theluyuan/DiffSynth-Studio
1
0
Fork 0
mirror of https://github.com/modelscope/DiffSynth-Studio.git synced 2026-03-22 16:50:47 +00:00
Code Issues Packages Projects Releases Wiki Activity

compatibility update

Browse Source
This commit is contained in:
Artiprocher
2023-12-23 20:13:41 +08:00
parent b30d0fa412
commit 66b3e995c2
27 changed files with 1051 additions and 398 deletions
Download Patch File Download Diff File Expand all files Collapse all files

89
diffsynth/pipelines/stable_diffusion.py
View File

@@ -1,14 +1,16 @@
from ..models import ModelManager, SDTextEncoder, SDUNet, SDVAEDecoder, SDVAEEncoder
from ..controlnets.controlnet_unit import MultiControlNetManager
from ..controlnets import MultiControlNetManager, ControlNetUnit, ControlNetConfigUnit, Annotator
from ..prompts import SDPrompter
from ..schedulers import EnhancedDDIMScheduler
from .dancer import lets_dance
from typing import List
import torch
from tqdm import tqdm
from PIL import Image
import numpy as np
class SDPipeline(torch.nn.Module):
class SDImagePipeline(torch.nn.Module):
def __init__(self, device="cuda", torch_dtype=torch.float16):
super().__init__()
@@ -23,6 +25,7 @@ class SDPipeline(torch.nn.Module):
self.vae_encoder: SDVAEEncoder = None
self.controlnet: MultiControlNetManager = None
def fetch_main_models(self, model_manager: ModelManager):
self.text_encoder = model_manager.text_encoder
self.unet = model_manager.unet
@@ -31,13 +34,48 @@ class SDPipeline(torch.nn.Module):
# load textual inversion
self.prompter.load_textual_inversion(model_manager.textual_inversion_dict)
def fetch_controlnet_models(self, controlnet_units=[]):
def fetch_controlnet_models(self, model_manager: ModelManager, controlnet_config_units: List[ControlNetConfigUnit]=[]):
controlnet_units = []
for config in controlnet_config_units:
controlnet_unit = ControlNetUnit(
Annotator(config.processor_id),
model_manager.get_model_with_model_path(config.model_path),
config.scale
)
controlnet_units.append(controlnet_unit)
self.controlnet = MultiControlNetManager(controlnet_units)
def fetch_beautiful_prompt(self, model_manager: ModelManager):
if "beautiful_prompt" in model_manager.model:
self.prompter.load_beautiful_prompt(model_manager.model["beautiful_prompt"], model_manager.model_path["beautiful_prompt"])
@staticmethod
def from_model_manager(model_manager: ModelManager, controlnet_config_units: List[ControlNetConfigUnit]=[]):
pipe = SDImagePipeline(
device=model_manager.device,
torch_dtype=model_manager.torch_dtype,
)
pipe.fetch_main_models(model_manager)
pipe.fetch_beautiful_prompt(model_manager)
pipe.fetch_controlnet_models(model_manager, controlnet_config_units)
return pipe
def preprocess_image(self, image):
image = torch.Tensor(np.array(image, dtype=np.float32) * (2 / 255) - 1).permute(2, 0, 1).unsqueeze(0)
return image
def decode_image(self, latent, tiled=False, tile_size=64, tile_stride=32):
image = self.vae_decoder(latent.to(self.device), tiled=tiled, tile_size=tile_size, tile_stride=tile_stride)[0]
image = image.cpu().permute(1, 2, 0).numpy()
image = Image.fromarray(((image / 2 + 0.5).clip(0, 1) * 255).astype("uint8"))
return image
@torch.no_grad()
def __call__(
self,
@@ -45,7 +83,7 @@ class SDPipeline(torch.nn.Module):
negative_prompt="",
cfg_scale=7.5,
clip_skip=1,
init_image=None,
input_image=None,
controlnet_image=None,
denoising_strength=1.0,
height=512,
@@ -57,48 +95,43 @@ class SDPipeline(torch.nn.Module):
progress_bar_cmd=tqdm,
progress_bar_st=None,
):
# Encode prompts
prompt_emb_posi = self.prompter.encode_prompt(self.text_encoder, prompt, clip_skip=clip_skip, device=self.device)
prompt_emb_nega = self.prompter.encode_prompt(self.text_encoder, negative_prompt, clip_skip=clip_skip, device=self.device)
# Prepare scheduler
self.scheduler.set_timesteps(num_inference_steps, denoising_strength)
# Prepare latent tensors
if init_image is not None:
image = self.preprocess_image(init_image).to(device=self.device, dtype=self.torch_dtype)
if input_image is not None:
image = self.preprocess_image(input_image).to(device=self.device, dtype=self.torch_dtype)
latents = self.vae_encoder(image, tiled=tiled, tile_size=tile_size, tile_stride=tile_stride)
noise = torch.randn((1, 4, height//8, width//8), device=self.device, dtype=self.torch_dtype)
latents = self.scheduler.add_noise(latents, noise, timestep=self.scheduler.timesteps[0])
else:
latents = torch.randn((1, 4, height//8, width//8), device=self.device, dtype=self.torch_dtype)
# Encode prompts
prompt_emb_posi = self.prompter.encode_prompt(self.text_encoder, prompt, clip_skip=clip_skip, device=self.device, positive=True)
prompt_emb_nega = self.prompter.encode_prompt(self.text_encoder, negative_prompt, clip_skip=clip_skip, device=self.device, positive=False)
# Prepare ControlNets
if controlnet_image is not None:
controlnet_image = self.controlnet.process_image(controlnet_image).to(device=self.device, dtype=self.torch_dtype)
controlnet_image = controlnet_image.unsqueeze(1)
# Denoise
for progress_id, timestep in enumerate(progress_bar_cmd(self.scheduler.timesteps)):
timestep = torch.IntTensor((timestep,))[0].to(self.device)
# ControlNet
if controlnet_image is not None:
additional_res_stack_posi = self.controlnet(latents, timestep, prompt_emb_posi, controlnet_image)
additional_res_stack_nega = self.controlnet(latents, timestep, prompt_emb_nega, controlnet_image)
else:
additional_res_stack_posi = None
additional_res_stack_nega = None
# Classifier-free guidance
noise_pred_posi = self.unet(
latents, timestep, prompt_emb_posi,
additional_res_stack=additional_res_stack_posi,
tiled=tiled, tile_size=tile_size, tile_stride=tile_stride
noise_pred_posi = lets_dance(
self.unet, motion_modules=None, controlnet=self.controlnet,
sample=latents, timestep=timestep, encoder_hidden_states=prompt_emb_posi, controlnet_frames=controlnet_image,
tiled=tiled, tile_size=tile_size, tile_stride=tile_stride,
device=self.device, vram_limit_level=0
)
noise_pred_nega = self.unet(
latents, timestep, prompt_emb_nega,
additional_res_stack=additional_res_stack_nega,
tiled=tiled, tile_size=tile_size, tile_stride=tile_stride
noise_pred_nega = lets_dance(
self.unet, motion_modules=None, controlnet=self.controlnet,
sample=latents, timestep=timestep, encoder_hidden_states=prompt_emb_nega, controlnet_frames=controlnet_image,
tiled=tiled, tile_size=tile_size, tile_stride=tile_stride,
device=self.device, vram_limit_level=0
)
noise_pred = noise_pred_nega + cfg_scale * (noise_pred_posi - noise_pred_nega)
@@ -110,8 +143,6 @@ class SDPipeline(torch.nn.Module):
progress_bar_st.progress(progress_id / len(self.scheduler.timesteps))
# Decode image
image = self.vae_decoder(latents, tiled=tiled, tile_size=tile_size, tile_stride=tile_stride)[0]
image = image.cpu().permute(1, 2, 0).numpy()
image = Image.fromarray(((image / 2 + 0.5).clip(0, 1) * 255).astype("uint8"))
image = self.decode_image(latents, tiled=tiled, tile_size=tile_size, tile_stride=tile_stride)
return image