1.3b

README.md

@@ -8,11 +8,11 @@ datasets:
 # Simple Diffusion XS
 
 *XS Size, Excess Quality*
-![promo](media/result_grid_base.jpg)
+![promo](media/girl.jpg)
 
 At AiArtLab, we strive to create a free, compact and fast model that can be trained on consumer graphics cards.
 
-- Unet: 1.16b parameters
+- Unet: 1.3b parameters
 - Clip: [LongCLIP with 248 tokens](https://huggingface.co/zer0int/CLIP-KO-LITE-TypoAttack-Attn-Dropout-ViT-L-14)
 - VAE: 16ch16x(8x-enc/16x-dec)
 - Speed: Sampling 100%|██████████| 40/40 [00:01<00:00, 30.72it/s] (1024x1280)

pipeline_sdxs-Copy1.py

@@ -1,335 +0,0 @@
-import torch
-import numpy as np
-from PIL import Image
-from typing import List, Union, Optional, Tuple
-from dataclasses import dataclass
-
-from diffusers import DiffusionPipeline
-from diffusers.utils import BaseOutput
-from tqdm import tqdm
-
-@dataclass
-class SdxsPipelineOutput(BaseOutput):
-    images: Union[List[Image.Image], np.ndarray]
-
-class SdxsPipeline(DiffusionPipeline):
-    def __init__(self, vae, text_encoder, tokenizer, unet, scheduler):
-        super().__init__()
-        self.register_modules(
-            vae=vae, 
-            text_encoder=text_encoder, 
-            tokenizer=tokenizer, 
-            unet=unet, 
-            scheduler=scheduler
-        )
-        self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
-        
-    def create_frequency_soft_cutoff_mask(self, height: int, width: int, cutoff_radius: float, 
-                                        transition_width: float = 5.0, device: torch.device = None) -> torch.Tensor:
-        """Создает плавную маску частотного среза для сохранения структуры."""
-        if device is None:
-            device = torch.device('cpu')
-        
-        u = torch.arange(height, device=device)
-        v = torch.arange(width, device=device)
-        u, v = torch.meshgrid(u, v, indexing='ij')
-        
-        center_u, center_v = height // 2, width // 2
-        frequency_radius = torch.sqrt((u - center_u)**2 + (v - center_v)**2)
-        
-        mask = torch.exp(-(frequency_radius - cutoff_radius)**2 / (2 * transition_width**2))
-        mask = torch.where(frequency_radius <= cutoff_radius, torch.ones_like(mask), mask)
-        
-        return mask
-
-    def generate_structured_noise(
-        self,
-        image_latents: torch.Tensor,
-        cutoff_radius: Optional[float] = None,
-        transition_width: float = 2.0,
-        noise_std: float = 1.0,
-    ) -> torch.Tensor:
-        """
-        Генерирует структурированный шум для латентов с сохранением низкочастотной структуры.
-        
-        Args:
-            image_latents: Чистые латенты изображения [B, C, H, W]
-            cutoff_radius: Радиус среза частот (None = авто-расчет на основе coef)
-            transition_width: Ширина плавного перехода
-            noise_std: Стандартное отклонение шума
-            
-        Returns:
-            Структурированный шум с той же размерностью
-        """
-        batch_size, channels, height, width = image_latents.shape
-        device = image_latents.device
-        dtype = image_latents.dtype
-        
-        # Автоматический расчет cutoff_radius если не задан
-        if cutoff_radius is None:
-            # Сохраняем больше низких частот для лучшей структуры
-            max_radius = min(height, width) / 2
-            cutoff_radius = max_radius * 0.7  # Сохраняем 70% низких частот
-        
-        # Создаем частотную маску
-        freq_mask = self.create_frequency_soft_cutoff_mask(
-            height, width, cutoff_radius, transition_width, device
-        )
-        freq_mask = freq_mask.unsqueeze(0).unsqueeze(0)  # [1, 1, H, W]
-        
-        # Преобразуем латенты в частотную область
-        fft_image = torch.fft.fft2(image_latents, dim=(-2, -1))
-        fft_shifted = torch.fft.fftshift(fft_image, dim=(-2, -1))
-        
-        # Извлекаем фазу изображения
-        image_phase = torch.angle(fft_shifted)
-        
-        # Генерируем гауссовский шум
-        noise = torch.randn_like(image_latents) * noise_std
-        
-        # Преобразуем шум в частотную область
-        fft_noise = torch.fft.fft2(noise, dim=(-2, -1))
-        fft_noise_shifted = torch.fft.fftshift(fft_noise, dim=(-2, -1))
-        
-        # Извлекаем амплитуду шума
-        noise_magnitude = torch.abs(fft_noise_shifted)
-        noise_phase = torch.angle(fft_noise_shifted)
-        
-        # Смешиваем фазы: низкие частоты - фаза изображения, высокие - фаза шума
-        mixed_phase = freq_mask * image_phase + (1 - freq_mask) * noise_phase
-        
-        # Собираем обратно: амплитуда шума + смешанная фаза
-        fft_combined = noise_magnitude * torch.exp(1j * mixed_phase)
-        fft_unshifted = torch.fft.ifftshift(fft_combined, dim=(-2, -1))
-        
-        # Обратное преобразование
-        structured_noise = torch.fft.ifft2(fft_unshifted, dim=(-2, -1))
-        structured_noise = torch.real(structured_noise)
-        
-        # Нормализуе�� для сохранения статистики гауссовского шума
-        current_std = torch.std(structured_noise)
-        if current_std > 0:
-            structured_noise = structured_noise / current_std * noise_std
-        
-        return structured_noise.to(dtype)
-
-    def preprocess_image(self, image: Image.Image, width: int, height: int):
-        """Ресайз и центрированный кроп изображения для асимметричного VAE."""
-        # Для энкодера с масштабом 8
-        target_height = ((height // self.vae_scale_factor) * self.vae_scale_factor)//2
-        target_width = ((width // self.vae_scale_factor) * self.vae_scale_factor)//2
-        
-        w, h = image.size
-        aspect_ratio = target_width / target_height
-        
-        if w / h > aspect_ratio:
-            new_w = int(h * aspect_ratio)
-            left = (w - new_w) // 2
-            image = image.crop((left, 0, left + new_w, h))
-        else:
-            new_h = int(w / aspect_ratio)
-            top = (h - new_h) // 2
-            image = image.crop((0, top, w, top + new_h))
-        
-        image = image.resize((target_width, target_height), resample=Image.LANCZOS)
-        image = np.array(image).astype(np.float32) / 255.0
-        image = image[None].transpose(0, 3, 1, 2)  # [1, C, H, W]
-        image = torch.from_numpy(image)
-        return 2.0 * image - 1.0  # [-1, 1]
-
-    def encode_prompt(self, prompt, negative_prompt, device, dtype):
-        def get_single_encode(texts, is_negative=False):
-            if texts is None or texts == "":
-                hidden_dim = self.text_encoder.config.hidden_size
-                shape = (1, self.text_encoder.config.max_position_embeddings, hidden_dim)
-                emb = torch.zeros(shape, dtype=dtype, device=device)
-                mask = torch.ones((1, self.text_encoder.config.max_position_embeddings), dtype=torch.int64, device=device)
-                return emb, mask
-
-            if isinstance(texts, str):
-                texts = [texts]
-
-            with torch.no_grad():
-                toks = self.tokenizer(
-                    texts,
-                    padding="max_length",
-                    max_length=self.text_encoder.config.max_position_embeddings,
-                    truncation=True,
-                    return_tensors="pt"
-                ).to(device)
-
-                outputs = self.text_encoder(
-                    input_ids=toks.input_ids, 
-                    attention_mask=toks.attention_mask,
-                    output_hidden_states=True
-                )
-
-                layer_index = -2
-                prompt_embeds = outputs.hidden_states[layer_index]
-                final_layer_norm = self.text_encoder.text_model.final_layer_norm
-                prompt_embeds = final_layer_norm(prompt_embeds)
-
-                return prompt_embeds, toks.attention_mask
-
-        pos_embeds, pos_mask = get_single_encode(prompt)
-        neg_embeds, neg_mask = get_single_encode(negative_prompt, is_negative=True)
-
-        batch_size = pos_embeds.shape[0]
-        if neg_embeds.shape[0] != batch_size:
-            neg_embeds = neg_embeds.repeat(batch_size, 1, 1)
-            neg_mask = neg_mask.repeat(batch_size, 1)
-
-        text_embeddings = torch.cat([neg_embeds, pos_embeds], dim=0)
-        final_mask = torch.cat([neg_mask, pos_mask], dim=0)
-
-        return text_embeddings.to(dtype=dtype), final_mask.to(dtype=torch.int64)
-
-    @torch.no_grad()
-    def __call__(
-        self,
-        prompt: Union[str, List[str]],
-        image: Optional[Union[Image.Image, List[Image.Image]]] = None,
-        coef: float = 0.5,  # strength: 1.0 - полный шум, 0.0 - оригинал
-        negative_prompt: Optional[Union[str, List[str]]] = None,
-        height: int = 1024,
-        width: int = 1024,
-        num_inference_steps: int = 40,
-        guidance_scale: float = 4.0,
-        generator: Optional[torch.Generator] = None,
-        seed: Optional[int] = None,
-        output_type: str = "pil",
-        return_dict: bool = True,
-        structure_preservation: float = 0.2,  # Новый параметр: сохранение структуры 0-1
-        **kwargs,
-    ):
-        device = self.device
-        dtype = self.unet.dtype
-        
-        if generator is None and seed is not None:
-            if torch.cuda.is_available():
-                generator = torch.Generator(device=device)
-            else:
-                generator = torch.Generator()
-            generator.manual_seed(seed)
-        
-        # 1. Encode Prompt
-        text_embeddings, attention_mask = self.encode_prompt(
-            prompt, negative_prompt, device, dtype
-        )
-        batch_size = 1 if isinstance(prompt, str) else len(prompt)
-        
-        # 2. Настройка таймстепов - ВСЕГДА используем ВСЕ шаги
-        self.scheduler.set_timesteps(num_inference_steps, device=device)
-        timesteps = self.scheduler.timesteps  # Все 40 шагов
-        
-        #print(f"Используем ВСЕ шаги: {len(timesteps)}")
-        #print(f"Диапазон таймстепов: [{timesteps[0].item():.3f}, {timesteps[-1].item():.3f}]")
-        #print(f"Коэффициент смешивания (coef): {coef}")
-        
-        # 3. Обработка img2img с структурированным шумом
-        if image is not None:
-            # Подготовка изображения
-            if isinstance(image, Image.Image):
-                image_tensor = self.preprocess_image(image, width, height).to(
-                    device=device, dtype=self.vae.dtype
-                )
-            else:
-                image_tensor = self.preprocess_image(image[0], width, height).to(
-                    device=device, dtype=self.vae.dtype
-                )
-            
-            # Кодируем в латенты
-            latents_clean = self.vae.encode(image_tensor).latent_dist.sample(generator=generator)
-            vae_scaling_factor = getattr(self.vae.config, "scaling_factor", 1.0)
-            vae_shift_factor = getattr(self.vae.config, "shift_factor", 0.0)
-            latents_clean = (latents_clean - vae_shift_factor) / vae_scaling_factor
-            latents_clean = latents_clean.to(dtype=dtype)
-            
-            # Автоматический расчет cutoff_radius на основе structure_preservation
-            latent_height, latent_width = latents_clean.shape[-2], latents_clean.shape[-1]
-            max_radius = min(latent_height, latent_width) / 2
-            cutoff_radius = max_radius * structure_preservation
-            
-            # Генерируем структурированный шум
-            structured_noise = self.generate_structured_noise(
-                image_latents=latents_clean,
-                cutoff_radius=cutoff_radius,
-                transition_width=2.0,
-                noise_std=1.0
-            )
-            
-            # Нормализуем шум
-            current_std = torch.std(structured_noise)
-            if current_std > 0:
-                structured_noise = structured_noise / current_std
-            
-            # КЛЮЧЕВОЕ ИЗМЕНЕНИЕ: Простое линейное смешивание
-            # coef=0.0 -> 100% оригинал, 0% шум
-            # coef=1.0 -> 0% оригинал, 100% шум
-            print(f"Смешивание: {100*(1-coef):.1f}% оригинал + {100*coef:.1f}% шум")
-            
-            # Важно: инвертируем coef, если хотим, чтобы coef=0.1 давал слабое изменение
-            # coef=0.1 -> 90% оригинал + 10% шум
-            # coef=0.9 -> 10% оригинал + 90% шум
-            latents = (1.0 - coef) * latents_clean + coef * structured_noise
-            
-        else:
-            # TXT2IMG: начинаем с чистого шума (coef=1.0 эквивалентно)
-            vae_scaling_factor = getattr(self.vae.config, "scaling_factor", 1.0)
-            vae_shift_factor = getattr(self.vae.config, "shift_factor", 0.0)
-            
-            latent_height = height // self.vae_scale_factor
-            latent_width = width // self.vae_scale_factor
-            
-            latents = torch.randn(
-                (batch_size, self.unet.config.in_channels, latent_height, latent_width),
-                generator=generator, 
-                device=device, 
-                dtype=dtype
-            )
-        
-        # 4. Denoising Loop
-        for i, t in enumerate(tqdm(timesteps, desc="Sampling")):
-            # CFG preparation
-            latent_model_input = torch.cat([latents] * 2) if guidance_scale > 1 else latents
-            
-            # Predict flow
-            model_out = self.unet(
-                latent_model_input,
-                t,
-                encoder_hidden_states=text_embeddings,
-                encoder_attention_mask=attention_mask,
-                return_dict=False,
-            )[0]
-            
-            # CFG
-            if guidance_scale > 1:
-                flow_uncond, flow_cond = model_out.chunk(2)
-                model_out = flow_uncond + guidance_scale * (flow_cond - flow_uncond)
-            
-            # Euler step для flow matching
-            latents = self.scheduler.step(model_out, t, latents, return_dict=False)[0]
-        
-        # 5. Decode
-        if output_type == "latent":
-            return SdxsPipelineOutput(images=latents)
-        
-        # Масштабируем обратно для VAE
-        latents = latents * vae_scaling_factor + vae_shift_factor
-        image_output = self.vae.decode(latents.to(self.vae.dtype), return_dict=False)[0]
-        
-        
-        # Нормализуем к [0, 1] для PIL
-        image_output = (image_output.clamp(-1, 1) + 1) / 2
-        image_np = image_output.cpu().permute(0, 2, 3, 1).float().numpy()
-        
-        if output_type == "pil":
-            image_np = (image_np * 255).round().astype("uint8")
-            images = [Image.fromarray(img) for img in image_np]
-        else:
-            images = image_np
-        
-        if not return_dict:
-            return images
-        
-        return SdxsPipelineOutput(images=images)

pipeline_sdxs.py

@@ -100,7 +100,7 @@ class SdxsPipeline(DiffusionPipeline):
         self,
         prompt: Union[str, List[str]],
         image: Optional[Union[Image.Image, List[Image.Image]]] = None,
-        coef: float = 0.5,                    # ← strength (0.0 = оригинал, 1.0 = полный шум)
+        coef: float = 0.97,                    # ← strength (0.0 = оригинал, 1.0 = полный шум)
         negative_prompt: Optional[Union[str, List[str]]] = None,
         height: int = 1024,
         width: int = 1024,

test.ipynb

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:28095b7ec189c24ef448aaf1f3e44fdf347533348d04d187f46b9599ff83a06c
-size 5721667
+oid sha256:632553d43b6346999891e3f671bbf197a6f4e1bca39362387e48e3eb6f357630
+size 5305844

unet/config.json

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:4a93a03ab94ebcdad5427326dae7fabbc74a7f46dca4a3804c3e5c11e667ff7e
-size 1848
+oid sha256:18ea6e42455fff8208c1a900f1e343224198079dc0814f90d4ff283209c3924a
+size 1843

unet/diffusion_pytorch_model.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:d91907509f145bade2cea2809952d4bdfcbef609efb23ec606e5b9bb10239d2f
-size 5163635688
+oid sha256:e342af4061c1f6cce4214d23962c1033cf61f2cfee942c42661f72dc29b179a3
+size 2581889304

unet_1.3b/config.json

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:4a93a03ab94ebcdad5427326dae7fabbc74a7f46dca4a3804c3e5c11e667ff7e
-size 1848

unet_1.3b/diffusion_pytorch_model.safetensors

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:0acc47c6590d2347959a9338ef56fefbd59b8b09b44bdf8db019d700e4cb3bef
-size 5163635688

vae/.gitattributes

@@ -0,0 +1,35 @@
+*.7z filter=lfs diff=lfs merge=lfs -text
+*.arrow filter=lfs diff=lfs merge=lfs -text
+*.bin filter=lfs diff=lfs merge=lfs -text
+*.bz2 filter=lfs diff=lfs merge=lfs -text
+*.ckpt filter=lfs diff=lfs merge=lfs -text
+*.ftz filter=lfs diff=lfs merge=lfs -text
+*.gz filter=lfs diff=lfs merge=lfs -text
+*.h5 filter=lfs diff=lfs merge=lfs -text
+*.joblib filter=lfs diff=lfs merge=lfs -text
+*.lfs.* filter=lfs diff=lfs merge=lfs -text
+*.mlmodel filter=lfs diff=lfs merge=lfs -text
+*.model filter=lfs diff=lfs merge=lfs -text
+*.msgpack filter=lfs diff=lfs merge=lfs -text
+*.npy filter=lfs diff=lfs merge=lfs -text
+*.npz filter=lfs diff=lfs merge=lfs -text
+*.onnx filter=lfs diff=lfs merge=lfs -text
+*.ot filter=lfs diff=lfs merge=lfs -text
+*.parquet filter=lfs diff=lfs merge=lfs -text
+*.pb filter=lfs diff=lfs merge=lfs -text
+*.pickle filter=lfs diff=lfs merge=lfs -text
+*.pkl filter=lfs diff=lfs merge=lfs -text
+*.pt filter=lfs diff=lfs merge=lfs -text
+*.pth filter=lfs diff=lfs merge=lfs -text
+*.rar filter=lfs diff=lfs merge=lfs -text
+*.safetensors filter=lfs diff=lfs merge=lfs -text
+saved_model/**/* filter=lfs diff=lfs merge=lfs -text
+*.tar.* filter=lfs diff=lfs merge=lfs -text
+*.tar filter=lfs diff=lfs merge=lfs -text
+*.tflite filter=lfs diff=lfs merge=lfs -text
+*.tgz filter=lfs diff=lfs merge=lfs -text
+*.wasm filter=lfs diff=lfs merge=lfs -text
+*.xz filter=lfs diff=lfs merge=lfs -text
+*.zip filter=lfs diff=lfs merge=lfs -text
+*.zst filter=lfs diff=lfs merge=lfs -text
+*tfevents* filter=lfs diff=lfs merge=lfs -text

vae/config.json

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:aca6cfc9bfab2f27ae086c52df8bc1d7b4ebede07c3d6d1b29c47e39cca1b753
-size 752
+oid sha256:12d252abeac321629cb81b908a6b49d1bf8d7f60247e00b6be83fd03b0f98b39
+size 852

vae/diffusion_pytorch_model.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:b6342c154910ff1065c25915b107ab1326f146c44055f3a4119ffb95f5159a4f
-size 382598708
+oid sha256:19bd4d341b7cc8d20893e2f257760c8b964bad447a243263191c4be1c89c1aaf
+size 427466716

vae/train_vae_fdl.py

@@ -0,0 +1,624 @@
+# -*- coding: utf-8 -*-
+import os
+import math
+import re
+import torch
+import numpy as np
+import random
+import gc
+from datetime import datetime
+from pathlib import Path
+
+import torchvision.transforms as transforms
+import torch.nn.functional as F
+from torch.utils.data import DataLoader, Dataset
+from torch.optim.lr_scheduler import LambdaLR
+from diffusers import AutoencoderKL, AsymmetricAutoencoderKL
+# QWEN: импорт класса
+from diffusers import AutoencoderKLQwenImage
+from diffusers import AutoencoderKLWan
+
+from accelerate import Accelerator
+from PIL import Image, UnidentifiedImageError
+from tqdm import tqdm
+import bitsandbytes as bnb
+import wandb
+import lpips   # pip install lpips
+from FDL_pytorch import FDL_loss # pip install fdl-pytorch
+from collections import deque
+
+# --------------------------- Параметры ---------------------------
+ds_path            = "/workspace/d23"
+project            = "vae10"
+batch_size         = 1
+base_learning_rate = 6e-6
+min_learning_rate  = 7e-7
+num_epochs         = 2
+sample_interval_share = 25
+use_wandb          = True
+save_model         = True
+use_decay          = True
+optimizer_type     = "adam8bit"
+dtype              = torch.float32
+
+model_resolution   = 512 #288
+high_resolution    = 1024 #576 
+limit              = 0
+save_barrier       = 1.3
+warmup_percent     = 0.005
+percentile_clipping = 99
+beta2              = 0.997
+eps                = 1e-8
+clip_grad_norm     = 1.0
+mixed_precision    = "no"
+gradient_accumulation_steps = 1
+generated_folder   = "samples"
+save_as            = "vae10"
+num_workers        = 0
+device = None
+torch.backends.cuda.matmul.allow_tf32 = True
+torch.backends.cudnn.allow_tf32 = True
+# Включение Flash Attention 2/SDPA #MAX_JOBS=4 pip install flash-attn --no-build-isolation
+torch.backends.cuda.enable_flash_sdp(True)
+torch.backends.cuda.enable_mem_efficient_sdp(True)
+torch.backends.cuda.enable_math_sdp(False) 
+
+# --- Режимы обучения ---
+# QWEN: учим только декодер
+train_decoder_only = True
+train_up_only      = False
+full_training      = False  # если True — учим весь VAE и добавляем KL (ниже)
+kl_ratio           = 0.00
+
+# Доли лоссов
+loss_ratios = {
+    "lpips": 0.70,#0.50,
+    "fdl" :  0.10,#0.25,   
+    "edge":  0.05,
+    "mse":   0.10,
+    "mae":   0.05,
+    "kl":    0.00,  # активируем при full_training=True
+}
+median_coeff_steps = 250
+
+resize_long_side = 1280  # ресайз длинной стороны исходных картинок
+
+# QWEN: конфиг загрузки модели
+vae_kind      = "kl"  # "qwen" или "kl" (обычный)
+
+Path(generated_folder).mkdir(parents=True, exist_ok=True)
+
+accelerator = Accelerator(
+    mixed_precision=mixed_precision,
+    gradient_accumulation_steps=gradient_accumulation_steps
+)
+device = accelerator.device
+
+# reproducibility
+seed = int(datetime.now().strftime("%Y%m%d")) + 13
+torch.manual_seed(seed); np.random.seed(seed); random.seed(seed)
+torch.backends.cudnn.benchmark = False
+
+# --------------------------- WandB ---------------------------
+if use_wandb and accelerator.is_main_process:
+    wandb.init(project=project, config={
+        "batch_size": batch_size,
+        "base_learning_rate": base_learning_rate,
+        "num_epochs": num_epochs,
+        "optimizer_type": optimizer_type,
+        "model_resolution": model_resolution,
+        "high_resolution": high_resolution,
+        "gradient_accumulation_steps": gradient_accumulation_steps,
+        "train_decoder_only": train_decoder_only,
+        "full_training": full_training,
+        "kl_ratio": kl_ratio,
+        "vae_kind": vae_kind,
+    })
+
+# --------------------------- VAE ---------------------------
+def get_core_model(model):
+    m = model
+    # если модель уже обёрнута torch.compile
+    if hasattr(m, "_orig_mod"):
+        m = m._orig_mod
+    return m
+
+def is_video_vae(model) -> bool:
+    # WAN/Qwen — это видео-VAEs
+    if vae_kind in ("wan", "qwen"):
+        return True
+    # fallback по структуре (если понадобится)
+    try:
+        core = get_core_model(model)
+        enc = getattr(core, "encoder", None)
+        conv_in = getattr(enc, "conv_in", None)
+        w = getattr(conv_in, "weight", None)
+        if isinstance(w, torch.nn.Parameter):
+            return w.ndim == 5
+    except Exception:
+        pass
+    return False
+
+# загрузка
+if vae_kind == "qwen":
+    vae = AutoencoderKLQwenImage.from_pretrained("Qwen/Qwen-Image", subfolder="vae")
+else:
+    if vae_kind == "wan":
+        vae = AutoencoderKLWan.from_pretrained(project)
+    else:
+        # старое поведение (пример)
+        if model_resolution==high_resolution:
+            vae = AutoencoderKL.from_pretrained(project)
+        else:
+            vae = AsymmetricAutoencoderKL.from_pretrained(project)
+
+vae = vae.to(dtype)
+
+# torch.compile (опцион��льно)
+if hasattr(torch, "compile"):
+    try:
+        vae = torch.compile(vae)
+    except Exception as e:
+        print(f"[WARN] torch.compile failed: {e}")
+
+# --------------------------- Freeze/Unfreeze ---------------------------
+core = get_core_model(vae)
+
+for p in core.parameters():
+    p.requires_grad = False
+
+unfrozen_param_names = []
+
+if full_training and not train_decoder_only:
+    for name, p in core.named_parameters():
+        p.requires_grad = True
+        unfrozen_param_names.append(name)
+        loss_ratios["kl"] = float(kl_ratio)
+        trainable_module = core
+else:
+    # учим только 0-й блок декодера + post_quant_conv
+    if hasattr(core, "decoder"):
+        if train_up_only:#hasattr(core.decoder, "up_blocks") and len(core.decoder.up_blocks) > 0:
+            # --- только 0-й up_block ---
+            for name, p in core.decoder.up_blocks[0].named_parameters():
+                p.requires_grad = True
+                unfrozen_param_names.append(f"{name}")
+        else:
+            print("Decoder — fallback to full decoder")
+            for name, p in core.decoder.named_parameters():
+                p.requires_grad = True
+                unfrozen_param_names.append(f"decoder.{name}")
+    if hasattr(core, "post_quant_conv"):
+        for name, p in core.post_quant_conv.named_parameters():
+            p.requires_grad = True
+            unfrozen_param_names.append(f"post_quant_conv.{name}")
+    trainable_module = core.decoder if hasattr(core, "decoder") else core
+
+
+print(f"[INFO] Разморожено параметров: {len(unfrozen_param_names)}. Первые 200 имён:")
+for nm in unfrozen_param_names[:200]:
+    print(" ", nm)
+
+# --------------------------- Датасет ---------------------------
+class PngFolderDataset(Dataset):
+    def __init__(self, root_dir, min_exts=('.png',), resolution=1024, limit=0):
+        self.root_dir = root_dir
+        self.resolution = resolution
+        self.paths = []
+        for root, _, files in os.walk(root_dir):
+            for fname in files:
+                if fname.lower().endswith(tuple(ext.lower() for ext in min_exts)):
+                    self.paths.append(os.path.join(root, fname))
+        if limit:
+            self.paths = self.paths[:limit]
+        valid = []
+        for p in self.paths:
+            try:
+                with Image.open(p) as im:
+                    im.verify()
+                valid.append(p)
+            except (OSError, UnidentifiedImageError):
+                continue
+        self.paths = valid
+        if len(self.paths) == 0:
+            raise RuntimeError(f"No valid PNG images found under {root_dir}")
+        random.shuffle(self.paths)
+
+    def __len__(self):
+        return len(self.paths)
+
+    def __getitem__(self, idx):
+        p = self.paths[idx % len(self.paths)]
+        with Image.open(p) as img:
+            img = img.convert("RGB")
+            if not resize_long_side or resize_long_side <= 0:
+                return img
+            w, h = img.size
+            long = max(w, h)
+            if long <= resize_long_side:
+                return img
+            scale = resize_long_side / float(long)
+            new_w = int(round(w * scale))
+            new_h = int(round(h * scale))
+            return img.resize((new_w, new_h), Image.BICUBIC)
+
+def random_crop(img, sz):
+    w, h = img.size
+    if w < sz or h < sz:
+        img = img.resize((max(sz, w), max(sz, h)), Image.BICUBIC)
+    x = random.randint(0, max(1, img.width - sz))
+    y = random.randint(0, max(1, img.height - sz))
+    return img.crop((x, y, x + sz, y + sz))
+
+tfm = transforms.Compose([
+    transforms.ToTensor(),
+    transforms.Normalize([0.5, 0.5, 0.5], [0.5, 0.5, 0.5])
+])
+
+dataset = PngFolderDataset(ds_path, min_exts=('.png',), resolution=high_resolution, limit=limit)
+print("len(dataset)",len(dataset))
+if len(dataset) < batch_size:
+    raise RuntimeError(f"Not enough valid images ({len(dataset)}) to form a batch of size {batch_size}")
+
+def collate_fn(batch):
+    imgs = []
+    for img in batch:
+        img = random_crop(img, high_resolution)
+        imgs.append(tfm(img))
+    return torch.stack(imgs)
+
+dataloader = DataLoader(
+    dataset,
+    batch_size=batch_size,
+    shuffle=True,
+    collate_fn=collate_fn,
+    num_workers=num_workers,
+    pin_memory=True,
+    drop_last=True
+)
+
+# --------------------------- Оптимизатор ---------------------------
+def get_param_groups(module, weight_decay=0.001):
+    no_decay = ["bias", "LayerNorm.weight", "layer_norm.weight", "ln_1.weight", "ln_f.weight"]
+    decay_params, no_decay_params = [], []
+    for n, p in vae.named_parameters():  # глобально по vae, с фильтром requires_grad
+        if not p.requires_grad:
+            continue
+        if any(nd in n for nd in no_decay):
+            no_decay_params.append(p)
+        else:
+            decay_params.append(p)
+    return [
+        {"params": decay_params, "weight_decay": weight_decay},
+        {"params": no_decay_params, "weight_decay": 0.0},
+    ]
+
+def get_param_groups(module, weight_decay=0.001):
+    no_decay_tokens = ("bias", "norm", "rms", "layernorm")
+    decay_params, no_decay_params = [], []
+    for n, p in module.named_parameters():
+        if not p.requires_grad:
+            continue
+        n_l = n.lower()
+        if any(t in n_l for t in no_decay_tokens):
+            no_decay_params.append(p)
+        else:
+            decay_params.append(p)
+    return [
+        {"params": decay_params, "weight_decay": weight_decay},
+        {"params": no_decay_params, "weight_decay": 0.0},
+    ]
+
+def create_optimizer(name, param_groups):
+    if name == "adam8bit":
+        return bnb.optim.AdamW8bit(param_groups, lr=base_learning_rate, betas=(0.9, beta2), eps=eps)
+    raise ValueError(name)
+
+param_groups = get_param_groups(get_core_model(vae), weight_decay=0.001)
+optimizer = create_optimizer(optimizer_type, param_groups)
+
+# --------------------------- LR schedule ---------------------------
+batches_per_epoch = len(dataloader)
+steps_per_epoch = int(math.ceil(batches_per_epoch / float(gradient_accumulation_steps)))
+total_steps = steps_per_epoch * num_epochs
+
+def lr_lambda(step):
+    if not use_decay:
+        return 1.0
+    x = float(step) / float(max(1, total_steps))
+    warmup = float(warmup_percent)
+    min_ratio = float(min_learning_rate) / float(base_learning_rate)
+    if x < warmup:
+        return min_ratio + (1.0 - min_ratio) * (x / warmup)
+    decay_ratio = (x - warmup) / (1.0 - warmup)
+    return min_ratio + 0.5 * (1.0 - min_ratio) * (1.0 + math.cos(math.pi * decay_ratio))
+
+scheduler = LambdaLR(optimizer, lr_lambda)
+
+# Подготовка
+dataloader, vae, optimizer, scheduler = accelerator.prepare(dataloader, vae, optimizer, scheduler)
+trainable_params = [p for p in vae.parameters() if p.requires_grad]
+
+# fdl
+fdl_loss = FDL_loss()
+fdl_loss = fdl_loss.to(accelerator.device)
+
+# --------------------------- LPIPS и вспомогательные ---------------------------
+_lpips_net = None
+def _get_lpips():
+    global _lpips_net
+    if _lpips_net is None:
+        _lpips_net = lpips.LPIPS(net='vgg', verbose=False).eval().to(accelerator.device).eval()
+    return _lpips_net
+
+_sobel_kx = torch.tensor([[[[-1,0,1],[-2,0,2],[-1,0,1]]]], dtype=torch.float32)
+_sobel_ky = torch.tensor([[[[-1,-2,-1],[0,0,0],[1,2,1]]]], dtype=torch.float32)
+def sobel_edges(x: torch.Tensor) -> torch.Tensor:
+    C = x.shape[1]
+    kx = _sobel_kx.to(x.device, x.dtype).repeat(C, 1, 1, 1)
+    ky = _sobel_ky.to(x.device, x.dtype).repeat(C, 1, 1, 1)
+    gx = F.conv2d(x, kx, padding=1, groups=C)
+    gy = F.conv2d(x, ky, padding=1, groups=C)
+    return torch.sqrt(gx * gx + gy * gy + 1e-12)
+
+class MedianLossNormalizer:
+    def __init__(self, desired_ratios: dict, window_steps: int):
+        s = sum(desired_ratios.values())
+        self.ratios = {k: (v / s) if s > 0 else 0.0 for k, v in desired_ratios.items()}
+        self.buffers = {k: deque(maxlen=window_steps) for k in self.ratios.keys()}
+        self.window = window_steps
+
+    def update_and_total(self, abs_losses: dict):
+        for k, v in abs_losses.items():
+            if k in self.buffers:
+                self.buffers[k].append(float(v.detach().abs().cpu()))
+        meds = {k: (np.median(self.buffers[k]) if len(self.buffers[k]) > 0 else 1.0) for k in self.buffers}
+        coeffs = {k: (self.ratios[k] / max(meds[k], 1e-12)) for k in self.ratios}
+        total = sum(coeffs[k] * abs_losses[k] for k in abs_losses if k in coeffs)
+        return total, coeffs, meds
+
+if full_training and not train_decoder_only:
+    loss_ratios["kl"] = float(kl_ratio)
+normalizer = MedianLossNormalizer(loss_ratios, median_coeff_steps)
+
+# --------------------------- Сэмплы ---------------------------
+@torch.no_grad()
+def get_fixed_samples(n=3):
+    idx = random.sample(range(len(dataset)), min(n, len(dataset)))
+    pil_imgs = [dataset[i] for i in idx]
+    tensors = []
+    for img in pil_imgs:
+        img = random_crop(img, high_resolution)
+        tensors.append(tfm(img))
+    return torch.stack(tensors).to(accelerator.device, dtype)
+
+fixed_samples = get_fixed_samples()
+
+@torch.no_grad()
+def _to_pil_uint8(img_tensor: torch.Tensor) -> Image.Image:
+    arr = ((img_tensor.float().clamp(-1, 1) + 1.0) * 127.5).clamp(0, 255).byte().cpu().numpy().transpose(1, 2, 0)
+    return Image.fromarray(arr)
+
+
+@torch.no_grad()
+def generate_and_save_samples(step=None):
+    try:
+        #temp_vae = accelerator.unwrap_model(vae).eval()
+        if hasattr(vae, "module"):
+            # Если это DDP или DistributedDataParallel
+            unwrapped_vae = vae.module
+        else:
+            unwrapped_vae = vae
+        
+        # Если использовался torch.compile, достаем оригинал
+        if hasattr(unwrapped_vae, "_orig_mod"):
+            temp_vae = unwrapped_vae._orig_mod
+        else:
+            temp_vae = unwrapped_vae
+        
+        temp_vae = temp_vae.eval()
+        lpips_net = _get_lpips()
+        with torch.no_grad():
+            orig_high = fixed_samples
+            orig_low = F.interpolate(
+                orig_high,
+                size=(model_resolution, model_resolution),
+                mode="bilinear",
+                align_corners=False
+            )
+            model_dtype = next(temp_vae.parameters()).dtype
+            orig_low = orig_low.to(dtype=model_dtype)
+
+            # Encode/decode с учётом видео-режима
+            if is_video_vae(temp_vae):
+                x_in = orig_low.unsqueeze(2)  # [B,3,1,H,W]
+                enc = temp_vae.encode(x_in)
+                latents_mean = enc.latent_dist.mean
+                dec = temp_vae.decode(latents_mean).sample  # [B,3,1,H,W]
+                rec = dec.squeeze(2)  # [B,3,H,W]
+            else:
+                enc = temp_vae.encode(orig_low)
+                latents_mean = enc.latent_dist.mean
+                rec = temp_vae.decode(latents_mean).sample
+
+        # Подгон размеров, если надо
+        #if rec.shape[-2:] != orig_high.shape[-2:]:
+        #    rec = F.interpolate(rec, size=orig_high.shape[-2:], mode="bilinear", align_corners=False)
+
+        # Сохраняем все real/decoded
+        for i in range(rec.shape[0]):
+            real_img = _to_pil_uint8(orig_high[i])
+            dec_img  = _to_pil_uint8(rec[i])
+            real_img.save(f"{generated_folder}/sample_real_{i}.png")
+            dec_img.save(f"{generated_folder}/sample_decoded_{i}.png")
+
+        # LPIPS
+        lpips_scores = []
+        for i in range(rec.shape[0]):
+            orig_full = orig_high[i:i+1].to(torch.float32)
+            rec_full  = rec[i:i+1].to(torch.float32)
+            #if rec_full.shape[-2:] != orig_full.shape[-2:]:
+            #    rec_full = F.interpolate(rec_full, size=orig_full.shape[-2:], mode="bilinear", align_corners=False)
+            lpips_val = lpips_net(orig_full, rec_full).item()
+            lpips_scores.append(lpips_val)
+        avg_lpips = float(np.mean(lpips_scores))
+
+        # W&B логирование
+        if use_wandb and accelerator.is_main_process:
+            log_data = {"lpips_mean": avg_lpips}
+            for i in range(rec.shape[0]):
+                log_data[f"sample/real_{i}"] = wandb.Image(f"{generated_folder}/sample_real_{i}.png", caption=f"real_{i}")
+                log_data[f"sample/decoded_{i}"] = wandb.Image(f"{generated_folder}/sample_decoded_{i}.png", caption=f"decoded_{i}")
+            wandb.log(log_data, step=step)
+
+    finally:
+        gc.collect()
+        torch.cuda.empty_cache()
+
+
+if accelerator.is_main_process and save_model:
+    print("Генерация сэмплов до старта обучения...")
+    generate_and_save_samples(0)
+
+accelerator.wait_for_everyone()
+
+# --------------------------- Тренировка ---------------------------
+progress = tqdm(total=total_steps, disable=not accelerator.is_local_main_process)
+global_step = 0
+min_loss = float("inf")
+sample_interval = max(1, total_steps // max(1, sample_interval_share * num_epochs))
+
+for epoch in range(num_epochs):
+    vae.train()
+    batch_losses, batch_grads = [], []
+    track_losses = {k: [] for k in loss_ratios.keys()}
+
+    for imgs in dataloader:
+        with accelerator.accumulate(vae):
+            imgs = imgs.to(accelerator.device)
+
+            if high_resolution != model_resolution:
+                imgs_low = F.interpolate(imgs, size=(model_resolution, model_resolution),mode="area") # mode="bilinear", align_corners=False)
+            else:
+                imgs_low = imgs
+
+            model_dtype = next(vae.parameters()).dtype
+            imgs_low_model = imgs_low.to(dtype=model_dtype) if imgs_low.dtype != model_dtype else imgs_low
+
+            # Вместо: current_vae = accelerator.unwrap_model(vae)
+            unwrapped = vae.module if hasattr(vae, "module") else vae
+            current_vae = getattr(unwrapped, "_orig_mod", unwrapped)
+
+
+            # QWEN: encode/decode с T=1
+            if is_video_vae(current_vae):
+                x_in = imgs_low_model.unsqueeze(2)             # [B,3,1,H,W]
+                enc = current_vae.encode(x_in)
+                latents = enc.latent_dist.mean if train_decoder_only else enc.latent_dist.sample()
+                dec = current_vae.decode(latents).sample               # [B,3,1,H,W]
+                rec = dec.squeeze(2)                           # [B,3,H,W]
+            else:
+                enc = current_vae.encode(imgs_low_model)
+                latents = enc.latent_dist.mean if train_decoder_only else enc.latent_dist.sample()
+                rec = current_vae.decode(latents).sample
+
+            #if rec.shape[-2:] != imgs.shape[-2:]:
+            #    rec = F.interpolate(rec, size=imgs.shape[-2:], mode="bilinear", align_corners=False)
+
+            rec_f32 = rec.to(torch.float32)
+            imgs_f32 = imgs.to(torch.float32)
+
+            abs_losses = {
+                "mae":   F.l1_loss(rec_f32, imgs_f32),
+                "mse":   F.mse_loss(rec_f32, imgs_f32),
+                "lpips": _get_lpips()(rec_f32, imgs_f32).mean(),
+                "fdl":   fdl_loss(rec_f32, imgs_f32),
+                "edge":  F.l1_loss(sobel_edges(rec_f32), sobel_edges(imgs_f32)),
+            }
+
+            if full_training and not train_decoder_only:
+                mean   = enc.latent_dist.mean
+                logvar = enc.latent_dist.logvar
+                kl = -0.5 * torch.mean(1 + logvar - mean.pow(2) - logvar.exp())
+                abs_losses["kl"] = kl
+            else:
+                abs_losses["kl"] = torch.tensor(0.0, device=accelerator.device, dtype=torch.float32)
+
+            total_loss, coeffs, meds = normalizer.update_and_total(abs_losses)
+
+            if torch.isnan(total_loss) or torch.isinf(total_loss):
+                raise RuntimeError("NaN/Inf loss")
+
+            accelerator.backward(total_loss)
+
+            grad_norm = torch.tensor(0.0, device=accelerator.device)
+            if accelerator.sync_gradients:
+                grad_norm = accelerator.clip_grad_norm_(trainable_params, clip_grad_norm)
+                optimizer.step()
+                scheduler.step()
+                optimizer.zero_grad(set_to_none=True)
+                global_step += 1
+                progress.update(1)
+
+            if accelerator.is_main_process:
+                try:
+                    current_lr = optimizer.param_groups[0]["lr"]
+                except Exception:
+                    current_lr = scheduler.get_last_lr()[0]
+
+                batch_losses.append(total_loss.detach().item())
+                batch_grads.append(float(grad_norm.detach().cpu().item()) if isinstance(grad_norm, torch.Tensor) else float(grad_norm))
+                for k, v in abs_losses.items():
+                    track_losses[k].append(float(v.detach().item()))
+
+                if use_wandb and accelerator.sync_gradients:
+                    log_dict = {
+                        "total_loss": float(total_loss.detach().item()),
+                        "learning_rate": current_lr,
+                        "epoch": epoch,
+                        "grad_norm": batch_grads[-1],
+                    }
+                    for k, v in abs_losses.items():
+                        log_dict[f"loss_{k}"] = float(v.detach().item())
+                    for k in coeffs:
+                        log_dict[f"coeff_{k}"] = float(coeffs[k])
+                        log_dict[f"median_{k}"] = float(meds[k])
+                    wandb.log(log_dict, step=global_step)
+
+            if global_step > 0 and global_step % sample_interval == 0:
+                if accelerator.is_main_process:
+                    generate_and_save_samples(global_step)
+                accelerator.wait_for_everyone()
+
+                n_micro = sample_interval * gradient_accumulation_steps
+                avg_loss = float(np.mean(batch_losses[-n_micro:])) if len(batch_losses) >= n_micro else float(np.mean(batch_losses)) if batch_losses else float("nan")
+                avg_grad = float(np.mean(batch_grads[-n_micro:])) if len(batch_grads) >= 1 else float(np.mean(batch_grads)) if batch_grads else 0.0
+
+                if accelerator.is_main_process:
+                    print(f"Epoch {epoch} step {global_step} loss: {avg_loss:.6f}, grad_norm: {avg_grad:.6f}, lr: {current_lr:.9f}")
+                    if save_model and avg_loss < min_loss * save_barrier:
+                        min_loss = avg_loss
+                        unwrapped = vae.module if hasattr(vae, "module") else vae
+                        current_vae = getattr(unwrapped, "_orig_mod", unwrapped)
+                        current_vae.save_pretrained(save_as)
+                    if use_wandb:
+                        wandb.log({"interm_loss": avg_loss, "interm_grad": avg_grad}, step=global_step)
+
+    if accelerator.is_main_process:
+        epoch_avg = float(np.mean(batch_losses)) if batch_losses else float("nan")
+        print(f"Epoch {epoch} done, avg loss {epoch_avg:.6f}")
+        if use_wandb:
+            wandb.log({"epoch_loss": epoch_avg, "epoch": epoch + 1}, step=global_step)
+
+# --------------------------- Финальное сохранение ---------------------------
+if accelerator.is_main_process:
+    print("Training finished – saving final model")
+    if save_model:
+        unwrapped = vae.module if hasattr(vae, "module") else vae
+        current_vae = getattr(unwrapped, "_orig_mod", unwrapped)
+        current_vae.save_pretrained(save_as)
+
+accelerator.free_memory()
+if torch.distributed.is_initialized():
+    torch.distributed.destroy_process_group()
+print("Готово!")