v14

README.md

@@ -7,16 +7,15 @@ pipeline_tag: text-to-image
 
 *XS Size, Excess Quality*
 
-At AiArtLab, we strive to create a free, compact (1.7b) and fast (3 sec/image) model that can be trained on consumer graphics cards.
+At AiArtLab, we strive to create a free, compact and fast model that can be trained on consumer graphics cards.
 
 - We use U-Net for its high efficiency.
-- We have chosen the multilingual/multimodal encoder Mexma-SigLIP, which supports 80 languages.
-- We use the AuraDiffusion 16ch-VAE architecture, which preserves details and anatomy.
-- The model was trained (~1 month on 4xA5000) on approximately 1 million images with various resolutions and styles, including anime and realistic photos.
+- We have chosen the Qwen0.6b wich support 100+ languages.
+- We train new SOTA 16ch Simple VAE, which preserves details and anatomy.
+- The model was trained (~2 month on 4xRTX5090) on approximately 1+ million images with various resolutions and styles, including anime and realistic photos.
 
 ### Model Limitations:
 - Limited concept coverage due to the small dataset.
-- The Image2Image functionality requires further training.
 
 ## Acknowledgments
 - **[Stan](https://t.me/Stangle)** — Key investor. Thank you for believing in us when others called it madness.
@@ -26,10 +25,6 @@ At AiArtLab, we strive to create a free, compact (1.7b) and fast (3 sec/image) m
 ## Datasets
 - **[CaptionEmporium](https://huggingface.co/CaptionEmporium)**
 
-## Training budget
-
-Around ~$1k for now, but research budget ~$10k
-
 ## Donations
 
 Please contact with us if you may provide some GPU's or money on training
@@ -42,7 +37,7 @@ BTC: 3JHv9Hb8kEW8zMAccdgCdZGfrHeMhH1rpN
 
 [recoilme](https://t.me/recoilme)
 
-Train status, in progress: [wandb](https://wandb.ai/recoilme/micro)
+Train status, in progress:
 
 ![result](result_grid.jpg)
 
@@ -79,7 +74,7 @@ def encode_prompt(prompt, negative_prompt, device, dtype):
             negative_prompt,
             return_tensors="pt",
             padding="max_length",
-            max_length=512,
+            max_length=150,
             truncation=True,
         ).to(device)
         negative_embeddings = text_model.encode_texts(negative_inputs.input_ids, negative_inputs.attention_mask)
@@ -141,7 +136,7 @@ if __name__ == "__main__":
     scheduler = DDPMScheduler.from_pretrained(pipeid, subfolder="scheduler")
 
 
-    height, width = 576, 384
+    height, width = 640, 576
     num_inference_steps = 40
     output_folder, project_name = "samples", "sdxs"
     latents = generate_latents(

dataset_fromfolder.py → dataset.py

@@ -1,34 +1,34 @@
 # pip install flash-attn --no-build-isolation
-from datasets import Dataset, load_from_disk, concatenate_datasets
-from diffusers import AutoencoderKL
-from torchvision.transforms import Resize, ToTensor, Normalize, Compose, InterpolationMode, Lambda
-from transformers import AutoModel, AutoImageProcessor, AutoTokenizer
 import torch
 import os
 import gc
 import numpy as np
-from PIL import Image
-from tqdm import tqdm
 import random
 import json
 import shutil
 import time
+
+from datasets import Dataset, load_from_disk, concatenate_datasets
+from diffusers import AutoencoderKL,AutoencoderKLWan
+from torchvision.transforms import Resize, ToTensor, Normalize, Compose, InterpolationMode, Lambda
+from transformers import AutoModel, AutoImageProcessor, AutoTokenizer, AutoModelForCausalLM
+from typing import Dict, List, Tuple, Optional, Any
+from PIL import Image
+from tqdm import tqdm
 from datetime import timedelta
 
 # ---------------- 1️⃣ Настройки ----------------
 dtype = torch.float16
 device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
 batch_size = 5
-min_size = 320 #192
-max_size = 576 #384
-step = 64
-img_share = 0.05
+min_size = 320 #192 #256 #192
+max_size = 640 #384 #256 #384
+step = 64 #64
 empty_share = 0.05
 limit = 0
-textemb_full = False
 # Основная процедура обработки
-folder_path = "/workspace/eshu"
-save_path = "/workspace/sdxs/datasets/eshu_576"
+folder_path = "/workspace/d23" #alchemist"
+save_path = "/workspace/d23_640" #"alchemist"
 os.makedirs(save_path, exist_ok=True)
 
 # Функция для очистки CUDA памяти
@@ -42,15 +42,32 @@ def clear_cuda_memory():
 # ---------------- 2️⃣ Загрузка моделей ----------------
 def load_models():
     print("Загрузка моделей...")
-    vae = AutoencoderKL.from_pretrained("/workspace/sdxs/vae", variant="fp16",torch_dtype=dtype).to(device).eval()
-    model = AutoModel.from_pretrained("visheratin/mexma-siglip", torch_dtype=dtype, trust_remote_code=True, optimized=True).to(device).eval()
-    processor = AutoImageProcessor.from_pretrained("visheratin/mexma-siglip", use_fast=True)
-    tokenizer = AutoTokenizer.from_pretrained("visheratin/mexma-siglip")
-    return vae, model, processor, tokenizer
-
-vae, model, processor, tokenizer = load_models()
-
-
+    vae = AutoencoderKL.from_pretrained("AiArtLab/sdxs3d",subfolder="vae",torch_dtype=dtype).to(device).eval()
+
+    model_name = "Qwen/Qwen3-0.6B"
+    tokenizer = AutoTokenizer.from_pretrained(model_name)
+    model = AutoModelForCausalLM.from_pretrained(
+        model_name,
+        torch_dtype=dtype,
+        device_map=device
+    ).eval()
+    #tokenizer = AutoTokenizer.from_pretrained('Qwen/Qwen3-Embedding-0.6B', padding_side='left')
+    #model = AutoModel.from_pretrained('Qwen/Qwen3-Embedding-0.6B').to("cuda")
+    return vae, model, tokenizer
+
+vae, model, tokenizer = load_models()
+
+shift_factor = getattr(vae.config, "shift_factor", 0.0)
+if shift_factor is None:
+    shift_factor = 0.0
+
+scaling_factor = getattr(vae.config, "scaling_factor", 1.0)
+if scaling_factor is None:
+    scaling_factor = 1.0
+    
+latents_mean = getattr(vae.config, "latents_mean", None)
+latents_std = getattr(vae.config, "latents_std", None)
+    
 # ---------------- 3️⃣ Трансформации ----------------
 def get_image_transform(min_size=256, max_size=512, step=64):
     def transform(img, dry_run=False):
@@ -107,30 +124,49 @@ def get_image_transform(min_size=256, max_size=512, step=64):
     return transform
 
 # ---------------- 4️⃣ Функции обработки ----------------
-def encode_images_batch(images, processor, model):
-    pixel_values = torch.stack([processor(images=img, return_tensors="pt")["pixel_values"].squeeze(0) for img in images]).to(device, dtype)
-
+def last_token_pool(last_hidden_states: torch.Tensor,
+                    attention_mask: torch.Tensor) -> torch.Tensor:
+    # Определяем, есть ли left padding
+    left_padding = (attention_mask[:, -1].sum() == attention_mask.shape[0])
+    if left_padding:
+        return last_hidden_states[:, -1]
+    else:
+        sequence_lengths = attention_mask.sum(dim=1) - 1
+        batch_size = last_hidden_states.shape[0]
+        return last_hidden_states[torch.arange(batch_size, device=last_hidden_states.device), sequence_lengths]
+
+def encode_texts_batch(texts, tokenizer, model, device="cuda", max_length=150, normalize=False):
     with torch.inference_mode():
-        image_embeddings = model.vision_model(pixel_values).pooler_output
+        # Токенизация
+        batch = tokenizer(
+            texts,
+            return_tensors="pt",
+            padding="max_length",
+            truncation=True,
+            max_length=max_length
+        ).to(device)
 
-    return image_embeddings.unsqueeze(1).cpu().numpy()
+        # Прогон через модель
+        #outputs = model(**batch)
 
-def encode_texts_batch(texts, tokenizer, model):
-    with torch.inference_mode():
-        text_tokenized = tokenizer(texts, return_tensors="pt", padding="max_length",
-            max_length=512,
-            truncation=True).to(device)
-        text_embeddings = model.encode_texts(text_tokenized.input_ids, text_tokenized.attention_mask)
-    return text_embeddings.unsqueeze(1).cpu().numpy()
+        # Пулинг по last token
+        #embeddings = last_token_pool(outputs.last_hidden_state, batch["attention_mask"])
 
-def encode_texts_batch_full(texts, tokenizer, model):
-    with torch.inference_mode():
-        text_tokenized = tokenizer(texts, return_tensors="pt", padding="max_length",max_length=512,truncation=True).to(device)
-        features = model.text_model(
-            input_ids=text_tokenized.input_ids, attention_mask=text_tokenized.attention_mask
-        ).last_hidden_state
-        features_proj = model.text_projector(features)
-    return features_proj.cpu().numpy()
+        # L2-нормализация (опционально, обычно нужна для семантического поиска)
+        #if normalize:
+        #    embeddings = F.normalize(embeddings, p=2, dim=1)
+
+        # Прогон через базовую модель (внутри CausalLM)
+        outputs = model.model(**batch, output_hidden_states=True)
+
+        # Берем последний слой (эмбеддинги всех токенов)
+        hidden_states = outputs.hidden_states[-1]  # [B, L, D]
+
+        # Можно применить нормализацию по каждому токену (как в CLIP)
+        if normalize:
+            hidden_states = F.normalize(hidden_states, p=2, dim=-1)
+
+    return hidden_states.cpu().numpy() # embeddings.unsqueeze(1).cpu().numpy()
 
 def clean_label(label):
     label = label.replace("Image 1", "").replace("Image 2", "").replace("Image 3", "").replace("Image 4", "")
@@ -186,54 +222,34 @@ def encode_to_latents(images, texts):
 
         # Создаём батч
         batch_tensor = torch.stack(transformed_tensors).to(device, dtype)
+        if batch_tensor.ndim==5:
+            batch_tensor = batch_tensor.unsqueeze(2)  # [B, C, 1, H, W]
         
         # Кодируем батч
         with torch.no_grad():
             posteriors = vae.encode(batch_tensor).latent_dist.mode()
-            latents = (posteriors - vae.config.shift_factor) * vae.config.scaling_factor
-        
-        latents_np = latents.cpu().numpy()
-
-        # Проверка однородности форм
-        base_shape = latents_np.shape[1:]  # Форма без батча
-        valid_indices = []
-        valid_latents = []
+            latents = (posteriors - shift_factor) / scaling_factor
         
-        for idx, latent in enumerate(latents_np):
-            if latent.shape != base_shape:
-                print(f"❌ Несоответствие формы в индексе {idx}: {latent.shape} vs {base_shape}")
-                continue
-            valid_indices.append(idx)
-            valid_latents.append(latent)
+        latents_np = latents.to(dtype).cpu().numpy()
 
-        # Фильтруем данные
-        valid_pil = [pil_images[i] for i in valid_indices]
-        valid_widths = [widths[i] for i in valid_indices]
-        valid_heights = [heights[i] for i in valid_indices]
-        
         # Обрабатываем тексты
-        text_labels = [clean_label(texts[i]) for i in valid_indices]
-        if random.random() < img_share:
-            embeddings = encode_images_batch(valid_pil, processor, model)
-            text_labels = [f"img: {text_labels[i]}" for i in valid_indices]
-        else:
-            model_prompts, text_labels = process_labels_for_guidance(text_labels, empty_share)                
-            if textemb_full:
-                embeddings = encode_texts_batch_full(model_prompts, tokenizer, model)
-            else:
-                embeddings = encode_texts_batch(model_prompts, tokenizer, model)
+        text_labels = [clean_label(text) for text in texts]
+
+        model_prompts, text_labels = process_labels_for_guidance(text_labels, empty_share)                
+        embeddings = encode_texts_batch(model_prompts, tokenizer, model)
 
         return {
-            "vae": np.array(valid_latents),
+            "vae": latents_np,
             "embeddings": embeddings,
             "text": text_labels,
-            "width": valid_widths,
-            "height": valid_heights
+            "width": widths,
+            "height": heights
         }
         
     except Exception as e:
         print(f"Критическая ошибка в encode_to_latents: {e}")
         raise
+        
 
 # ---------------- 5️⃣ Обработка папки с изображениями и текстами ----------------
 def process_folder(folder_path, limit=None):
@@ -282,7 +298,7 @@ def process_folder(folder_path, limit=None):
     print(f"Найдено {len(image_paths)} изображений с текстовыми описаниями")
     return image_paths, text_paths, width, height
     
-def process_in_chunks(image_paths, text_paths, width, height, chunk_size=50000, batch_size=1):
+def process_in_chunks(image_paths, text_paths, width, height, chunk_size=10000, batch_size=1):
     total_files = len(image_paths)
     start_time = time.time()
     chunks = range(0, total_files, chunk_size)
@@ -373,7 +389,7 @@ image_paths, text_paths, width, height = process_folder(folder_path,limit)
 print(f"Всего найдено {len(image_paths)} изображений")
 
 # Обработка с чанкованием
-process_in_chunks(image_paths, text_paths, width, height, chunk_size=100000, batch_size=batch_size)
+process_in_chunks(image_paths, text_paths, width, height, chunk_size=10000, batch_size=batch_size)
 
 # Объединение чанков в финальный датасет
 combine_chunks(temp_path, save_path)

model_index.json

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:85e884f29f7a6282a634d90350933aa68021326035a0980072667757c3bc9112
-size 476
+oid sha256:5b101caa6806403e974c161e47a59251ab322afe5ef583bcc7bb40fba0b3a25a
+size 415

pipeline_sdxs.py

@@ -25,44 +25,6 @@ class SdxsPipeline(DiffusionPipeline):
             unet=unet, scheduler=scheduler
         )
 
-        # Get the model path, which is either provided directly or from internal dict
-        model_path = None
-        if hasattr(self, '_internal_dict') and self._internal_dict.get('_name_or_path'):
-            model_path = self._internal_dict.get('_name_or_path')
-        
-        # Get device and dtype from existing components
-        device = "cuda"
-        dtype = torch.float16
-        
-        # Always load text_projector, regardless of whether one was provided
-        projector_path = None
-        
-        # Try to find projector path
-        if model_path and os.path.exists(f"{model_path}/text_projector"):
-            projector_path = f"{model_path}/text_projector"
-        elif os.path.exists("./text_projector"):
-            projector_path = "./text_projector"
-        
-        if projector_path:
-            # Create and load projector
-            try:
-                with open(f"{projector_path}/config.json", "r") as f:
-                    projector_config = json.load(f)
-                
-                # Create Linear layer with bias=False
-                self.text_projector = nn.Linear(
-                    in_features=projector_config["in_features"],
-                    out_features=projector_config["out_features"],
-                    bias=False
-                )
-                
-                # Load the state dict using safetensors
-                self.text_projector.load_state_dict(load_file(f"{projector_path}/model.safetensors"))
-                self.text_projector.to(device=device, dtype=dtype)
-                print(f"Successfully loaded text_projector from {projector_path}",device, dtype)
-            except Exception as e:
-                print(f"Error loading text_projector: {e}")
-
         self.vae_scale_factor = 8
 
     
@@ -88,7 +50,7 @@ class SdxsPipeline(DiffusionPipeline):
                 
                 text_inputs = self.tokenizer(
                     prompt, return_tensors="pt", padding="max_length",
-                    max_length=512, truncation=True
+                    max_length=150, truncation=True
                 ).to(device)
                 
                 # Получаем эмбеддинги

requirements.txt

@@ -1,6 +1,3 @@
-# torch>=2.6.0
-# torchvision>=0.21.0
-# torchaudio>=2.6.0
 diffusers>=0.32.2
 accelerate>=1.5.2
 datasets>=3.5.0

scheduler/scheduler_config.json

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:2e919ad3cde5f0bdf9529c68ee7c3306b1ceef40245778d29050d58ebc074158
-size 507
+oid sha256:ad559f5b938d5eac3c1f95f9d96dcfc8d0e4f9780f65755aa0d7e924c0f981c9
+size 482

text_encoder/config.json

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:87131a858ee394af6afae023f733cdebc36eda2ccbed27c36bc887cfae427392
-size 721
+oid sha256:39738c799f70b98f75dace818f14bfffb0b34fad25d41b80b6236e20a731c740
+size 1359

text_projector/config.json → text_encoder/generation_config.json

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:ae2f211593cd2cc736bf8617bcb0a5e6abd4db0265170de82ae03b7a6664feda
-size 83
+oid sha256:d3e057bbca66b92f33a8bdc6a1301014e0e4ab69b3b3fd2e442d9fe0c69f3431
+size 214

text_encoder/model.fp16.safetensors

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:107fe15da52fe6d13d877512fa36861d1100534d1b9b88015ad9fd017db095a7
-size 1119825680

text_encoder/model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a3184342876e1e64cd98164dc01aaaf69177a9d0ed80725816e3225977ab2982
+size 1192134784

vae/config-orig-wrong.json → tokenizer/added_tokens.json

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:992be219657d2876df23100f817d060c9f9aa497358ad30d7282103d96823d1f
-size 819
+oid sha256:c0284b582e14987fbd3d5a2cb2bd139084371ed9acbae488829a1c900833c680
+size 707

tokenizer/chat_template.jinja

@@ -0,0 +1,89 @@
+{%- if tools %}
+    {{- '<|im_start|>system\n' }}
+    {%- if messages[0].role == 'system' %}
+        {{- messages[0].content + '\n\n' }}
+    {%- endif %}
+    {{- "# Tools\n\nYou may call one or more functions to assist with the user query.\n\nYou are provided with function signatures within <tools></tools> XML tags:\n<tools>" }}
+    {%- for tool in tools %}
+        {{- "\n" }}
+        {{- tool | tojson }}
+    {%- endfor %}
+    {{- "\n</tools>\n\nFor each function call, return a json object with function name and arguments within <tool_call></tool_call> XML tags:\n<tool_call>\n{\"name\": <function-name>, \"arguments\": <args-json-object>}\n</tool_call><|im_end|>\n" }}
+{%- else %}
+    {%- if messages[0].role == 'system' %}
+        {{- '<|im_start|>system\n' + messages[0].content + '<|im_end|>\n' }}
+    {%- endif %}
+{%- endif %}
+{%- set ns = namespace(multi_step_tool=true, last_query_index=messages|length - 1) %}
+{%- for message in messages[::-1] %}
+    {%- set index = (messages|length - 1) - loop.index0 %}
+    {%- if ns.multi_step_tool and message.role == "user" and message.content is string and not(message.content.startswith('<tool_response>') and message.content.endswith('</tool_response>')) %}
+        {%- set ns.multi_step_tool = false %}
+        {%- set ns.last_query_index = index %}
+    {%- endif %}
+{%- endfor %}
+{%- for message in messages %}
+    {%- if message.content is string %}
+        {%- set content = message.content %}
+    {%- else %}
+        {%- set content = '' %}
+    {%- endif %}
+    {%- if (message.role == "user") or (message.role == "system" and not loop.first) %}
+        {{- '<|im_start|>' + message.role + '\n' + content + '<|im_end|>' + '\n' }}
+    {%- elif message.role == "assistant" %}
+        {%- set reasoning_content = '' %}
+        {%- if message.reasoning_content is string %}
+            {%- set reasoning_content = message.reasoning_content %}
+        {%- else %}
+            {%- if '</think>' in content %}
+                {%- set reasoning_content = content.split('</think>')[0].rstrip('\n').split('<think>')[-1].lstrip('\n') %}
+                {%- set content = content.split('</think>')[-1].lstrip('\n') %}
+            {%- endif %}
+        {%- endif %}
+        {%- if loop.index0 > ns.last_query_index %}
+            {%- if loop.last or (not loop.last and reasoning_content) %}
+                {{- '<|im_start|>' + message.role + '\n<think>\n' + reasoning_content.strip('\n') + '\n</think>\n\n' + content.lstrip('\n') }}
+            {%- else %}
+                {{- '<|im_start|>' + message.role + '\n' + content }}
+            {%- endif %}
+        {%- else %}
+            {{- '<|im_start|>' + message.role + '\n' + content }}
+        {%- endif %}
+        {%- if message.tool_calls %}
+            {%- for tool_call in message.tool_calls %}
+                {%- if (loop.first and content) or (not loop.first) %}
+                    {{- '\n' }}
+                {%- endif %}
+                {%- if tool_call.function %}
+                    {%- set tool_call = tool_call.function %}
+                {%- endif %}
+                {{- '<tool_call>\n{"name": "' }}
+                {{- tool_call.name }}
+                {{- '", "arguments": ' }}
+                {%- if tool_call.arguments is string %}
+                    {{- tool_call.arguments }}
+                {%- else %}
+                    {{- tool_call.arguments | tojson }}
+                {%- endif %}
+                {{- '}\n</tool_call>' }}
+            {%- endfor %}
+        {%- endif %}
+        {{- '<|im_end|>\n' }}
+    {%- elif message.role == "tool" %}
+        {%- if loop.first or (messages[loop.index0 - 1].role != "tool") %}
+            {{- '<|im_start|>user' }}
+        {%- endif %}
+        {{- '\n<tool_response>\n' }}
+        {{- content }}
+        {{- '\n</tool_response>' }}
+        {%- if loop.last or (messages[loop.index0 + 1].role != "tool") %}
+            {{- '<|im_end|>\n' }}
+        {%- endif %}
+    {%- endif %}
+{%- endfor %}
+{%- if add_generation_prompt %}
+    {{- '<|im_start|>assistant\n' }}
+    {%- if enable_thinking is defined and enable_thinking is false %}
+        {{- '<think>\n\n</think>\n\n' }}
+    {%- endif %}
+{%- endif %}

tokenizer/special_tokens_map.json

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:8c785abebea9ae3257b61681b4e6fd8365ceafde980c21970d001e834cf10835
-size 964
+oid sha256:76862e765266b85aa9459767e33cbaf13970f327a0e88d1c65846c2ddd3a1ecd
+size 613

tokenizer/tokenizer.json

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:3ffb37461c391f096759f4a9bbbc329da0f36952f88bab061fcf84940c022e98
-size 17082999
+oid sha256:aeb13307a71acd8fe81861d94ad54ab689df773318809eed3cbe794b4492dae4
+size 11422654

tokenizer/tokenizer_config.json

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:ccf223ba3d5b3cc7fa6c3bf451f3bb40557a5c92b0aa33f63d17802ff1a96fd9
-size 1178
+oid sha256:443bfa629eb16387a12edbf92a76f6a6f10b2af3b53d87ba1550adfcf45f7fa0
+size 5404

text_projector/model.safetensors → tokenizer/vocab.json

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:6e7060a387b4a6419f9d1d852759cb5b94541a1845e996f6062a07462d8b7b6a
-size 2359384
+oid sha256:ca10d7e9fb3ed18575dd1e277a2579c16d108e32f27439684afa0e10b1440910
+size 2776833

train.py

@@ -8,7 +8,7 @@ from torch.utils.data.distributed import DistributedSampler
 from torch.optim.lr_scheduler import LambdaLR
 from collections import defaultdict
 from torch.optim.lr_scheduler import LambdaLR
-from diffusers import UNet2DConditionModel, AutoencoderKL, DDPMScheduler
+from diffusers import UNet2DConditionModel, AutoencoderKLWan,AutoencoderKL
 from accelerate import Accelerator
 from datasets import load_from_disk
 from tqdm import tqdm
@@ -26,15 +26,16 @@ import torch.nn.functional as F
 from collections import deque
 
 # --------------------------- Параметры ---------------------------
-ds_path = "datasets/384"
-project = "micro"
-batch_size = 32
-base_learning_rate = 6e-5
-min_learning_rate = 1e-5
-num_epochs = 20
+ds_path = "/workspace/sdxs3d/datasets/mjnj_640"
+project = "unet"
+batch_size = 48
+base_learning_rate = 5e-5
+min_learning_rate = 6e-6
+num_epochs = 40
 # samples/save per epoch
-sample_interval_share = 10
-use_wandb = True
+sample_interval_share = 4
+use_wandb = False
+use_comet_ml = True
 save_model = True
 use_decay = True
 fbp = False # fused backward pass
@@ -42,19 +43,19 @@ optimizer_type = "adam8bit"
 torch_compile = False
 unet_gradient = True
 clip_sample = False #Scheduler
-fixed_seed = False
+fixed_seed = True
 shuffle = True
+comet_ml_api_key = "Agctp26mbqnoYrrlvQuKSTk6r"  # Добавлен API ключ для Comet ML
+comet_ml_workspace = "recoilme"  # Добавлен workspace для Comet ML
 torch.backends.cuda.matmul.allow_tf32 = True
 torch.backends.cudnn.allow_tf32 = True
 torch.backends.cuda.enable_mem_efficient_sdp(False)
 dtype = torch.float32
-save_barrier = 1.03
-warmup_percent = 0.01
-dispersive_temperature=0.5
-dispersive_weight= 0.05
-percentile_clipping = 95 # 8bit optim
-betta2 = 0.97
-eps = 1e-6
+save_barrier = 1.006
+warmup_percent = 0.005
+percentile_clipping = 99 # 8bit optim
+betta2 = 0.99
+eps = 1e-8
 clip_grad_norm = 1.0
 steps_offset = 0 # Scheduler
 limit = 0
@@ -70,7 +71,7 @@ device = accelerator.device
 # Параметры для диффузии
 n_diffusion_steps = 50
 samples_to_generate = 12
-guidance_scale = 5
+guidance_scale = 4
 
 # Папки для сохранения результатов
 generated_folder = "samples"
@@ -86,16 +87,6 @@ if fixed_seed:
     if torch.cuda.is_available():
         torch.cuda.manual_seed_all(seed)
 
-# --- Пропорции лоссов и окно медианного нормирования (КОЭФ., не значения) ---
-# CHANGED: добавлен huber и dispersive в пропорции, суммы = 1.0
-loss_ratios = {
-    "mse":   0.50,
-    "mae":   0.25,
-    "huber": 0.20,
-    "dispersive": 0.05,
-}
-median_coeff_steps = 256  # за сколько шагов считать медианные коэффициенты
-
 # --------------------------- Параметры LoRA ---------------------------
 lora_name = "" 
 lora_rank = 32
@@ -109,116 +100,68 @@ def sample_timesteps_bias(
     progress: float,           # [0..1]
     num_train_timesteps: int,  # обычно 1000
     steps_offset: int = 0,
-    device=None
+    device=None,
+    mode: str = "beta",        # "beta", "uniform"
 ) -> torch.Tensor:
     """
-    Возвращает псевдослучайные timesteps во всём диапазоне,
-    но с bias: на старте больше вероятности брать max (999),
-    к концу — больше вероятности брать min (0).
-
-    FIX: исправлена формула alpha/beta (раньше было перевёрнуто).
+    Возвращает timesteps с разным bias:
+    - beta   : как раньше (сдвиг в начало или конец в зависимости от progress)
+    - normal : около середины (гауссовое распределение)
+    - uniform: равномерно по всем timestep’ам
     """
-    # Параметры Beta-распределения (FIX: alpha и beta поменяны местами по логике)
-    alpha = 1.0 + 2.0 * (1.0 - progress)  # при progress=0 -> alpha ~10 (сдвиг к 1.0)
-    beta  = 1.0 + 2.0 * progress          # при progress=0 -> beta ~1
-
-    samples = torch.distributions.Beta(alpha, beta).sample((batch_size,)).to(device)
 
     max_idx = num_train_timesteps - 1 - steps_offset
-    timesteps = steps_offset + (samples * max_idx).long()
+
+    if mode == "beta":
+        alpha = 1.0 + .5 * (1.0 - progress)
+        beta  = 1.0 + .5 * progress
+        samples = torch.distributions.Beta(alpha, beta).sample((batch_size,))
+
+    elif mode == "uniform":
+        samples = torch.rand(batch_size)
+
+    else:
+        raise ValueError(f"Unknown mode: {mode}")
+
+    timesteps = steps_offset + (samples * max_idx).long().to(device)
     return timesteps
 
-# Нормализация лоссов по медианам: считаем КОЭФФИЦИЕНТЫ
-class MedianLossNormalizer:
-    def __init__(self, desired_ratios: dict, window_steps: int):
-        # нормируем доли на случ��й, если сумма != 1
-        s = sum(desired_ratios.values())
-        self.ratios = {k: (v / s) 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, losses: dict):
-        """
-        losses: dict ключ->тензор (значения лоссов)
-        Поведение:
-          - буферим ABS(l) только для активных (ratio>0) лоссов
-          - coeff = ratio / median(abs(loss))
-          - total = sum(coeff * loss) по активным лоссам
-        CHANGED: буферим abs() — чтобы медиана была положительной и не ломала деление.
-        """
-        # буферим только активные лоссы
-        for k, v in losses.items():
-            if k in self.buffers and self.ratios.get(k, 0) > 0:
-                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}
-
-        # суммируем только по активным (ratio>0)
-        total = sum(coeffs[k] * losses[k] for k in coeffs if self.ratios.get(k, 0) > 0)
-        return total, coeffs, meds
-
-# создаём normalizer после определения loss_ratios
-normalizer = MedianLossNormalizer(loss_ratios, median_coeff_steps)
-
-class AccelerateDispersiveLoss:
-    def __init__(self, accelerator, temperature=0.5, weight=0.5):
-        self.accelerator = accelerator
-        self.temperature = temperature
-        self.weight = weight
-        self.activations = []
-        self.hooks = []
-    
-    def register_hooks(self, model, target_layer="down_blocks.0"):
-        unwrapped_model = self.accelerator.unwrap_model(model)
-        print("=== Поиск слоев в unwrapped модели ===")
-        for name, module in unwrapped_model.named_modules():
-            if target_layer in name:
-                hook = module.register_forward_hook(self.hook_fn)
-                self.hooks.append(hook)
-                print(f"✅ Хук зарегистрирован на: {name}")
-                break
-    
-    def hook_fn(self, module, input, output):
-        if isinstance(output, tuple):
-            activation = output[0]
-        else:
-            activation = output
-        if len(activation.shape) > 2:
-            activation = activation.view(activation.shape[0], -1)
-        self.activations.append(activation.detach().clone())
-        
-    def compute_dispersive_loss(self):
-        if not self.activations:
-            return torch.tensor(0.0, requires_grad=True, device=device)
-        local_activations = self.activations[-1].float()
-        batch_size = local_activations.shape[0]
-        if batch_size < 2:
-            return torch.tensor(0.0, requires_grad=True, device=device)
-        sf = local_activations / torch.norm(local_activations, dim=1, keepdim=True)
-        distance = torch.nn.functional.pdist(sf.float(), p=2) ** 2
-        exp_neg_dist = torch.exp(-distance / self.temperature) + 1e-5
-        dispersive_loss = torch.log(torch.mean(exp_neg_dist))
-        return dispersive_loss
-    
-    def clear_activations(self):
-        self.activations.clear()
-    
-    def remove_hooks(self):
-        for hook in self.hooks:
-            hook.remove()
-        self.hooks.clear()
+def logit_normal_samples(shape, mu=0.0, sigma=1.0, device=None, dtype=None):
+    normal_samples = torch.normal(mean=mu, std=sigma, size=shape, device=device, dtype=dtype)
+
+    logit_normal_samples = torch.sigmoid(normal_samples)
 
+    return logit_normal_samples
 
 # --------------------------- Инициализация WandB ---------------------------
-if use_wandb and accelerator.is_main_process:
-    wandb.init(project=project+lora_name, config={
-        "batch_size": batch_size,
-        "base_learning_rate": base_learning_rate,
-        "num_epochs": num_epochs,
-        "fbp": fbp,
-        "optimizer_type": optimizer_type,
-    })
+if accelerator.is_main_process:
+    if use_wandb:
+        wandb.init(project=project+lora_name, config={
+            "batch_size": batch_size,
+            "base_learning_rate": base_learning_rate,
+            "num_epochs": num_epochs,
+            "fbp": fbp,
+            "optimizer_type": optimizer_type,
+        })
+    if use_comet_ml:
+        from comet_ml import Experiment
+        comet_experiment = Experiment(
+            api_key=comet_ml_api_key,
+            project_name=project,
+            workspace=comet_ml_workspace
+        )
+        # Логируем гиперпараметры в Comet ML
+        hyper_params = {
+            "batch_size": batch_size,
+            "base_learning_rate": base_learning_rate,
+            "min_learning_rate": min_learning_rate,
+            "num_epochs": num_epochs,
+            "n_diffusion_steps": n_diffusion_steps,
+            "guidance_scale": guidance_scale,
+            "optimizer_type": optimizer_type,
+            "mixed_precision": mixed_precision,
+        }
+        comet_experiment.log_parameters(hyper_params)
 
 # Включение Flash Attention 2/SDPA
 torch.backends.cuda.enable_flash_sdp(True)
@@ -228,15 +171,28 @@ gen.manual_seed(seed)
 
 # --------------------------- Загрузка моделей ---------------------------
 # VAE загружается на CPU для экономии GPU-памяти (как в твоём оригинальном коде)
-vae = AutoencoderKL.from_pretrained("vae", variant="fp16").to(device="cpu", dtype=torch.float16).eval()
-
-# DDPMScheduler с V_Prediction и Zero-SNR
-scheduler = DDPMScheduler(
-    num_train_timesteps=1000,
-    prediction_type="v_prediction",
-    rescale_betas_zero_snr=True,
-    clip_sample = clip_sample,
-    steps_offset = steps_offset
+vae = AutoencoderKL.from_pretrained("AiArtLab/simplevae", subfolder="vae", torch_dtype=dtype).to("cpu").eval()
+
+shift_factor = getattr(vae.config, "shift_factor", 0.0)
+if shift_factor is None:
+    shift_factor = 0.0
+
+scaling_factor = getattr(vae.config, "scaling_factor", 1.0)
+if scaling_factor is None:
+    scaling_factor = 1.0
+    
+latents_mean = getattr(vae.config, "latents_mean", None)
+latents_std = getattr(vae.config, "latents_std", None)
+
+from diffusers import FlowMatchEulerDiscreteScheduler
+
+# Подстрой под свои параметры
+num_train_timesteps = 1000
+
+scheduler = FlowMatchEulerDiscreteScheduler(
+    num_train_timesteps=num_train_timesteps,
+    #shift=3.0,               # пример; подбирается при необходимости
+    #use_dynamic_shifting=True
 )
 
 
@@ -343,22 +299,6 @@ def collate_fn_simple(batch):
     embeddings = torch.tensor(np.array([item["embeddings"] for item in batch])).to(device,dtype=dtype)
     return latents, embeddings
 
-def collate_fn(batch):
-    if not batch:
-        return [], []
-    ref_vae_shape = np.array(batch[0]["vae"]).shape
-    ref_embed_shape = np.array(batch[0]["embeddings"]).shape
-    valid_latents = []
-    valid_embeddings = []
-    for item in batch:
-        if (np.array(item["vae"]).shape == ref_vae_shape and 
-            np.array(item["embeddings"]).shape == ref_embed_shape):
-            valid_latents.append(item["vae"])
-            valid_embeddings.append(item["embeddings"])
-    latents = torch.tensor(np.array(valid_latents)).to(device,dtype=dtype)
-    embeddings = torch.tensor(np.array(valid_embeddings)).to(device,dtype=dtype)
-    return latents, embeddings
-
 batch_sampler = DistributedResolutionBatchSampler(
         dataset=dataset,
         batch_size=batch_size,
@@ -381,11 +321,6 @@ latest_checkpoint = os.path.join(checkpoints_folder, project)
 if os.path.isdir(latest_checkpoint):
     print("Загружаем UNet из чекпоинта:", latest_checkpoint)
     unet = UNet2DConditionModel.from_pretrained(latest_checkpoint).to(device=device,dtype=dtype)
-    if torch_compile:
-        print("compiling")
-        torch.set_float32_matmul_precision('high')
-        unet = torch.compile(unet)
-        print("compiling - ok")
     if unet_gradient:
         unet.enable_gradient_checkpointing()
     unet.set_use_memory_efficient_attention_xformers(False)
@@ -395,13 +330,6 @@ if os.path.isdir(latest_checkpoint):
         print(f"Ошибка при включении SDPA: {e}")
         unet.set_use_memory_efficient_attention_xformers(True)
 
-    # Создаём hook для dispersive только если нужно
-    if loss_ratios.get("dispersive", 0) > 0:
-        dispersive_hook = AccelerateDispersiveLoss(
-            accelerator=accelerator,
-            temperature=dispersive_temperature,
-            weight=dispersive_weight
-        )
 else:
     # FIX: если чекпоинта нет — прекращаем с понятной ошибкой (лучше, чем неожиданные NameError дальше)
     raise FileNotFoundError(f"UNet checkpoint not found at {latest_checkpoint}. Положи UNet чекпоинт в {latest_checkpoint} или укажи другой путь.")
@@ -454,25 +382,13 @@ else:
 def create_optimizer(name, params):
     if name == "adam8bit":
         return bnb.optim.AdamW8bit(
-            params, lr=base_learning_rate, betas=(0.9, betta2), eps=eps, weight_decay=0.001,
+            params, lr=base_learning_rate, betas=(0.9, betta2), eps=eps, weight_decay=0.01,
             percentile_clipping=percentile_clipping
         )
     elif name == "adam":
         return torch.optim.AdamW(
             params, lr=base_learning_rate, betas=(0.9, 0.999), eps=1e-8, weight_decay=0.01
         )
-    elif name == "lion8bit":
-        return bnb.optim.Lion8bit(
-            params, lr=base_learning_rate, betas=(0.9, 0.97), weight_decay=0.01,
-            percentile_clipping=percentile_clipping
-        )
-    elif name == "adafactor":
-        from transformers import Adafactor
-        return Adafactor(
-            params, lr=base_learning_rate, scale_parameter=True, relative_step=False,
-            warmup_init=False, eps=(1e-30, 1e-3), clip_threshold=1.0,
-            beta1=0.9, weight_decay=0.01
-        )
     else:
         raise ValueError(f"Unknown optimizer: {name}")
 
@@ -504,23 +420,96 @@ else:
         if torch.isnan(param).any() or torch.isinf(param).any():
             print(f"[rank {accelerator.process_index}] NaN/Inf in {name}")
     unet, optimizer, lr_scheduler = accelerator.prepare(unet, optimizer, lr_scheduler)
-    
-# Регистрация хуков ПОСЛЕ prepare
-if loss_ratios.get("dispersive", 0) > 0:
-    dispersive_hook.register_hooks(unet, "down_blocks.2")
+
+if torch_compile:
+    print("compiling")
+    torch.set_float32_matmul_precision('high')
+    torch.backends.cudnn.allow_tf32 = True
+    torch.backends.cuda.matmul.allow_tf32 = True
+    unet = torch.compile(unet)#, mode='max-autotune')
+    print("compiling - ok")
 
 # --------------------------- Фиксированные семплы для генерации ---------------------------
 fixed_samples = get_fixed_samples_by_resolution(dataset)
 
+def get_negative_embedding(neg_prompt="", batch_size=1):
+    """
+    Возвращает эмбеддинг негативного промпта с батчем.
+    Загружает модели, вычисляет эмбеддинг, выгружает модели на CPU.
+    """
+    import torch
+    from transformers import AutoTokenizer, AutoModel
+
+    # Настройки
+    dtype = torch.float16
+    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+
+    # Загрузка моделей (если ещё не загружены)
+    if not hasattr(get_negative_embedding, "tokenizer"):
+        get_negative_embedding.tokenizer = AutoTokenizer.from_pretrained(
+            "Qwen/Qwen3-Embedding-0.6B", padding_side="left"
+        )
+        get_negative_embedding.text_model = AutoModel.from_pretrained(
+            "Qwen/Qwen3-Embedding-0.6B"
+        ).to(device).eval()
+
+    # Вспомогательная функция для пулинга
+    def last_token_pool(last_hidden_states, attention_mask):
+        left_padding = (attention_mask[:, -1].sum() == attention_mask.shape[0])
+        if left_padding:
+            return last_hidden_states[:, -1]
+        else:
+            sequence_lengths = attention_mask.sum(dim=1) - 1
+            batch_size = last_hidden_states.shape[0]
+            return last_hidden_states[
+                torch.arange(batch_size, device=last_hidden_states.device),
+                sequence_lengths
+            ]
+
+    # Вычисление эмбеддинга
+    def encode_texts(texts, max_length=512):
+        with torch.inference_mode():
+            toks = get_negative_embedding.tokenizer(
+                texts, return_tensors="pt", padding="max_length", truncation=True, max_length=max_length
+            ).to(device)
+            outs = get_negative_embedding.text_model(**toks)
+            emb = last_token_pool(outs.last_hidden_state, toks["attention_mask"])
+            emb = emb.unsqueeze(1)  # Добавляем размерность sequence_length
+            return emb
+
+    # Возвращаем эмбеддинг
+    if not neg_prompt:
+        hidden_dim = 1024  # Размерность эмбеддинга Qwen3-Embedding-0.6B
+        seq_len = 150
+        return torch.zeros((batch_size, seq_len, hidden_dim), dtype=dtype, device=device)
+
+    uncond_emb = encode_texts([neg_prompt]).to(dtype=dtype, device=device)
+    uncond_emb = uncond_emb.repeat(batch_size, 1, 1)  # Добавляем батч
+
+    # Выгружаем модели
+    if hasattr(get_negative_embedding, "text_model"):
+        get_negative_embedding.text_model = get_negative_embedding.text_model.to("cpu")
+    if hasattr(get_negative_embedding, "tokenizer"):
+        del get_negative_embedding.tokenizer  # Освобождаем память
+    torch.cuda.empty_cache()
+
+    return uncond_emb
+    
+uncond_emb = get_negative_embedding("low quality")
+
 @torch.compiler.disable()
 @torch.no_grad()
-def generate_and_save_samples(fixed_samples_cpu, step):
+def generate_and_save_samples(fixed_samples_cpu,empty_embeddings, step):
     original_model = None
     try:
-        original_model = accelerator.unwrap_model(unet, keep_torch_compile=True).eval()
+        # безопасный unwrap: если компилировано, unwrap не нужен
+        if not torch_compile:
+            original_model = accelerator.unwrap_model(unet, keep_torch_compile=True).eval()
+        else:
+            original_model = unet.eval()
+
         vae.to(device=device).eval()  # временно подгружаем VAE на GPU для декодинга
 
-        scheduler.set_timesteps(n_diffusion_steps)
         
         all_generated_images = []
         all_captions = [] 
@@ -529,43 +518,66 @@ def generate_and_save_samples(fixed_samples_cpu, step):
             width, height = size
             sample_latents = sample_latents.to(dtype=dtype, device=device)
             sample_text_embeddings = sample_text_embeddings.to(dtype=dtype, device=device)
-
-            noise = torch.randn(
+        
+            # начальный шум
+            latents = torch.randn(
                 sample_latents.shape,
-                generator=gen,
                 device=device,
-                dtype=sample_latents.dtype
+                dtype=sample_latents.dtype,
+                generator=torch.Generator(device=device).manual_seed(seed)
             )
-            current_latents = noise.clone()
-            
-            if guidance_scale > 0:
-                empty_embeddings = torch.zeros_like(sample_text_embeddings, dtype=sample_text_embeddings.dtype, device=device)
-                text_embeddings_batch = torch.cat([empty_embeddings, sample_text_embeddings], dim=0)
-            else:
-                text_embeddings_batch = sample_text_embeddings
-            
+        
+            # подготовим timesteps через шедулер
+            scheduler.set_timesteps(n_diffusion_steps, device=device)
+        
             for t in scheduler.timesteps:
-                t_batch = t.repeat(current_latents.shape[0]).to(device)
-                if guidance_scale > 0:
-                    latent_model_input = torch.cat([current_latents] * 2)
-                else:
-                    latent_model_input = current_latents
+                # guidance: удваиваем батч
+                if guidance_scale != 1:
+                    latent_model_input = torch.cat([latents, latents], dim=0)
                 
-                latent_model_input_scaled = scheduler.scale_model_input(latent_model_input, t_batch)
-                noise_pred = original_model(latent_model_input_scaled, t_batch, text_embeddings_batch).sample               
+                    # empty_embeddings: [1, 1, hidden_dim] → повторяем по seq_len и batch
+                    seq_len = sample_text_embeddings.shape[1]
+                    hidden_dim = sample_text_embeddings.shape[2]
+                    empty_embeddings_exp = empty_embeddings.expand(-1, seq_len, hidden_dim)  # [1, seq_len, hidden_dim]
+                    empty_embeddings_exp = empty_embeddings_exp.repeat(sample_text_embeddings.shape[0], 1, 1)  # [batch, seq_len, hidden_dim]
                 
-                if guidance_scale > 0:
-                    noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
-                    noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
+                    text_embeddings_batch = torch.cat([empty_embeddings_exp, sample_text_embeddings], dim=0)
+                else:
+                    latent_model_input = latents
+                    text_embeddings_batch = sample_text_embeddings
+
+
+        
+                # предсказание потока (velocity)
+                model_out = original_model(latent_model_input, t, encoder_hidden_states=text_embeddings_batch)
+                flow = getattr(model_out, "sample", model_out)
+        
+                # guidance объединение
+                if guidance_scale != 1:
+                    flow_uncond, flow_cond = flow.chunk(2)
+                    flow = flow_uncond + guidance_scale * (flow_cond - flow_uncond)
+        
+                # шаг через scheduler
+                latents = scheduler.step(flow, t, latents).prev_sample
+        
+            current_latents = latents
 
-                current_latents = scheduler.step(noise_pred, t, current_latents).prev_sample
 
-            latent_for_vae = (current_latents.detach() / vae.config.scaling_factor) + getattr(vae.config, "shift_factor", 0.0)
-            decoded = vae.decode(latent_for_vae.to(torch.float16)).sample
+            # Параметры нормализации
+            latent_for_vae = current_latents.detach() / scaling_factor + shift_factor
+            
+            decoded = vae.decode(latent_for_vae.to(torch.float32)).sample
+            #decoded = decoded[:, :, 0, :, :]  # [3, H, W]
+            #print(decoded.ndim, decoded.shape)
             
             decoded_fp32 = decoded.to(torch.float32)
             for img_idx, img_tensor in enumerate(decoded_fp32):
-                img = (img_tensor / 2 + 0.5).clamp(0, 1).cpu().numpy().transpose(1, 2, 0)
+
+                # Форма: [3, H, W] -> преобразуем в [H, W, 3]
+                img = (img_tensor / 2 + 0.5).clamp(0, 1).cpu().numpy()
+                img = img.transpose(1, 2, 0)  # Из [3, H, W] в [H, W, 3]
+    
+                #img = (img_tensor / 2 + 0.5).clamp(0, 1).cpu().numpy().transpose(1, 2, 0)
                 if np.isnan(img).any():
                     print("NaNs found, saving stopped! Step:", step)
                 pil_img = Image.fromarray((img * 255).astype("uint8"))
@@ -589,7 +601,20 @@ def generate_and_save_samples(fixed_samples_cpu, step):
                 wandb.Image(img, caption=f"{all_captions[i]}")
                 for i, img in enumerate(all_generated_images)
             ]
-            wandb.log({"generated_images": wandb_images, "global_step": step})
+            wandb.log({"generated_images": wandb_images})
+        if use_comet_ml and accelerator.is_main_process:
+            for i, img in enumerate(all_generated_images):
+                comet_experiment.log_image(
+                    image_data=img,
+                    name=f"step_{step}_img_{i}",
+                    step=step,
+                    metadata={
+                        "caption": all_captions[i],
+                        "width": img.width,
+                        "height": img.height,
+                        "global_step": step
+                    }
+                )
     finally:
         # вернуть VAE на CPU (как было в твоём коде)
         vae.to("cpu")
@@ -603,44 +628,30 @@ def generate_and_save_samples(fixed_samples_cpu, step):
 if accelerator.is_main_process:
     if save_model:
         print("Генерация сэмплов до старта обучения...")
-        generate_and_save_samples(fixed_samples,0)
+        generate_and_save_samples(fixed_samples,uncond_emb,0)
 accelerator.wait_for_everyone()
 
 # Модифицируем функцию сохранения модели для поддержки LoRA
-def save_checkpoint(unet,variant=""):
+def save_checkpoint(unet, variant=""):
     if accelerator.is_main_process:
         if lora_name:
             save_lora_checkpoint(unet)
         else:
-            if variant!="":
-                accelerator.unwrap_model(unet.to(dtype=torch.float16)).save_pretrained(os.path.join(checkpoints_folder, f"{project}"),variant=variant)
+            # безопасный unwrap для компилированной модели
+            model_to_save = None
+            if not torch_compile:
+                model_to_save = accelerator.unwrap_model(unet)
             else:
-                accelerator.unwrap_model(unet).save_pretrained(os.path.join(checkpoints_folder, f"{project}"))
-            unet = unet.to(dtype=dtype)
+                model_to_save = unet
 
-def batch_pred_original_from_step(model_outputs, timesteps_tensor, noisy_latents, scheduler):
-    device = noisy_latents.device
-    dtype = noisy_latents.dtype
+            if variant != "":
+                model_to_save.to(dtype=torch.float16).save_pretrained(
+                    os.path.join(checkpoints_folder, f"{project}"), variant=variant
+                )
+            else:
+                model_to_save.save_pretrained(os.path.join(checkpoints_folder, f"{project}"))
 
-    available_ts = scheduler.timesteps
-    if not isinstance(available_ts, torch.Tensor):
-        available_ts = torch.tensor(available_ts, device="cpu")
-    else:
-        available_ts = available_ts.cpu()
-
-    B = model_outputs.shape[0]
-    preds = []
-    for i in range(B):
-        t_i = int(timesteps_tensor[i].item())
-        diffs = torch.abs(available_ts - t_i)
-        idx = int(torch.argmin(diffs).item())
-        t_for_step = int(available_ts[idx].item())
-        model_out_i = model_outputs[i:i+1]
-        noisy_latent_i = noisy_latents[i:i+1]
-        step_out = scheduler.step(model_out_i, t_for_step, noisy_latent_i)
-        preds.append(step_out.pred_original_sample)
-
-    return torch.cat(preds, dim=0).to(device=device, dtype=dtype)
+            unet = unet.to(dtype=dtype)
 
 # --------------------------- Тренировочный цикл ---------------------------
 if accelerator.is_main_process:
@@ -651,61 +662,41 @@ progress_bar = tqdm(total=total_training_steps, disable=not accelerator.is_local
 
 steps_per_epoch = len(dataloader)
 sample_interval = max(1, steps_per_epoch // sample_interval_share)
-min_loss = 1.
+min_loss = 2.
 
 for epoch in range(start_epoch, start_epoch + num_epochs):
     batch_losses = []
-    batch_tlosses = []
     batch_grads = []
     batch_sampler.set_epoch(epoch)
     accelerator.wait_for_everyone()
     unet.train()
-    print("epoch:",epoch)
+    #print("epoch:",epoch)
     for step, (latents, embeddings) in enumerate(dataloader):
         with accelerator.accumulate(unet):
             if save_model == False and step == 5 :
                 used_gb = torch.cuda.max_memory_allocated() / 1024**3
                 print(f"Шаг {step}: {used_gb:.2f} GB")
-
+                
+            # шум
             noise = torch.randn_like(latents, dtype=latents.dtype)
-
-            progress = global_step / max(1, total_training_steps - 1)
-            timesteps = sample_timesteps_bias(
-                batch_size=latents.shape[0],
-                progress=progress,
-                num_train_timesteps=scheduler.config.num_train_timesteps,
-                steps_offset=steps_offset,
-                device=device
-            )
-
-            noisy_latents = scheduler.add_noise(latents, noise, timesteps)
-
-            if loss_ratios.get("dispersive", 0) > 0:
-                dispersive_hook.clear_activations()
-        
+    
+            # берём t из [0, 1]
+            t = torch.rand(latents.shape[0], device=latents.device, dtype=latents.dtype)
+    
+            # интерполяция между x0 и шумом
+            noisy_latents = (1.0 - t.view(-1, 1, 1, 1)) * latents + t.view(-1, 1, 1, 1) * noise
+    
+            # делаем integer timesteps для UNet
+            timesteps = (t * scheduler.config.num_train_timesteps).long()
+    
+            # предсказание потока (Flow)
             model_pred = unet(noisy_latents, timesteps, embeddings).sample
-            target_pred = scheduler.get_velocity(latents, noise, timesteps)
-
-            # === Losses ===
-            losses_dict = {}
-
-            mse_loss = F.mse_loss(model_pred.float(), target_pred.float())
-            losses_dict["mse"] = mse_loss
-            losses_dict["mae"] = F.l1_loss(model_pred.float(), target_pred.float())
-
-            # CHANGED: Huber (smooth_l1) loss added
-            losses_dict["huber"] = F.smooth_l1_loss(model_pred.float(), target_pred.float())
-
-            # === Dispersive loss ===
-            if loss_ratios.get("dispersive", 0) > 0:
-                disp_raw = dispersive_hook.compute_dispersive_loss().to(device)  # может быть отрицательным
-                losses_dict["dispersive"] = dispersive_hook.weight * disp_raw
-            else:
-                losses_dict["dispersive"] = torch.tensor(0.0, device=device)
-
-            # === Нормализация всех лоссов ===
-            abs_for_norm = {k: losses_dict.get(k, torch.tensor(0.0, device=device)) for k in normalizer.ratios.keys()}
-            total_loss, coeffs, meds = normalizer.update_and_total(abs_for_norm)
+    
+            # таргет — векторное поле (= разность между конечными точками)
+            target = noise - latents   # или latents - noise?
+    
+            # MSE лосс
+            mse_loss = F.mse_loss(model_pred.float(), target.float())
 
             # Сохраняем для логов (мы сохраняем MSE отдельно — как показатель)
             batch_losses.append(mse_loss.detach().item())
@@ -714,7 +705,7 @@ for epoch in range(start_epoch, start_epoch + num_epochs):
                 accelerator.wait_for_everyone()
                 
             # Backward
-            accelerator.backward(total_loss)
+            accelerator.backward(mse_loss)
 
             if (global_step % 100 == 0) or (global_step % sample_interval == 0):
                 accelerator.wait_for_everyone()
@@ -729,68 +720,63 @@ for epoch in range(start_epoch, start_epoch + num_epochs):
                     lr_scheduler.step()
                     optimizer.zero_grad(set_to_none=True)
 
-            global_step += 1
-            progress_bar.update(1)
+            if accelerator.sync_gradients:
+                global_step += 1
+                progress_bar.update(1)
+                # Логируем метрики
+                if accelerator.is_main_process:
+                    if fbp:
+                        current_lr = base_learning_rate
+                    else:
+                        current_lr = lr_scheduler.get_last_lr()[0]
+                    batch_grads.append(grad)
+    
+                    log_data = {}
+                    log_data["loss"] = mse_loss.detach().item()
+                    log_data["lr"] = current_lr
+                    log_data["grad"] = grad
+                    if accelerator.sync_gradients:
+                        if use_wandb:
+                            wandb.log(log_data, step=global_step)
+                        if use_comet_ml:
+                            comet_experiment.log_metrics(log_data, step=global_step)
+    
+                    # Генерируем сэмплы с заданным интервалом
+                    if global_step % sample_interval == 0:
+                        generate_and_save_samples(fixed_samples,uncond_emb, global_step)
+                        last_n = sample_interval
+    
+                        if save_model:
+                            avg_sample_loss = np.mean(batch_losses[-sample_interval:]) if len(batch_losses) > 0 else 0.0
+                            print("saving:", avg_sample_loss < min_loss * save_barrier, "Avg:", avg_sample_loss)
+                            if avg_sample_loss is not None and avg_sample_loss < min_loss * save_barrier:
+                                min_loss = avg_sample_loss
+                                save_checkpoint(unet)
 
-            # Логируем метрики
-            if accelerator.is_main_process:
-                if fbp:
-                    current_lr = base_learning_rate
-                else:
-                    current_lr = lr_scheduler.get_last_lr()[0]
-                batch_tlosses.append(total_loss.detach().item())
-                batch_grads.append(grad)
-
-                # Логируем только активные лоссы (ratio>0)
-                active_keys = [k for k, v in loss_ratios.items() if v > 0]
-                log_data = {}
-                for k in active_keys:
-                    v = losses_dict.get(k, None)
-                    if v is None:
-                        continue
-                    log_data[f"loss/{k}"] = (v.item() if isinstance(v, torch.Tensor) else float(v))
-
-                log_data["loss/total"] = float(total_loss.item())
-                log_data["loss/lr"] = current_lr
-                for k, c in coeffs.items():
-                    log_data[f"coeff/{k}"] = float(c)
-                if use_wandb and accelerator.sync_gradients:
-                    wandb.log(log_data, step=global_step)
-
-                # Генерируем сэмплы с заданным интервалом
-                if global_step % sample_interval == 0:
-                    generate_and_save_samples(fixed_samples,global_step)
-                    last_n = sample_interval
-                    avg_loss = float(np.mean(batch_losses[-last_n:])) if len(batch_losses) > 0 else 0.0
-                    avg_tloss = float(np.mean(batch_tlosses[-last_n:])) if len(batch_tlosses) > 0 else 0.0
-                    avg_grad = float(np.mean(batch_grads[-last_n:])) if len(batch_grads) > 0 else 0.0
-                    print(f"Эпоха {epoch}, шаг {global_step}, средний лосс: {avg_loss:.6f}, grad: {avg_grad:.6f}")
-
-                    if save_model:
-                        print("saving:",avg_loss < min_loss*save_barrier)
-                        if avg_loss < min_loss*save_barrier:
-                            min_loss = avg_loss
-                            save_checkpoint(unet)
-                    if use_wandb:
-                        avg_data = {}
-                        avg_data["avg/loss"] = avg_loss
-                        avg_data["avg/tloss"] = avg_tloss
-                        avg_data["avg/grad"] = avg_grad
-                        wandb.log(avg_data, step=global_step)
 
     if accelerator.is_main_process:
-        avg_epoch_loss = np.mean(batch_losses) if len(batch_losses)>0 else 0.0
+        # local averages
+        avg_epoch_loss = np.mean(batch_losses) if len(batch_losses) > 0 else 0.0
+        avg_epoch_grad = np.mean(batch_grads) if len(batch_grads) > 0 else 0.0
+
         print(f"\nЭпоха {epoch} завершена. Средний лосс: {avg_epoch_loss:.6f}")
+        log_data_ep = {
+                        "epoch_loss": avg_epoch_loss,
+                        "epoch_grad": avg_epoch_grad,
+                        "epoch": epoch + 1,
+                    }
         if use_wandb:
-            wandb.log({"epoch_loss": avg_epoch_loss, "epoch": epoch+1})
+            wandb.log(log_data_ep)
+        if use_comet_ml:
+            comet_experiment.log_metrics(log_data_ep)
 
 # Завершение обучения - сохраняем финальную модель
-if loss_ratios.get("dispersive", 0) > 0:
-    dispersive_hook.remove_hooks()
 if accelerator.is_main_process:
     print("Обучение завершено! Сохраняем финальную модель...")
     if save_model:
         save_checkpoint(unet,"fp16")
+    if use_comet_ml:
+        comet_experiment.end()
 accelerator.free_memory()
 if torch.distributed.is_initialized():
     torch.distributed.destroy_process_group()

train_flow_test.py

@@ -1,770 +0,0 @@
-import os
-import math
-import torch
-import numpy as np
-import matplotlib.pyplot as plt
-from torch.utils.data import DataLoader, Sampler
-from collections import defaultdict
-from torch.optim.lr_scheduler import LambdaLR
-from diffusers import UNet2DConditionModel, AutoencoderKL, DDPMScheduler
-from accelerate import Accelerator
-from datasets import load_from_disk
-from tqdm import tqdm
-from PIL import Image,ImageOps
-import wandb
-import random
-import gc
-from accelerate.state import DistributedType
-from torch.distributed import broadcast_object_list
-from torch.utils.checkpoint import checkpoint
-from diffusers.models.attention_processor import AttnProcessor2_0
-from datetime import datetime
-import bitsandbytes as bnb
-
-
-# region scheduler start
-
-
-#@title scheduler
-# Copyright 2024 The HuggingFace Team. All rights reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-
-# DISCLAIMER: This code is strongly influenced by https://github.com/leffff/euler-scheduler
-
-from dataclasses import dataclass
-from typing import Tuple, Any, Optional, Union
-
-import torch
-from diffusers.configuration_utils import ConfigMixin, register_to_config
-from diffusers.utils import BaseOutput
-from diffusers.schedulers.scheduling_utils import KarrasDiffusionSchedulers, SchedulerMixin
-
-
-@dataclass
-class FlowMatchingEulerSchedulerOutput(BaseOutput):
-    """
-    Output class for the scheduler's `step` function output.
-
-    Args:
-        prev_sample (`torch.Tensor` of shape `(batch_size, num_channels, height, width)` for images):
-            Computed sample `(x_{t-1})` of previous timestep (which in flow-matching notation should be noted as
-            `(x_{t+h})`). `prev_sample` should be used as next model input in the denoising loop.
-        pred_original_sample (`torch.Tensor` of shape `(batch_size, num_channels, height, width)` for images):
-            The predicted denoised sample `(x_{0})` (which in flow-matching notation should be noted as
-            `(x_{1})`) based on the model output from the current timestep.
-            `pred_original_sample` can be used to preview progress or for guidance.
-    """
-
-    prev_sample: torch.Tensor
-    pred_original_sample: Optional[torch.Tensor] = None
-
-
-def get_time_coefficients(timestep: torch.Tensor, ndim: int) -> torch.Tensor:
-    """
-    Convert timestep to time coefficients.
-    Args:
-        timestep (`torch.Tensor`): Timestep tensor.
-        ndim (`int`): Number of dimensions.
-    Returns:
-        `torch.Tensor`: Time coefficients.
-    """
-    return timestep.reshape((timestep.shape[0], *([1] * (ndim - 1) )))
-
-
-class FlowMatchingEulerScheduler(SchedulerMixin, ConfigMixin):
-    """
-    `FlowMatchingEulerScheduler` is a scheduler for training and inferencing Conditional Flow Matching models (CFMs).
-
-    Flow Matching (FM) is a novel, simulation-free methodology for training Continuous Normalizing Flows (CNFs) by
-    regressing vector fields of predetermined conditional probability paths, facilitating scalable training and
-    efficient sample generation through the utilization of various probability paths, including Gaussian and
-    Optimal Transport (OT) paths, thereby enhancing model performance and generalization capabilities
-
-    Args:
-        num_inference_steps (`int`, defaults to 100):
-            The number of steps on inference.
-    """
-
-    @register_to_config
-    def __init__(self, num_inference_steps: int = 100):
-        self.timesteps = None
-        self.num_inference_steps = None
-        self.h = None
-
-        if num_inference_steps is not None:
-            self.set_timesteps(num_inference_steps)
-
-    @staticmethod
-    def add_noise(original_samples: torch.Tensor, noise: torch.Tensor, timestep: torch.Tensor) -> torch.Tensor:
-        """
-        Add noise to the given sample
-
-        Args:
-            original_samples (`torch.Tensor`):
-                The original sample that is to be noised
-            noise (`torch.Tensor`):
-                The noise that is used to noise the image
-            timestep (`torch.Tensor`):
-                Timestep used to create linear interpolation `x_t = t * x_1 + (1 - t) * x_0`.
-                Where x_1 is a target distribution, x_0 is a source distribution and t (timestep) ∈ [0, 1]
-        """
-
-        t = get_time_coefficients(timestep, original_samples.ndim)
-
-        noised_sample = t * original_samples + (1 - t) * noise
-
-        return noised_sample
-
-    def set_timesteps(self, num_inference_steps: int = 100) -> None:
-        """
-        Set number of inference steps (Euler intagration steps)
-
-        Args:
-            num_inference_steps (`int`, defaults to 100):
-                The number of steps on inference.
-        """
-
-        self.num_inference_steps = num_inference_steps
-        self.h = 1 / num_inference_steps
-        self.timesteps = torch.arange(0, 1, self.h)
-
-    def step(self, model_output: torch.Tensor, timestep: torch.Tensor, sample: torch.Tensor,
-             return_dict: bool = True) -> Union[FlowMatchingEulerSchedulerOutput, Tuple]:
-        """
-        Predict the sample from the previous timestep by reversing the SDE. This function propagates the diffusion
-        process from the learned model outputs (most often the predicted noise).
-
-        Args:
-            model_output (`torch.Tensor`):
-                The direct output from learned diffusion model.
-            timestep (`float`):
-                Timestep used to perform Euler Method `x_t = h * f(x_t, t) + x_{t-1}`.
-                Where x_1 is a target distribution, x_0 is a source distribution and t (timestep) ∈ [0, 1]
-            sample (`torch.Tensor`):
-                A current instance of a sample created by the diffusion process.
-            return_dict (`bool`, *optional*, defaults to `True`):
-                Whether or not to return a [`~schedulers.scheduling_ddpm.DDPMSchedulerOutput`] or `tuple`.
-
-        Returns:
-            [`~schedulers.scheduling_ddpm.DDPMSchedulerOutput`] or `tuple`:
-                If return_dict is `True`, [`~schedulers.scheduling_ddpm.DDPMSchedulerOutput`] is returned, otherwise a
-                tuple is returned where the first element is the sample tensor.
-        """
-
-        step = FlowMatchingEulerSchedulerOutput(
-            prev_sample=sample + self.h * model_output,
-            pred_original_sample=sample + (1 - get_time_coefficients(timestep, model_output.ndim)) * model_output
-        )
-
-        if return_dict:
-            return step
-
-        return step.prev_sample,
-
-    @staticmethod
-    def get_velocity(original_samples: torch.Tensor, noise: torch.Tensor) -> torch.Tensor:
-        """
-        Predict the sample from the previous timestep by reversing the SDE. This function propagates the diffusion
-        process from the learned model outputs (most often the predicted noise).
-
-        Args:
-            original_samples (`torch.Tensor`):
-                The original sample that is to be noised
-            noise (`torch.Tensor`):
-                The noise that is used to noise the image
-
-        Returns:
-            `torch.Tensor`
-        """
-
-        return original_samples - noise
-
-    @staticmethod
-    def scale_model_input(sample: torch.Tensor, timestep: Optional[int] = None) -> torch.Tensor:
-        """
-         Ensures interchangeability with schedulers that need to scale the denoising model input depending on the
-         current timestep.
-
-        Args:
-            sample (`torch.Tensor`):
-                The input sample.
-            timestep (`int`, *optional*):
-                The current timestep in the diffusion chain.
-
-        Returns:
-            `torch.Tensor`:
-                A scaled input sample.
-        """
-
-        return sample
-
-
-
-
-
-# region scheduler end
-
-# --------------------------- Параметры ---------------------------
-save_path = "datasets/768" # "datasets/576" #"datasets/576p2" #"datasets/1152p2" #"datasets/576p2" #"datasets/dataset384_temp" #"datasets/dataset384" #"datasets/imagenet-1kk" #"datasets/siski576" #"datasets/siski384" #"datasets/siski64" #"datasets/mnist"
-batch_size = 30 #26 #45 #11 #45 #555 #35 #7
-base_learning_rate = 4e-6 #9.5e-7 #9e-7 #2e-6 #1e-6 #9e-7 #1e-6 #2e-6 #1e-6 #2e-6 #6e-6 #2e-6 #8e-7 #6e-6 #2e-5 #4e-5 #3e-5 #5e-5 #8e-5
-min_learning_rate = 2.5e-5 #2e-5
-num_epochs = 1 #2 #36 #18
-project = "sdxs"
-<<<<<<< HEAD
-use_wandb = True
-save_model = True
-=======
-use_wandb = False
-save_model = False
->>>>>>> d0c94e4 (sdxxxs)
-limit = 0 #200000 #0
-checkpoints_folder = ""
-
-# Параметры для диффузии
-n_diffusion_steps = 40 
-samples_to_generate = 12
-guidance_scale = 5
-sample_interval_share = 25 # samples/save per epoch
-
-# Папки для сохранения результатов
-generated_folder = "samples"
-os.makedirs(generated_folder, exist_ok=True)
-
-# Настройка seed для воспроизводимости
-current_date = datetime.now()
-seed = int(current_date.strftime("%Y%m%d"))
-fixed_seed = True
-if fixed_seed:
-    torch.manual_seed(seed)
-    np.random.seed(seed)
-    random.seed(seed)
-    if torch.cuda.is_available():
-        torch.cuda.manual_seed_all(seed)
-
-# --------------------------- Параметры LoRA ---------------------------
-# pip install peft
-lora_name = "" #"nusha"  # Имя для сохранения/загрузки LoRA адаптеров
-lora_rank = 32   # Ранг LoRA (чем меньше, тем компактнее модель)
-lora_alpha = 64  # Альфа параметр LoRA, определ��ющий масштаб
-
-print("init")
-# Включение Flash Attention 2/SDPA
-torch.backends.cuda.enable_flash_sdp(True)
-# --------------------------- Инициализация Accelerator --------------------
-dtype = torch.bfloat16
-accelerator = Accelerator(mixed_precision="bf16")
-device = accelerator.device
-gen = torch.Generator(device=device)
-gen.manual_seed(seed)
-
-# --------------------------- Инициализация WandB ---------------------------
-if use_wandb and accelerator.is_main_process:
-    wandb.init(project=project+lora_name, config={
-        "batch_size": batch_size,
-        "base_learning_rate": base_learning_rate,
-        "num_epochs": num_epochs,
-        "n_diffusion_steps": n_diffusion_steps,
-        "samples_to_generate": samples_to_generate,
-        "dtype": str(dtype)
-    })
-
-# --------------------------- Загрузка датасета ---------------------------
-class ResolutionBatchSampler(Sampler):
-    """Сэмплер, который группирует примеры по одинаковым размерам"""
-    def __init__(self, dataset, batch_size, shuffle=True, drop_last=False):
-        self.dataset = dataset
-        self.batch_size = batch_size
-        self.shuffle = shuffle
-        self.drop_last = drop_last
-        
-        # Группируем примеры по размерам
-        self.size_groups = defaultdict(list)
-
-        try:
-            widths = dataset["width"]
-            heights = dataset["height"]
-        except KeyError:
-            widths = [0] * len(dataset)
-            heights = [0] * len(dataset)
-            
-        for i, (w, h) in enumerate(zip(widths, heights)):
-            size = (w, h)
-            self.size_groups[size].append(i)
-        
-        # Печатаем статистику по размерам
-        print(f"Найдено {len(self.size_groups)} уникальных размеров:")
-        for size, indices in sorted(self.size_groups.items(), key=lambda x: len(x[1]), reverse=True):
-            width, height = size
-            print(f"  {width}x{height}: {len(indices)} примеров")
-        
-        # Формируем батчи
-        self.reset()
-        
-    def reset(self):
-        """Сбрасывает и перемешивает индексы"""
-        self.batches = []
-        
-        for size, indices in self.size_groups.items():
-            if self.shuffle:
-                indices_copy = indices.copy()
-                random.shuffle(indices_copy)
-            else:
-                indices_copy = indices
-            
-            # Разбиваем на батчи
-            for i in range(0, len(indices_copy), self.batch_size):
-                batch_indices = indices_copy[i:i + self.batch_size]
-                
-                # Пропускаем неполные батчи если drop_last=True
-                if self.drop_last and len(batch_indices) < self.batch_size:
-                    continue
-                    
-                self.batches.append(batch_indices)
-        
-        # Перемешиваем батчи между собой
-        if self.shuffle:
-            random.shuffle(self.batches)
-    
-    def __iter__(self):
-        self.reset()  # Сбрасываем и перемешиваем в начале каждой эпохи
-        return iter(self.batches)
-    
-    def __len__(self):
-        return len(self.batches)
-
-# Функция для выборки фиксированных семплов по размерам
-def get_fixed_samples_by_resolution(dataset, samples_per_group=1):
-    """Выбирает фиксированные семплы для каждого уникального разрешения"""
-    # Группируем по размерам
-    size_groups = defaultdict(list)
-    try:
-        widths = dataset["width"]
-        heights = dataset["height"]
-    except KeyError:
-        widths = [0] * len(dataset)
-        heights = [0] * len(dataset)
-    for i, (w, h) in enumerate(zip(widths, heights)):
-        size = (w, h)
-        size_groups[size].append(i)
-    
-    # Выбираем фиксированные примеры из каждой группы
-    fixed_samples = {}
-    for size, indices in size_groups.items():
-        # Определяем сколько семплов брать из этой группы
-        n_samples = min(samples_per_group, len(indices))
-        if len(size_groups)==1:
-            n_samples = samples_to_generate
-        if n_samples == 0:
-            continue
-            
-        # Выбираем случайные индексы
-        sample_indices = random.sample(indices, n_samples)
-        samples_data = [dataset[idx] for idx in sample_indices]
-        
-        # Собираем данные
-        latents = torch.tensor(np.array([item["vae"] for item in samples_data]), dtype=dtype).to(device)
-        embeddings = torch.tensor(np.array([item["embeddings"] for item in samples_data]), dtype=dtype).to(device)
-        texts = [item["text"] for item in samples_data]
-        
-        # Сохраняем для этого размера
-        fixed_samples[size] = (latents, embeddings, texts)
-    
-    print(f"Создано {len(fixed_samples)} групп фиксированных семплов по разрешениям")
-    return fixed_samples
-
-if limit > 0:
-    dataset = load_from_disk(save_path).select(range(limit))
-else:
-    dataset = load_from_disk(save_path)
-
-
-
-def collate_fn(batch):
-    # Преобразуем список в тензоры и перемещаем на девайс
-    latents = torch.tensor(np.array([item["vae"] for item in batch]), dtype=dtype).to(device)
-    embeddings = torch.tensor(np.array([item["embeddings"] for item in batch]), dtype=dtype).to(device)
-    return latents, embeddings
-    
-# Используем наш ResolutionBatchSampler
-batch_sampler = ResolutionBatchSampler(dataset, batch_size=batch_size, shuffle=True)
-dataloader = DataLoader(dataset, batch_sampler=batch_sampler)#, collate_fn=collate_fn)
-
-print("Total samples",len(dataloader))
-dataloader = accelerator.prepare(dataloader)
-
-# --------------------------- Загрузка моделей ---------------------------
-# VAE загружается на CPU для экономии GPU-памяти
-vae = AutoencoderKL.from_pretrained("AuraDiffusion/16ch-vae").to("cpu", dtype=dtype)
-
-# DDPMScheduler с V_Prediction и Zero-SNR
-# scheduler = DDPMScheduler(
-#     num_train_timesteps=1000,       # Полный график шагов для обучения
-#     prediction_type="v_prediction", # V-Prediction
-#     rescale_betas_zero_snr=True,    # Включение Zero-SNR
-#     timestep_spacing="leading",     # Добавляем улучшенное распределение шагов
-#     steps_offset=1                  # Избегаем проблем с нулевым timestep
-# )
-
-# Flow Matching
-scheduler = FlowMatchingEulerScheduler(
-<<<<<<< HEAD
-    num_train_timesteps=1000,
-=======
-   # num_train_timesteps=1000,
->>>>>>> d0c94e4 (sdxxxs)
-)
-
-# Инициализация переменных для возобновления обучения
-start_epoch = 0
-global_step = 0
-
-# Расчёт общего количества шагов 
-total_training_steps = (len(dataloader) * num_epochs)
-# Get the world size
-world_size = accelerator.state.num_processes
-print(f"World Size: {world_size}")
-
-# Опция загрузки модели из последнего чекпоинта (если существует)
-latest_checkpoint = os.path.join(checkpoints_folder, project)
-if os.path.isdir(latest_checkpoint):
-    print("Загружаем UNet из чекпоинта:", latest_checkpoint)
-    unet = UNet2DConditionModel.from_pretrained(latest_checkpoint).to(device, dtype=dtype)
-    unet.enable_gradient_checkpointing()
-    unet.set_use_memory_efficient_attention_xformers(False) # отключаем xformers
-    try:
-        unet.set_attn_processor(AttnProcessor2_0())  # Используем стандартный AttnProcessor
-        print("SDPA включен через set_attn_processor.")
-    except Exception as e:
-        print(f"Ошибка при включении SDPA: {e}")
-        print("Попытка использовать enable_xformers_memory_efficient_attention.")
-        unet.set_use_memory_efficient_attention_xformers(True)
-
-if lora_name:
-    print(f"--- Настройка LoRA через PEFT (Rank={lora_rank}, Alpha={lora_alpha}) ---")
-    from peft import LoraConfig, get_peft_model, prepare_model_for_kbit_training
-    from peft.tuners.lora import LoraModel
-    import os
-    # 1. Замораживаем все параметры UNet
-    unet.requires_grad_(False)
-    print("Параметры базового UNet заморожены.")
-
-    # 2. Создаем конфигурацию LoRA
-    lora_config = LoraConfig(
-        r=lora_rank,
-        lora_alpha=lora_alpha,
-        target_modules=["to_q", "to_k", "to_v", "to_out.0"],
-    )
-    unet.add_adapter(lora_config)
-
-    # 3. Оборачиваем UNet в PEFT-модель
-    from peft import get_peft_model
-    
-    peft_unet = get_peft_model(unet, lora_config)
-
-    # 4. Получаем параметры для оптимизации
-    params_to_optimize = list(p for p in peft_unet.parameters() if p.requires_grad)
-    
-
-    # 5. Выводим информацию о количестве параметров
-    if accelerator.is_main_process:
-        lora_params_count = sum(p.numel() for p in params_to_optimize)
-        total_params_count = sum(p.numel() for p in unet.parameters())
-        print(f"Количество обучаемых параметров (LoRA): {lora_params_count:,}")
-        print(f"Общее количество параметров UNet: {total_params_count:,}")
-
-    # 6. Путь для сохранения
-    lora_save_path = os.path.join("lora", lora_name)
-    os.makedirs(lora_save_path, exist_ok=True) 
-
-    # 7. Функция для сохранения
-    def save_lora_checkpoint(model):
-        if accelerator.is_main_process:
-            print(f"Сохраняем LoRA адаптеры  в {lora_save_path}")
-            from peft.utils.save_and_load import get_peft_model_state_dict
-            # Получаем state_dict только LoRA
-            lora_state_dict = get_peft_model_state_dict(model)
-
-            # Сохраняем веса
-            torch.save(lora_state_dict, os.path.join(lora_save_path, "adapter_model.bin"))
-    
-            # Сохраняем конфиг
-            model.peft_config["default"].save_pretrained(lora_save_path)
-            # SDXL must be compatible
-            from diffusers import StableDiffusionXLPipeline
-            StableDiffusionXLPipeline.save_lora_weights(lora_save_path, lora_state_dict)
-
-# --------------------------- Оптимизатор ---------------------------
-# Определяем параметры для оптимизации
-if lora_name:
-    # Если используется LoRA, оптимизируем только параметры LoRA
-    trainable_params = [p for p in unet.parameters() if p.requires_grad]
-else:
-    # Иначе оптимизируем все параметры
-    trainable_params = list(unet.parameters())
-    
-# [1] Создаем словарь оптимизаторов (fused backward)
-optimizer_dict = {
-    p: bnb.optim.AdamW8bit(
-        [p],  # Каждый параметр получает свой оптимизатор
-        lr=base_learning_rate,
-        betas=(0.9, 0.999),
-        weight_decay=1e-5,
-        eps=1e-8
-    ) for p in trainable_params
-}
-
-# [2] Определяем hook для применения оптимизатора сразу после накопления градиента
-def optimizer_hook(param):
-    optimizer_dict[param].step()
-    optimizer_dict[param].zero_grad(set_to_none=True)
-
-# [3] Регистрируем hook для trainable параметров модели
-for param in trainable_params:
-    param.register_post_accumulate_grad_hook(optimizer_hook)
-
-# Подготовка через Accelerator
-unet, optimizer = accelerator.prepare(unet, optimizer_dict)
-
-# --------------------------- Фиксированные семплы для генерации ---------------------------
-# Примеры фиксированных семплов по размерам
-fixed_samples = get_fixed_samples_by_resolution(dataset)
-
-
-@torch.no_grad()
-def generate_and_save_samples(fixed_samples,step):
-    """
-    Генерирует семплы для каждого из разрешений и сохраняет их.
-    
-    Args:
-        step: Текущий шаг обучения
-        fixed_samples: Словарь, где ключи - размеры (width, height), 
-                       а значения - кортежи (latents, embeddings)
-    """
-    try:
-        original_model = accelerator.unwrap_model(unet)
-        # Перемещаем VAE на device для семплирования
-        vae.to(accelerator.device, dtype=dtype)
-        
-        # Устанавливаем количество diffusion шагов
-        scheduler.set_timesteps(n_diffusion_steps)
-        
-        all_generated_images = []
-        size_info = []  # Для хранения информации о размере для каждого изображения
-        all_captions = [] 
-        
-        # Проходим по всем группам размеров
-        for size, (sample_latents, sample_text_embeddings, sample_text) in fixed_samples.items():
-            width, height = size
-            size_info.append(f"{width}x{height}")
-            #print(f"Генерация {sample_latents.shape[0]} изображений размером {width}x{height}")
-            
-            # Инициализируем латенты случайным шумом для этой группы
-            noise = torch.randn(
-                sample_latents.shape,
-                generator=gen,
-                device=sample_latents.device,
-                dtype=sample_latents.dtype
-            )
-            
-            # Начинаем с шума
-            current_latents = noise.clone()
-            
-            # Подготовка текстовых эмбеддингов для guidance
-            if guidance_scale > 0:
-                empty_embeddings = torch.zeros_like(sample_text_embeddings)
-                text_embeddings = torch.cat([empty_embeddings, sample_text_embeddings], dim=0)
-            else:
-                text_embeddings = sample_text_embeddings
-            
-            # Генерация изображений
-            for t in scheduler.timesteps:
-                # Подготовка входных данных для UNet
-                t =  t.unsqueeze(dim=0).to(device) # Добавляем размерность для батча
-                if guidance_scale > 0:
-                    latent_model_input = torch.cat([current_latents] * 2)
-                    latent_model_input = scheduler.scale_model_input(latent_model_input, t)
-                else:
-                    latent_model_input = scheduler.scale_model_input(current_latents, t)
-                
-                # Предсказание шума
-                noise_pred = original_model(latent_model_input, t, text_embeddings).sample
-                
-                # Применение guidance scale
-                if guidance_scale > 0:
-                    noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
-                    noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
-                
-                # Обновление латентов
-                current_latents = scheduler.step(noise_pred, t, current_latents).prev_sample
-            
-            # Декодирование через VAE
-            latent = (current_latents.detach() / vae.config.scaling_factor) + vae.config.shift_factor
-            latent = latent.to(accelerator.device, dtype=dtype)
-            decoded = vae.decode(latent).sample
-            
-            # Преобразуем тензоры в PIL-изображения и сохраняем
-            for img_idx, img_tensor in enumerate(decoded):
-                img = (img_tensor.to(torch.float32) / 2 + 0.5).clamp(0, 1).cpu().numpy().transpose(1, 2, 0)
-                pil_img = Image.fromarray((img * 255).astype("uint8"))
-                # Определяем максимальные ширину и высоту
-                max_width = max(size[0] for size in fixed_samples.keys())
-                max_height = max(size[1] for size in fixed_samples.keys())
-                max_width = max(255,max_width)
-                max_height = max(255,max_height)
-            
-                # Добавляем padding, чтобы изображение стало размером max_width x max_height
-                padded_img = ImageOps.pad(pil_img, (max_width, max_height), color='white')
-            
-                all_generated_images.append(padded_img)
-
-                caption_text = sample_text[img_idx][:200] if img_idx < len(sample_text) else ""
-                all_captions.append(caption_text)
-                
-                # Сохраняем с информацией о размере в имени файла
-                save_path = f"{generated_folder}/{project}_{width}x{height}_{img_idx}.jpg"
-                pil_img.save(save_path, "JPEG", quality=96)
-        
-        # Отправляем изображения на WandB с информацией о размере
-        if use_wandb and accelerator.is_main_process:
-            wandb_images = [
-                wandb.Image(img, caption=f"{all_captions[i]}")
-                for i, img in enumerate(all_generated_images)
-            ]
-            wandb.log({"generated_images": wandb_images, "global_step": step})
-    
-    finally:        
-        # Гарантированное перемещение VAE обратно на CPU
-        vae.to("cpu")
-        if original_model is not None:
-            del original_model
-        # Очистка всех тензоров
-        for var in list(locals().keys()):
-            if isinstance(locals()[var], torch.Tensor):
-                del locals()[var]
-        torch.cuda.empty_cache()
-        gc.collect()
-        
-# --------------------------- Генерация сэмплов перед обучением ---------------------------
-if accelerator.is_main_process:
-    if save_model:
-        print("Генерация сэмплов до старта обучения...")
-        generate_and_save_samples(fixed_samples,0)
-
-# Модифицируем функцию сохранения модели для поддержки LoRA
-def save_checkpoint(unet):
-    if accelerator.is_main_process:
-        if lora_name:
-            # Сохраняем только LoRA адаптеры
-            save_lora_checkpoint(unet)
-        else:
-            # Сохраняем полную модель
-            accelerator.unwrap_model(unet).save_pretrained(os.path.join(checkpoints_folder, f"{project}"))
-
-# --------------------------- Тренировочный цикл ---------------------------
-# Для логирования среднего лосса каждые % эпохи
-if accelerator.is_main_process:
-    print(f"Total steps per GPU: {total_training_steps}")
-print(f"[GPU {accelerator.process_index}] Total steps: {total_training_steps}")
-
-epoch_loss_points = []
-progress_bar = tqdm(total=total_training_steps, disable=not accelerator.is_local_main_process, desc="Training", unit="step")
-
-# Определяем интервал для сэмплирования и логирования в пределах эпохи (10% эпохи)
-steps_per_epoch = len(dataloader)
-sample_interval = max(1, steps_per_epoch // sample_interval_share)
-
-# Начинаем с указанной эпохи (полезно при возобновлении)
-for epoch in range(start_epoch, start_epoch + num_epochs):
-    batch_losses = []
-    unet.train()
-    
-    for step, (latents, embeddings) in enumerate(dataloader):
-        with accelerator.accumulate(unet):
-            if save_model == False and step == 3 :
-                used_gb = torch.cuda.max_memory_allocated() / 1024**3
-                print(f"Шаг {step}: {used_gb:.2f} GB")
-            # Forward pass 
-            noise = torch.randn_like(latents)
-            
-            timesteps = torch.randint(
-                0,
-                1000, 
-                (latents.shape[0],), 
-                device=device
-            ) / 1000 # Кастим в float
-            
-            # Добавляем шум к латентам
-            noisy_latents = scheduler.add_noise(latents, noise, timesteps)
-
-            # Получаем предсказание шума
-            noise_pred = unet(noisy_latents, timesteps, embeddings).sample #.to(dtype=torch.bfloat16)
-        
-            # Используем целевое значение v_prediction
-            target = scheduler.get_velocity(latents, noise)
-
-            # Считаем лосс
-            loss = torch.nn.functional.mse_loss(noise_pred.float(), target.float())
-
-            # Делаем backward через Accelerator
-            accelerator.backward(loss)
-            
-            # Увеличиваем счетчик глобальных шагов
-            global_step += 1
-            
-            # Обновляем прогресс-бар
-            progress_bar.update(1)
-            
-            # Логируем метрики
-            if accelerator.is_main_process:
-                current_lr = base_learning_rate
-                batch_losses.append(loss.detach().item())
-                
-                # Логируем в Wandb 
-                if use_wandb:
-                    wandb.log({
-                        "loss": loss.detach().item(),
-                        "learning_rate": current_lr,
-                        "epoch": epoch,
-                        "global_step": global_step
-                    })
-            
-                # Генерируем сэмплы с заданным интервалом
-                if global_step % sample_interval == 0:
-                    if save_model:
-                        save_checkpoint(unet)
-                        
-                    generate_and_save_samples(fixed_samples,global_step)
-                    
-                    # Выводим текущий лосс
-                    avg_loss = np.mean(batch_losses[-sample_interval:])
-                    #print(f"Эпоха {epoch}, шаг {global_step}, средний лосс: {avg_loss:.6f}, LR: {current_lr:.8f}")
-                    if use_wandb:
-                        wandb.log({"intermediate_loss": avg_loss})
-            
-    
-    # По окончании эпохи
-    if accelerator.is_main_process:
-        avg_epoch_loss = np.mean(batch_losses)
-        print(f"\nЭпоха {epoch} завершена. Средний лосс: {avg_epoch_loss:.6f}")
-        if use_wandb:
-            wandb.log({"epoch_loss": avg_epoch_loss, "epoch": epoch+1})
-
-# Завершение обучения - сохраняем финальную модель
-if accelerator.is_main_process:
-    print("Обучение завершено! Сохраняем финальную модель...")
-    # Сохраняем основную модель
-    #if save_model:
-    save_checkpoint(unet)
-    print("Готово!")

unet/config.json

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:18891ea8f81c705422b12d7575493f5abb949d60f2814a989a436fde65e84c82
-size 1873
+oid sha256:636ed60d2a73b31fe88849234ff578446cd018092fdc9210575893551220bfc3
+size 1887

unet/diffusion_pytorch_model.fp16.safetensors

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:756e18a3addbf8b5b8d1fabc9bae9e03d710318eb9698346ba3ed70b4e07af72
-size 3929714960

{micro → unet}/diffusion_pytorch_model.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:756e18a3addbf8b5b8d1fabc9bae9e03d710318eb9698346ba3ed70b4e07af72
-size 3929714960
+oid sha256:b8ec1c74913276f691c838b5a85cd65939c4437c0ce1741f69c8c8931dd39112
+size 3092571208

vae/config-Copy1.1json

@@ -1,37 +0,0 @@
-{
-  "_class_name": "AutoencoderKL",
-  "_diffusers_version": "0.30.0.dev0",
-  "act_fn": "silu",
-  "block_out_channels": [
-    128,
-    256,
-    512,
-    512
-  ],
-  "down_block_types": [
-    "DownEncoderBlock2D",
-    "DownEncoderBlock2D",
-    "DownEncoderBlock2D",
-    "DownEncoderBlock2D"
-  ],
-  "force_upcast": false,
-  "in_channels": 3,
-  "latent_channels": 16,
-  "latents_mean": [0.2539, 0.1431, 0.1484, -0.3048, -0.0985, -0.162, 0.1403, 0.2034, -0.1419, 0.2646, 0.0655, 0.0061, 0.1555, 0.0506, 0.0129, -0.1948],
-  "latents_std": [0.8123, 0.7376, 0.7354, 1.1827, 0.8387, 0.8735, 0.8705, 0.8142, 0.8076, 0.7409, 0.7655, 0.8731, 0.8087, 0.7058, 0.8087, 0.7615],
-  "layers_per_block": 2,
-  "mid_block_add_attention": false,
-  "norm_num_groups": 32,
-  "out_channels": 3,
-  "sample_size": 1024,
-  "scaling_factor": 1,
-  "shift_factor": 0,
-  "up_block_types": [
-    "UpDecoderBlock2D",
-    "UpDecoderBlock2D",
-    "UpDecoderBlock2D",
-    "UpDecoderBlock2D"
-  ],
-  "use_post_quant_conv": true,
-  "use_quant_conv": true
-}

vae/config.json

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:45e70f6e76d6094b0a93bc35630c31e0efe2617879e1aff5395e75ca5cd8bf7a
-size 718
+oid sha256:ca78c49c02cf8afd8f022b99418bec3e4da9882cfed036c3664897af6d4589f0
+size 801

vae/diffusion_pytorch_model.fp16.safetensors

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:a27d33cadf456598eee3cf5638ba3e8e01dd020a5f9c80c1a7b162bf97096701
-size 163460798

result_grid.jpg → vae/diffusion_pytorch_model.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:b9c9f658508ed3b80da550599f8b23699dfefed03a67f6908bb525f59b1d6b26
-size 7023173
+oid sha256:2ff1aa1cce9dcf003c7eb23548fdb62101fe9bdeedf30aa4a897b91a6ef13f93
+size 167669678