", prompt_embeds.detach().cpu().numpy())
# prompt_embeds = torch.from_numpy(np.load("/workspace/gt_prompt_embeds.npy")).cuda()
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 = self.scheduler.scale_model_input(
latent_model_input, t)
latent_model_input = latent_model_input.permute(
0, 2, 3, 1).contiguous()
noise_pred = self.unet.forward(latent_model_input, prompt_embeds, t, aug_emb, None, None,
None, None, None, extra_tensor_dict2, param_scale_dict).permute(0, 3, 1, 2).contiguous()
# perform guidance
if 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)
if do_classifier_free_guidance and guidance_rescale > 0.0:
# Based on 3.4. in https://arxiv.org/pdf/2305.08891.pdf
noise_pred = rescale_noise_cfg(
noise_pred, noise_pred_text, guidance_rescale=guidance_rescale)
# compute the previous noisy sample x_t -> x_t-1
latents = self.scheduler.step(
noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0]
# 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)
if output_type == "latent":
return latents
image = self.vae.decode(1 / self.vae.scaling_factor * latents)
image = self.image_processor.postprocess(
image, output_type=output_type)
# Offload last model to CPU
if hasattr(self, "final_offload_hook") and self.final_offload_hook is not None:
self.final_offload_hook.offload()
return image
|