upd

eval_alchemist2.py

@@ -27,25 +27,27 @@ from scipy.stats import skew, kurtosis
 # ========================== Конфиг ==========================
 DEVICE = "cuda"
 DTYPE = torch.float16
-IMAGE_FOLDER = "/home/recoilme/dataset/alchemist"
+IMAGE_FOLDER = "/workspace/alchemist"
 MIN_SIZE = 1280
 CROP_SIZE = 512
 BATCH_SIZE = 10
-MAX_IMAGES = 500
+MAX_IMAGES = 0
 NUM_WORKERS = 4
 SAMPLES_DIR = "vaetest"
 
 VAE_LIST = [
     # ("SD15 VAE", AutoencoderKL, "stable-diffusion-v1-5/stable-diffusion-v1-5", "vae"),
     # ("SDXL VAE fp16 fix", AutoencoderKL, "madebyollin/sdxl-vae-fp16-fix", None),
-    ("Wan2.2-TI2V-5B", AutoencoderKLWan, "Wan-AI/Wan2.2-TI2V-5B-Diffusers", "vae"),
-    ("Wan2.2-T2V-A14B", AutoencoderKLWan, "Wan-AI/Wan2.2-T2V-A14B-Diffusers", "vae"),
+    #("Wan2.2-TI2V-5B", AutoencoderKLWan, "Wan-AI/Wan2.2-TI2V-5B-Diffusers", "vae"),
+    #("Wan2.2-T2V-A14B", AutoencoderKLWan, "Wan-AI/Wan2.2-T2V-A14B-Diffusers", "vae"),
     #("SimpleVAE1", AutoencoderKL, "/home/recoilme/simplevae/simplevae", "simple_vae_nightly"),
-    #("SimpleVAE2", AutoencoderKL, "/home/recoilme/simplevae/simplevae", "simple_vae_nightly2"),    
+    #("SimpleVAE2", AutoencoderKL, "/home/recoilme/simplevae/simplevae", "simple_vae_nightly2"),
+    #("SimpleVAE ", AutoencoderKL, "AiArtLab/simplevae", None),
     #("SimpleVAE nightly", AutoencoderKL, "AiArtLab/simplevae", "simple_vae_nightly"),
-    #("FLUX.1-schnell VAE", AutoencoderKL, "black-forest-labs/FLUX.1-schnell", "vae"),
+    ("FLUX.1-schnell VAE", AutoencoderKL, "black-forest-labs/FLUX.1-schnell", "vae"),
+    ("SimpleVAE nightly", AutoencoderKL, "AiArtLab/simplevae", "simple_vae_nightly"),
     # ("LTX-Video VAE", AutoencoderKLLTXVideo, "Lightricks/LTX-Video", "vae"),
-    ("QwenImage", AutoencoderKLQwenImage, "Qwen/Qwen-Image", "vae"),
+    #("QwenImage", AutoencoderKLQwenImage, "Qwen/Qwen-Image", "vae"),
 ]
 
 

simple_vae/diffusion_pytorch_model.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:f5f0a20e403669e880b510514ee575a2a9cb74a1b36ab0e31fc68ef66c2173d7
+oid sha256:8ba1d500c4bd376a7c8662a35fa1857c7e577da0635414b524180852143ef2f6
 size 335311892

simple_vae_nightly/diffusion_pytorch_model.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:7b705da7f401289eefa22570514d7c1b9b2f9fd32a71159e2d3d5888f74e41cd
+oid sha256:b39620d0953839362425f03674e6c3e37f03d20be3fbd7f281baea4dfc336a40
 size 335311892

train_sdxl_vae_full.py

@@ -26,9 +26,9 @@ from collections import deque
 ds_path            = "/workspace/png"
 project            = "simple_vae"
 batch_size         = 3
-base_learning_rate = 5e-5
-min_learning_rate  = 9e-7
-num_epochs         = 16
+base_learning_rate = 2e-6
+min_learning_rate  = 8e-7
+num_epochs         = 8
 sample_interval_share = 10
 use_wandb          = True
 save_model         = True
@@ -58,20 +58,20 @@ device = None  # accelerator задаст устройство
 # CHANGED: добавлен параметр для полного обучения VAE (а не только декодера).
 # Если False — поведение прежнее: учим только decoder.* (up_blocks + mid_block).
 # Если True  — размораживаем ВСЮ модель и добавляем KL-loss для энкодера.
-full_training      = True
+full_training      = False
 
 # CHANGED: добавлен вес (через долю в нормализаторе) для KL, используется только при full_training=True.
-kl_ratio           = 0.05  # простая доля для KL в общей смеси (KISS). Игнорируется, если full_training=False.
+kl_ratio           = 0.00  # простая доля для KL в общей смеси (KISS). Игнорируется, если full_training=False.
 
 # --- Пропорции лоссов и окно медианного нормирования (КОЭФ., не значения) ---
 # Итоговые доли в total loss (сумма = 1.0 после нормализации).
 loss_ratios = {
-    "lpips": 0.80,
-    "edge":  0.05,
-    "mse":   0.05,
-    "mae":   0.05,
+    "lpips": 0.60,
+    "edge":  0.10,
+    "mse":   0.15,
+    "mae":   0.15,
     # CHANGED: заранее добавлен ключ "kl" (по умолчанию 0.0). Если включаем full_training — активируем ниже.
-    "kl":    0.05,
+    "kl":    0.00,
 }
 median_coeff_steps = 256  # за сколько шагов считать медианные коэффициенты
 
@@ -490,7 +490,11 @@ for epoch in range(num_epochs):
                 mean   = enc.latent_dist.mean
                 logvar = enc.latent_dist.logvar
                 # стабильное усреднение по батчу и пространству
-                kl = -0.5 * torch.mean(1 + logvar - mean.pow(2) - logvar.exp())
+                # СТАРОЕ (неправильное):
+                #kl = -0.5 * torch.mean(1 + logvar - mean.pow(2) - logvar.exp())
+                # НОВОЕ (правильное):
+                kl_per_sample = -0.5 * torch.sum(1 + logvar - mean.pow(2) - logvar.exp(), dim=[1, 2, 3])
+                kl = torch.mean(kl_per_sample)
                 abs_losses["kl"] = kl
             else:
                 # ключ присутствует в ratios, но при partial-training его доля = 0 и он не влияет

train_sdxl_vae_wan.py

@@ -0,0 +1,565 @@
+# -*- 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 collections import deque
+
+# --------------------------- Параметры ---------------------------
+ds_path            = "/workspace/png"
+project            = "wan16x_vae"
+batch_size         = 4
+base_learning_rate = 6e-6
+min_learning_rate  = 9e-7
+num_epochs         = 8
+sample_interval_share = 10
+use_wandb          = True
+save_model         = True
+use_decay          = True
+optimizer_type     = "adam8bit"
+dtype              = torch.float32
+
+model_resolution   = 512
+high_resolution    = 512 
+limit              = 0
+save_barrier       = 1.03
+warmup_percent     = 0.01
+percentile_clipping = 95
+beta2              = 0.97
+eps                = 1e-6
+clip_grad_norm     = 1.0
+mixed_precision    = "no"
+gradient_accumulation_steps = 4
+generated_folder   = "samples"
+save_as            = "wan16x_vae_nightly"
+num_workers        = 0
+device = None
+
+# --- Режимы обучения ---
+# QWEN: учим только декодер
+train_decoder_only = True
+full_training      = False  # если True — учим весь VAE и добавляем KL (ниже)
+kl_ratio           = 0.05
+
+# Доли лоссов
+loss_ratios = {
+    "lpips": 0.75,
+    "edge":  0.05,
+    "mse":   0.10,
+    "mae":   0.10,
+    "kl":    0.00,  # активируем при full_training=True
+}
+median_coeff_steps = 256
+
+resize_long_side = 1280  # ресайз длинной стороны исходных картинок
+
+# QWEN: конфиг загрузки модели
+vae_kind      = "wan"  # "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"))
+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 (опционально)
+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 декодер + post_quant_conv на "ядре" модели
+    if hasattr(core, "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.LANCZOS)
+
+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.LANCZOS)
+    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)
+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]
+
+# --------------------------- 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()
+        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)
+
+            # QWEN: добавляем T=1 на encode/decode и снимаем при сравнении
+            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
+
+        if rec.shape[-2:] != orig_high.shape[-2:]:
+            rec = F.interpolate(rec, size=orig_high.shape[-2:], mode="bilinear", align_corners=False)
+
+        first_real = _to_pil_uint8(orig_high[0])
+        first_dec  = _to_pil_uint8(rec[0])
+        first_real.save(f"{generated_folder}/sample_real.jpg", quality=95)
+        first_dec.save(f"{generated_folder}/sample_decoded.jpg", quality=95)
+
+        for i in range(rec.shape[0]):
+            _to_pil_uint8(rec[i]).save(f"{generated_folder}/sample_{i}.jpg", quality=95)
+
+        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))
+
+        if use_wandb and accelerator.is_main_process:
+            wandb.log({"lpips_mean": avg_lpips}, 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="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
+
+            # QWEN: encode/decode с T=1
+            if is_video_vae(vae):
+                x_in = imgs_low_model.unsqueeze(2)             # [B,3,1,H,W]
+                enc = vae.encode(x_in)
+                latents = enc.latent_dist.mean if train_decoder_only else enc.latent_dist.sample()
+                dec = vae.decode(latents).sample               # [B,3,1,H,W]
+                rec = dec.squeeze(2)                           # [B,3,H,W]
+            else:
+                enc = vae.encode(imgs_low_model)
+                latents = enc.latent_dist.mean if train_decoder_only else enc.latent_dist.sample()
+                rec = 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(),
+                "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
+                        accelerator.unwrap_model(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:
+        accelerator.unwrap_model(vae).save_pretrained(save_as)
+
+accelerator.free_memory()
+if torch.distributed.is_initialized():
+    torch.distributed.destroy_process_group()
+print("Готово!")