Spaces:
Running
on
Zero
Running
on
Zero
File size: 7,721 Bytes
703e263 |
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 |
from ..models import SDTextEncoder, SDUNet, SDVAEDecoder, SDVAEEncoder, SDIpAdapter, IpAdapterCLIPImageEmbedder
from ..models.model_manager import ModelManager
from ..controlnets import MultiControlNetManager, ControlNetUnit, ControlNetConfigUnit, Annotator
from ..prompters import SDPrompter
from ..schedulers import EnhancedDDIMScheduler
from .base import BasePipeline
from .dancer import lets_dance
from typing import List
import torch
from tqdm import tqdm
class SDImagePipeline(BasePipeline):
def __init__(self, device="cuda", torch_dtype=torch.float16):
super().__init__(device=device, torch_dtype=torch_dtype)
self.scheduler = EnhancedDDIMScheduler()
self.prompter = SDPrompter()
# models
self.text_encoder: SDTextEncoder = None
self.unet: SDUNet = None
self.vae_decoder: SDVAEDecoder = None
self.vae_encoder: SDVAEEncoder = None
self.controlnet: MultiControlNetManager = None
self.ipadapter_image_encoder: IpAdapterCLIPImageEmbedder = None
self.ipadapter: SDIpAdapter = None
def denoising_model(self):
return self.unet
def fetch_models(self, model_manager: ModelManager, controlnet_config_units: List[ControlNetConfigUnit]=[], prompt_refiner_classes=[]):
# Main models
self.text_encoder = model_manager.fetch_model("sd_text_encoder")
self.unet = model_manager.fetch_model("sd_unet")
self.vae_decoder = model_manager.fetch_model("sd_vae_decoder")
self.vae_encoder = model_manager.fetch_model("sd_vae_encoder")
self.prompter.fetch_models(self.text_encoder)
self.prompter.load_prompt_refiners(model_manager, prompt_refiner_classes)
# ControlNets
controlnet_units = []
for config in controlnet_config_units:
controlnet_unit = ControlNetUnit(
Annotator(config.processor_id, device=self.device),
model_manager.fetch_model("sd_controlnet", config.model_path),
config.scale
)
controlnet_units.append(controlnet_unit)
self.controlnet = MultiControlNetManager(controlnet_units)
# IP-Adapters
self.ipadapter = model_manager.fetch_model("sd_ipadapter")
self.ipadapter_image_encoder = model_manager.fetch_model("sd_ipadapter_clip_image_encoder")
@staticmethod
def from_model_manager(model_manager: ModelManager, controlnet_config_units: List[ControlNetConfigUnit]=[], prompt_refiner_classes=[]):
pipe = SDImagePipeline(
device=model_manager.device,
torch_dtype=model_manager.torch_dtype,
)
pipe.fetch_models(model_manager, controlnet_config_units, prompt_refiner_classes=[])
return pipe
def encode_image(self, image, tiled=False, tile_size=64, tile_stride=32):
latents = self.vae_encoder(image, tiled=tiled, tile_size=tile_size, tile_stride=tile_stride)
return latents
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)
image = self.vae_output_to_image(image)
return image
def encode_prompt(self, prompt, clip_skip=1, positive=True):
prompt_emb = self.prompter.encode_prompt(prompt, clip_skip=clip_skip, device=self.device, positive=positive)
return {"encoder_hidden_states": prompt_emb}
def prepare_extra_input(self, latents=None):
return {}
@torch.no_grad()
def __call__(
self,
prompt,
local_prompts=[],
masks=[],
mask_scales=[],
negative_prompt="",
cfg_scale=7.5,
clip_skip=1,
input_image=None,
ipadapter_images=None,
ipadapter_scale=1.0,
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,
):
# Tiler parameters
tiler_kwargs = {"tiled": tiled, "tile_size": tile_size, "tile_stride": tile_stride}
# Prepare scheduler
self.scheduler.set_timesteps(num_inference_steps, denoising_strength)
# Prepare latent tensors
if input_image is not None:
image = self.preprocess_image(input_image).to(device=self.device, dtype=self.torch_dtype)
latents = self.encode_image(image, **tiler_kwargs)
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.encode_prompt(prompt, clip_skip=clip_skip, positive=True)
prompt_emb_nega = self.encode_prompt(negative_prompt, clip_skip=clip_skip, positive=False)
prompt_emb_locals = [self.encode_prompt(prompt_local, clip_skip=clip_skip, positive=True) for prompt_local in local_prompts]
# IP-Adapter
if ipadapter_images is not None:
ipadapter_image_encoding = self.ipadapter_image_encoder(ipadapter_images)
ipadapter_kwargs_list_posi = {"ipadapter_kwargs_list": self.ipadapter(ipadapter_image_encoding, scale=ipadapter_scale)}
ipadapter_kwargs_list_nega = {"ipadapter_kwargs_list": self.ipadapter(torch.zeros_like(ipadapter_image_encoding))}
else:
ipadapter_kwargs_list_posi, ipadapter_kwargs_list_nega = {"ipadapter_kwargs_list": {}}, {"ipadapter_kwargs_list": {}}
# 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)
controlnet_kwargs = {"controlnet_frames": controlnet_image}
else:
controlnet_kwargs = {"controlnet_frames": None}
# Denoise
for progress_id, timestep in enumerate(progress_bar_cmd(self.scheduler.timesteps)):
timestep = timestep.unsqueeze(0).to(self.device)
# Classifier-free guidance
inference_callback = lambda prompt_emb_posi: lets_dance(
self.unet, motion_modules=None, controlnet=self.controlnet,
sample=latents, timestep=timestep,
**prompt_emb_posi, **controlnet_kwargs, **tiler_kwargs, **ipadapter_kwargs_list_posi,
device=self.device,
)
noise_pred_posi = self.control_noise_via_local_prompts(prompt_emb_posi, prompt_emb_locals, masks, mask_scales, inference_callback)
noise_pred_nega = lets_dance(
self.unet, motion_modules=None, controlnet=self.controlnet,
sample=latents, timestep=timestep, **prompt_emb_nega, **controlnet_kwargs, **tiler_kwargs, **ipadapter_kwargs_list_nega,
device=self.device,
)
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.decode_image(latents, tiled=tiled, tile_size=tile_size, tile_stride=tile_stride)
return image
|