Instructions to use AiArtLab/sdxs-1b with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Diffusers
How to use AiArtLab/sdxs-1b with Diffusers:
pip install -U diffusers transformers accelerate
import torch from diffusers import DiffusionPipeline # switch to "mps" for apple devices pipe = DiffusionPipeline.from_pretrained("AiArtLab/sdxs-1b", dtype=torch.bfloat16, device_map="cuda") prompt = "sdxs-1b" image = pipe(prompt).images[0] - Notebooks
- Google Colab
- Kaggle
- Local Apps Settings
- Draw Things
- DiffusionBee
2601
Browse files- pipeline_sdxs-Copy2.py +196 -0
- pipeline_sdxs.py +94 -42
pipeline_sdxs-Copy2.py
ADDED
|
@@ -0,0 +1,196 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import torch
|
| 2 |
+
import numpy as np
|
| 3 |
+
from PIL import Image
|
| 4 |
+
from typing import List, Union, Optional, Tuple
|
| 5 |
+
from dataclasses import dataclass
|
| 6 |
+
|
| 7 |
+
from diffusers import DiffusionPipeline
|
| 8 |
+
from diffusers.utils import BaseOutput
|
| 9 |
+
from tqdm import tqdm
|
| 10 |
+
|
| 11 |
+
@dataclass
|
| 12 |
+
class SdxsPipelineOutput(BaseOutput):
|
| 13 |
+
images: Union[List[Image.Image], np.ndarray]
|
| 14 |
+
|
| 15 |
+
class SdxsPipeline(DiffusionPipeline):
|
| 16 |
+
def __init__(self, vae, text_encoder, tokenizer, unet, scheduler):
|
| 17 |
+
super().__init__()
|
| 18 |
+
self.register_modules(
|
| 19 |
+
vae=vae,
|
| 20 |
+
text_encoder=text_encoder,
|
| 21 |
+
tokenizer=tokenizer,
|
| 22 |
+
unet=unet,
|
| 23 |
+
scheduler=scheduler
|
| 24 |
+
)
|
| 25 |
+
self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
|
| 26 |
+
|
| 27 |
+
# --- ВСПОМОГАТЕЛЬНАЯ ФУНКЦИЯ ДЛЯ ПОДГОТОВКИ ИЗОБРАЖЕНИЯ (Img2Img) ---
|
| 28 |
+
def preprocess_image(self, image: Image.Image, width: int, height: int):
|
| 29 |
+
"""Ресайз и центрированный кроп изображения под нужный размер"""
|
| 30 |
+
height = (height // self.vae_scale_factor) * self.vae_scale_factor // 2
|
| 31 |
+
width= (width // self.vae_scale_factor) * self.vae_scale_factor // 2
|
| 32 |
+
w, h = image.size
|
| 33 |
+
aspect_ratio = width / height
|
| 34 |
+
if w / h > aspect_ratio:
|
| 35 |
+
new_w = int(h * aspect_ratio)
|
| 36 |
+
left = (w - new_w) // 2
|
| 37 |
+
image = image.crop((left, 0, left + new_w, h))
|
| 38 |
+
else:
|
| 39 |
+
new_h = int(w / aspect_ratio)
|
| 40 |
+
top = (h - new_h) // 2
|
| 41 |
+
image = image.crop((0, top, w, top + new_h))
|
| 42 |
+
|
| 43 |
+
image = image.resize((width, height ), resample=Image.LANCZOS)
|
| 44 |
+
image = np.array(image).astype(np.float32) / 255.0
|
| 45 |
+
image = image[None].transpose(0, 3, 1, 2) # [1, C, H, W]
|
| 46 |
+
image = torch.from_numpy(image)
|
| 47 |
+
return 2.0 * image - 1.0 # В диапазон [-1, 1]
|
| 48 |
+
|
| 49 |
+
def encode_prompt(self, prompt, negative_prompt, device, dtype):
|
| 50 |
+
def get_single_encode(texts, is_negative=False):
|
| 51 |
+
if texts is None or texts == "":
|
| 52 |
+
hidden_dim = self.text_encoder.config.hidden_size
|
| 53 |
+
shape = (1, self.text_encoder.config.max_position_embeddings, hidden_dim)
|
| 54 |
+
emb = torch.zeros(shape, dtype=dtype, device=device)
|
| 55 |
+
mask = torch.ones((1, self.text_encoder.config.max_position_embeddings), dtype=torch.int64, device=device)
|
| 56 |
+
return emb, mask
|
| 57 |
+
|
| 58 |
+
if isinstance(texts, str):
|
| 59 |
+
texts = [texts]
|
| 60 |
+
|
| 61 |
+
with torch.no_grad():
|
| 62 |
+
toks = self.tokenizer(
|
| 63 |
+
texts,
|
| 64 |
+
padding="max_length",
|
| 65 |
+
max_length=self.text_encoder.config.max_position_embeddings,
|
| 66 |
+
truncation=True,
|
| 67 |
+
return_tensors="pt"
|
| 68 |
+
).to(device)
|
| 69 |
+
|
| 70 |
+
outputs = self.text_encoder(
|
| 71 |
+
input_ids=toks.input_ids,
|
| 72 |
+
attention_mask=toks.attention_mask,
|
| 73 |
+
output_hidden_states=True
|
| 74 |
+
)
|
| 75 |
+
|
| 76 |
+
layer_index = -2
|
| 77 |
+
prompt_embeds = outputs.hidden_states[layer_index]
|
| 78 |
+
final_layer_norm = self.text_encoder.text_model.final_layer_norm
|
| 79 |
+
prompt_embeds = final_layer_norm(prompt_embeds)
|
| 80 |
+
|
| 81 |
+
return prompt_embeds, toks.attention_mask
|
| 82 |
+
|
| 83 |
+
pos_embeds, pos_mask = get_single_encode(prompt)
|
| 84 |
+
neg_embeds, neg_mask = get_single_encode(negative_prompt, is_negative=True)
|
| 85 |
+
|
| 86 |
+
batch_size = pos_embeds.shape[0]
|
| 87 |
+
if neg_embeds.shape[0] != batch_size:
|
| 88 |
+
neg_embeds = neg_embeds.repeat(batch_size, 1, 1)
|
| 89 |
+
neg_mask = neg_mask.repeat(batch_size, 1)
|
| 90 |
+
|
| 91 |
+
text_embeddings = torch.cat([neg_embeds, pos_embeds], dim=0)
|
| 92 |
+
final_mask = torch.cat([neg_mask, pos_mask], dim=0)
|
| 93 |
+
|
| 94 |
+
return text_embeddings.to(dtype=dtype), final_mask.to(dtype=torch.int64)
|
| 95 |
+
|
| 96 |
+
@torch.no_grad()
|
| 97 |
+
def __call__(
|
| 98 |
+
self,
|
| 99 |
+
prompt: Union[str, List[str]],
|
| 100 |
+
image: Optional[Union[Image.Image, List[Image.Image]]] = None, # Добавлен параметр изображения
|
| 101 |
+
coef: float = 0.5, # Коэффициент влияния (strength): 1.0 - полный шум, 0.0 - оригинал
|
| 102 |
+
negative_prompt: Optional[Union[str, List[str]]] = None,
|
| 103 |
+
height: int = 1024,
|
| 104 |
+
width: int = 1024,
|
| 105 |
+
num_inference_steps: int = 40,
|
| 106 |
+
guidance_scale: float = 4.0,
|
| 107 |
+
generator: Optional[torch.Generator] = None,
|
| 108 |
+
output_type: str = "pil",
|
| 109 |
+
return_dict: bool = True,
|
| 110 |
+
**kwargs,
|
| 111 |
+
):
|
| 112 |
+
device = self.device
|
| 113 |
+
self.vae.to(device)
|
| 114 |
+
vae_scaling_factor = getattr(self.vae.config, "scaling_factor", 1.0)
|
| 115 |
+
vae_shift_factor = getattr(self.vae.config, "shift_factor", 0.0)
|
| 116 |
+
device = self.device
|
| 117 |
+
dtype = self.unet.dtype
|
| 118 |
+
|
| 119 |
+
# 1. Encode Prompt
|
| 120 |
+
text_embeddings, attention_mask = self.encode_prompt(prompt, negative_prompt, device, dtype)
|
| 121 |
+
batch_size = 1 if isinstance(prompt, str) else len(prompt)
|
| 122 |
+
|
| 123 |
+
# 2. Настройка таймстепов
|
| 124 |
+
self.scheduler.set_timesteps(num_inference_steps, device=device)
|
| 125 |
+
timesteps = self.scheduler.timesteps # Обычно от 1.0 до 0.001
|
| 126 |
+
|
| 127 |
+
# --- ИСПРАВЛЕННАЯ ЛОГИКА IMG2IMG ---
|
| 128 |
+
if image is not None:
|
| 129 |
+
# Превращаем картинку в латенты
|
| 130 |
+
image_tensor = self.preprocess_image(image, width, height).to(device=device, dtype=self.vae.dtype)
|
| 131 |
+
|
| 132 |
+
# ВАЖНО: Кодируем и применяем масштабирование VAE СРАЗУ
|
| 133 |
+
latents_clean = self.vae.encode(image_tensor).latent_dist.sample(generator=generator)
|
| 134 |
+
latents_clean = (latents_clean - vae_shift_factor) / vae_scaling_factor
|
| 135 |
+
latents_clean = latents_clean.to(dtype=dtype)
|
| 136 |
+
|
| 137 |
+
# Определяем начальный таймстеп t_start.
|
| 138 |
+
# coef=1.0 -> t=1.0 (шум), coef=0.0 -> t=0.0 (картинка)
|
| 139 |
+
t_start = coef
|
| 140 |
+
|
| 141 |
+
# Находим индекс шага в расписании шедулера, который ближе всего к t_start
|
| 142 |
+
# Это исключает "одинаковость" при разной силе
|
| 143 |
+
idx = (timesteps - t_start).abs().argmin().item()
|
| 144 |
+
timesteps = timesteps[idx:]
|
| 145 |
+
|
| 146 |
+
# Добавляем шум строго по формуле Flow Matching: x_t = (1-t)x_0 + t*eps
|
| 147 |
+
noise = torch.randn(latents_clean.shape, generator=generator, device=device, dtype=dtype)
|
| 148 |
+
actual_t = timesteps[0] # Берем точное значение из шедулера
|
| 149 |
+
latents = (1.0 - actual_t) * latents_clean + actual_t * noise
|
| 150 |
+
else:
|
| 151 |
+
# TXT2IMG: Начинаем с чистого шума (t=1.0)
|
| 152 |
+
latents = torch.randn(
|
| 153 |
+
(batch_size, self.unet.config.in_channels, height // self.vae_scale_factor, width // self.vae_scale_factor),
|
| 154 |
+
generator=generator, device=device, dtype=dtype
|
| 155 |
+
)
|
| 156 |
+
|
| 157 |
+
# 3. Denoising Loop
|
| 158 |
+
for t in tqdm(timesteps, desc="Sampling"):
|
| 159 |
+
latent_model_input = torch.cat([latents] * 2) if guidance_scale > 1 else latents
|
| 160 |
+
|
| 161 |
+
# Predict
|
| 162 |
+
model_out = self.unet(
|
| 163 |
+
latent_model_input,
|
| 164 |
+
t,
|
| 165 |
+
encoder_hidden_states=text_embeddings,
|
| 166 |
+
encoder_attention_mask=attention_mask,
|
| 167 |
+
return_dict=False,
|
| 168 |
+
)[0]
|
| 169 |
+
|
| 170 |
+
# CFG
|
| 171 |
+
if guidance_scale > 1:
|
| 172 |
+
flow_uncond, flow_cond = model_out.chunk(2)
|
| 173 |
+
model_out = flow_uncond + guidance_scale * (flow_cond - flow_uncond)
|
| 174 |
+
|
| 175 |
+
# Step (Эйлеров шаг: latents = latents + (t_next - t_curr) * model_out)
|
| 176 |
+
latents = self.scheduler.step(model_out, t, latents, return_dict=False)[0]
|
| 177 |
+
|
| 178 |
+
# 5. Decode
|
| 179 |
+
if output_type == "latent":
|
| 180 |
+
return SdxsPipelineOutput(images=latents)
|
| 181 |
+
|
| 182 |
+
latents = latents * vae_scaling_factor + vae_shift_factor
|
| 183 |
+
image = self.vae.decode(latents.to(self.vae.dtype), return_dict=False)[0]
|
| 184 |
+
|
| 185 |
+
# Пост-процессинг
|
| 186 |
+
image = (image / 2 + 0.5).clamp(0, 1)
|
| 187 |
+
image = image.cpu().permute(0, 2, 3, 1).float().numpy()
|
| 188 |
+
|
| 189 |
+
if output_type == "pil":
|
| 190 |
+
image = (image * 255).round().astype("uint8")
|
| 191 |
+
image = [Image.fromarray(img) for img in image]
|
| 192 |
+
|
| 193 |
+
if not return_dict:
|
| 194 |
+
return image
|
| 195 |
+
|
| 196 |
+
return SdxsPipelineOutput(images=image)
|
pipeline_sdxs.py
CHANGED
|
@@ -97,8 +97,8 @@ class SdxsPipeline(DiffusionPipeline):
|
|
| 97 |
def __call__(
|
| 98 |
self,
|
| 99 |
prompt: Union[str, List[str]],
|
| 100 |
-
image: Optional[Union[Image.Image, List[Image.Image]]] = None,
|
| 101 |
-
coef: float = 0.5,
|
| 102 |
negative_prompt: Optional[Union[str, List[str]]] = None,
|
| 103 |
height: int = 1024,
|
| 104 |
width: int = 1024,
|
|
@@ -110,55 +110,106 @@ class SdxsPipeline(DiffusionPipeline):
|
|
| 110 |
**kwargs,
|
| 111 |
):
|
| 112 |
device = self.device
|
| 113 |
-
self.vae.to(device)
|
| 114 |
-
vae_scaling_factor = getattr(self.vae.config, "scaling_factor", 1.0)
|
| 115 |
-
vae_shift_factor = getattr(self.vae.config, "shift_factor", 0.0)
|
| 116 |
-
device = self.device
|
| 117 |
dtype = self.unet.dtype
|
| 118 |
|
| 119 |
# 1. Encode Prompt
|
| 120 |
-
text_embeddings, attention_mask = self.encode_prompt(
|
|
|
|
|
|
|
| 121 |
batch_size = 1 if isinstance(prompt, str) else len(prompt)
|
| 122 |
-
|
| 123 |
# 2. Настройка таймстепов
|
| 124 |
self.scheduler.set_timesteps(num_inference_steps, device=device)
|
| 125 |
-
timesteps = self.scheduler.timesteps
|
| 126 |
-
|
| 127 |
-
#
|
| 128 |
if image is not None:
|
| 129 |
-
# П
|
| 130 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 131 |
|
| 132 |
-
#
|
| 133 |
latents_clean = self.vae.encode(image_tensor).latent_dist.sample(generator=generator)
|
|
|
|
|
|
|
| 134 |
latents_clean = (latents_clean - vae_shift_factor) / vae_scaling_factor
|
| 135 |
latents_clean = latents_clean.to(dtype=dtype)
|
| 136 |
-
|
| 137 |
-
# Определяем начальный таймстеп
|
| 138 |
-
#
|
| 139 |
-
|
| 140 |
-
|
| 141 |
-
|
| 142 |
-
#
|
| 143 |
-
|
| 144 |
-
|
| 145 |
-
|
| 146 |
-
#
|
| 147 |
-
|
| 148 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 149 |
latents = (1.0 - actual_t) * latents_clean + actual_t * noise
|
|
|
|
| 150 |
else:
|
| 151 |
-
# TXT2IMG:
|
|
|
|
|
|
|
|
|
|
| 152 |
latents = torch.randn(
|
| 153 |
-
(batch_size, self.unet.config.in_channels,
|
| 154 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
| 155 |
)
|
| 156 |
-
|
| 157 |
# 3. Denoising Loop
|
| 158 |
-
for t in tqdm(timesteps, desc="Sampling"):
|
|
|
|
| 159 |
latent_model_input = torch.cat([latents] * 2) if guidance_scale > 1 else latents
|
| 160 |
|
| 161 |
-
# Predict
|
| 162 |
model_out = self.unet(
|
| 163 |
latent_model_input,
|
| 164 |
t,
|
|
@@ -166,31 +217,32 @@ device = self.device
|
|
| 166 |
encoder_attention_mask=attention_mask,
|
| 167 |
return_dict=False,
|
| 168 |
)[0]
|
| 169 |
-
|
| 170 |
# CFG
|
| 171 |
if guidance_scale > 1:
|
| 172 |
flow_uncond, flow_cond = model_out.chunk(2)
|
| 173 |
model_out = flow_uncond + guidance_scale * (flow_cond - flow_uncond)
|
| 174 |
-
|
| 175 |
-
#
|
| 176 |
latents = self.scheduler.step(model_out, t, latents, return_dict=False)[0]
|
| 177 |
-
|
| 178 |
-
#
|
| 179 |
if output_type == "latent":
|
| 180 |
return SdxsPipelineOutput(images=latents)
|
| 181 |
-
|
|
|
|
| 182 |
latents = latents * vae_scaling_factor + vae_shift_factor
|
| 183 |
image = self.vae.decode(latents.to(self.vae.dtype), return_dict=False)[0]
|
| 184 |
|
| 185 |
# Пост-процессинг
|
| 186 |
image = (image / 2 + 0.5).clamp(0, 1)
|
| 187 |
image = image.cpu().permute(0, 2, 3, 1).float().numpy()
|
| 188 |
-
|
| 189 |
if output_type == "pil":
|
| 190 |
image = (image * 255).round().astype("uint8")
|
| 191 |
image = [Image.fromarray(img) for img in image]
|
| 192 |
-
|
| 193 |
if not return_dict:
|
| 194 |
return image
|
| 195 |
-
|
| 196 |
return SdxsPipelineOutput(images=image)
|
|
|
|
| 97 |
def __call__(
|
| 98 |
self,
|
| 99 |
prompt: Union[str, List[str]],
|
| 100 |
+
image: Optional[Union[Image.Image, List[Image.Image]]] = None,
|
| 101 |
+
coef: float = 0.5, # strength: 1.0 - полный шум, 0.0 - оригинал
|
| 102 |
negative_prompt: Optional[Union[str, List[str]]] = None,
|
| 103 |
height: int = 1024,
|
| 104 |
width: int = 1024,
|
|
|
|
| 110 |
**kwargs,
|
| 111 |
):
|
| 112 |
device = self.device
|
|
|
|
|
|
|
|
|
|
|
|
|
| 113 |
dtype = self.unet.dtype
|
| 114 |
|
| 115 |
# 1. Encode Prompt
|
| 116 |
+
text_embeddings, attention_mask = self.encode_prompt(
|
| 117 |
+
prompt, negative_prompt, device, dtype
|
| 118 |
+
)
|
| 119 |
batch_size = 1 if isinstance(prompt, str) else len(prompt)
|
| 120 |
+
|
| 121 |
# 2. Настройка таймстепов
|
| 122 |
self.scheduler.set_timesteps(num_inference_steps, device=device)
|
| 123 |
+
timesteps = self.scheduler.timesteps
|
| 124 |
+
|
| 125 |
+
# ИСПРАВЛЕННАЯ ЛОГИКА IMG2IMG
|
| 126 |
if image is not None:
|
| 127 |
+
# Подготовка изображения
|
| 128 |
+
if isinstance(image, Image.Image):
|
| 129 |
+
image = [image]
|
| 130 |
+
|
| 131 |
+
# Обработка каждого изображения
|
| 132 |
+
image_tensors = []
|
| 133 |
+
for img in image:
|
| 134 |
+
img_tensor = self.preprocess_image(img, width, height)
|
| 135 |
+
image_tensors.append(img_tensor)
|
| 136 |
+
|
| 137 |
+
image_tensor = torch.cat(image_tensors, dim=0).to(
|
| 138 |
+
device=device, dtype=self.vae.dtype
|
| 139 |
+
)
|
| 140 |
+
|
| 141 |
+
# Проверка batch size
|
| 142 |
+
if image_tensor.shape[0] != batch_size:
|
| 143 |
+
if image_tensor.shape[0] == 1:
|
| 144 |
+
image_tensor = image_tensor.repeat(batch_size, 1, 1, 1)
|
| 145 |
+
else:
|
| 146 |
+
raise ValueError(
|
| 147 |
+
f"Batch size mismatch: prompt has {batch_size} elements, "
|
| 148 |
+
f"but image has {image_tensor.shape[0]} elements"
|
| 149 |
+
)
|
| 150 |
|
| 151 |
+
# Кодируем в латенты
|
| 152 |
latents_clean = self.vae.encode(image_tensor).latent_dist.sample(generator=generator)
|
| 153 |
+
vae_scaling_factor = getattr(self.vae.config, "scaling_factor", 1.0)
|
| 154 |
+
vae_shift_factor = getattr(self.vae.config, "shift_factor", 0.0)
|
| 155 |
latents_clean = (latents_clean - vae_shift_factor) / vae_scaling_factor
|
| 156 |
latents_clean = latents_clean.to(dtype=dtype)
|
| 157 |
+
|
| 158 |
+
# Определяем начальный таймстеп
|
| 159 |
+
# Flow matching: x_t = (1-t)x_0 + t*eps
|
| 160 |
+
# coef=1.0 -> t=1.0 (полный шум)
|
| 161 |
+
# coef=0.0 -> t=0.0 (оригинал)
|
| 162 |
+
|
| 163 |
+
# Если scheduler использует [0, 1] диапазон
|
| 164 |
+
t_start = coef # или 1.0 - coef в зависимости от направления
|
| 165 |
+
|
| 166 |
+
# Находим ближайший шаг в расписании
|
| 167 |
+
# Некоторые scheduler имеют reversed порядок (от 1 к 0)
|
| 168 |
+
if timesteps[0] > timesteps[-1]:
|
| 169 |
+
# Обычный порядок: от шума к данным
|
| 170 |
+
idx = torch.searchsorted(timesteps, t_start, right=True)
|
| 171 |
+
idx = min(idx, len(timesteps) - 1)
|
| 172 |
+
else:
|
| 173 |
+
# Обратный порядок: от данных к шуму
|
| 174 |
+
idx = torch.searchsorted(timesteps.flip(0), 1.0 - t_start, right=True)
|
| 175 |
+
idx = len(timesteps) - 1 - min(idx, len(timesteps) - 1)
|
| 176 |
+
|
| 177 |
+
timesteps = timesteps[idx:]
|
| 178 |
+
|
| 179 |
+
# Генерируем шум и применя flow matching формулу
|
| 180 |
+
noise = torch.randn(
|
| 181 |
+
latents_clean.shape,
|
| 182 |
+
generator=generator,
|
| 183 |
+
device=device,
|
| 184 |
+
dtype=dtype
|
| 185 |
+
)
|
| 186 |
+
|
| 187 |
+
# Берём актуальное время из scheduler
|
| 188 |
+
actual_t = timesteps[0].item() if isinstance(timesteps[0], torch.Tensor) else timesteps[0]
|
| 189 |
+
|
| 190 |
+
# Применяем формулу flow matching
|
| 191 |
latents = (1.0 - actual_t) * latents_clean + actual_t * noise
|
| 192 |
+
|
| 193 |
else:
|
| 194 |
+
# TXT2IMG: начинаем с чистого шума
|
| 195 |
+
vae_scaling_factor = getattr(self.vae.config, "scaling_factor", 1.0)
|
| 196 |
+
vae_shift_factor = getattr(self.vae.config, "shift_factor", 0.0)
|
| 197 |
+
|
| 198 |
latents = torch.randn(
|
| 199 |
+
(batch_size, self.unet.config.in_channels,
|
| 200 |
+
height // self.vae_scale_factor,
|
| 201 |
+
width // self.vae_scale_factor),
|
| 202 |
+
generator=generator,
|
| 203 |
+
device=device,
|
| 204 |
+
dtype=dtype
|
| 205 |
)
|
| 206 |
+
|
| 207 |
# 3. Denoising Loop
|
| 208 |
+
for i, t in enumerate(tqdm(timesteps, desc="Sampling")):
|
| 209 |
+
# CFG preparation
|
| 210 |
latent_model_input = torch.cat([latents] * 2) if guidance_scale > 1 else latents
|
| 211 |
|
| 212 |
+
# Predict flow
|
| 213 |
model_out = self.unet(
|
| 214 |
latent_model_input,
|
| 215 |
t,
|
|
|
|
| 217 |
encoder_attention_mask=attention_mask,
|
| 218 |
return_dict=False,
|
| 219 |
)[0]
|
| 220 |
+
|
| 221 |
# CFG
|
| 222 |
if guidance_scale > 1:
|
| 223 |
flow_uncond, flow_cond = model_out.chunk(2)
|
| 224 |
model_out = flow_uncond + guidance_scale * (flow_cond - flow_uncond)
|
| 225 |
+
|
| 226 |
+
# Euler step для flow matching
|
| 227 |
latents = self.scheduler.step(model_out, t, latents, return_dict=False)[0]
|
| 228 |
+
|
| 229 |
+
# 4. Decode
|
| 230 |
if output_type == "latent":
|
| 231 |
return SdxsPipelineOutput(images=latents)
|
| 232 |
+
|
| 233 |
+
# Масштабируем обратно
|
| 234 |
latents = latents * vae_scaling_factor + vae_shift_factor
|
| 235 |
image = self.vae.decode(latents.to(self.vae.dtype), return_dict=False)[0]
|
| 236 |
|
| 237 |
# Пост-процессинг
|
| 238 |
image = (image / 2 + 0.5).clamp(0, 1)
|
| 239 |
image = image.cpu().permute(0, 2, 3, 1).float().numpy()
|
| 240 |
+
|
| 241 |
if output_type == "pil":
|
| 242 |
image = (image * 255).round().astype("uint8")
|
| 243 |
image = [Image.fromarray(img) for img in image]
|
| 244 |
+
|
| 245 |
if not return_dict:
|
| 246 |
return image
|
| 247 |
+
|
| 248 |
return SdxsPipelineOutput(images=image)
|