mirror of
https://github.com/modelscope/DiffSynth-Studio.git
synced 2026-03-19 14:58:12 +00:00
118 lines
4.8 KiB
Python
118 lines
4.8 KiB
Python
from ..models import ModelManager, SDTextEncoder, SDUNet, SDVAEDecoder, SDVAEEncoder
|
|
from ..controlnets.controlnet_unit import MultiControlNetManager
|
|
from ..prompts import SDPrompter
|
|
from ..schedulers import EnhancedDDIMScheduler
|
|
import torch
|
|
from tqdm import tqdm
|
|
from PIL import Image
|
|
import numpy as np
|
|
|
|
|
|
class SDPipeline(torch.nn.Module):
|
|
|
|
def __init__(self, device="cuda", torch_dtype=torch.float16):
|
|
super().__init__()
|
|
self.scheduler = EnhancedDDIMScheduler()
|
|
self.prompter = SDPrompter()
|
|
self.device = device
|
|
self.torch_dtype = torch_dtype
|
|
# models
|
|
self.text_encoder: SDTextEncoder = None
|
|
self.unet: SDUNet = None
|
|
self.vae_decoder: SDVAEDecoder = None
|
|
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
|
|
self.vae_decoder = model_manager.vae_decoder
|
|
self.vae_encoder = model_manager.vae_encoder
|
|
# load textual inversion
|
|
self.prompter.load_textual_inversion(model_manager.textual_inversion_dict)
|
|
|
|
def fetch_controlnet_models(self, controlnet_units=[]):
|
|
self.controlnet = MultiControlNetManager(controlnet_units)
|
|
|
|
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
|
|
|
|
@torch.no_grad()
|
|
def __call__(
|
|
self,
|
|
prompt,
|
|
negative_prompt="",
|
|
cfg_scale=7.5,
|
|
clip_skip=1,
|
|
init_image=None,
|
|
controlnet_image=None,
|
|
denoising_strength=1.0,
|
|
height=512,
|
|
width=512,
|
|
num_inference_steps=20,
|
|
tiled=False,
|
|
tile_size=64,
|
|
tile_stride=32,
|
|
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)
|
|
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)
|
|
|
|
# Prepare ControlNets
|
|
if controlnet_image is not None:
|
|
controlnet_image = self.controlnet.process_image(controlnet_image).to(device=self.device, dtype=self.torch_dtype)
|
|
|
|
# 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_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 = noise_pred_nega + cfg_scale * (noise_pred_posi - noise_pred_nega)
|
|
|
|
# DDIM
|
|
latents = self.scheduler.step(noise_pred, timestep, latents)
|
|
|
|
# UI
|
|
if progress_bar_st is not None:
|
|
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"))
|
|
|
|
return image
|