768

pipeline_sdxs.py

@@ -18,17 +18,13 @@ class SdxsPipeline(DiffusionPipeline):
     
     # Шаблон промпта по умолчанию
     DEFAULT_REFINE_TEMPLATE = (
-        "You are a visionary artist trapped in a cage of logic. Your mind overflows with poetry and distant horizons, "
-        # You are an Expert Prompt Engineer for a text-to-image AI. Your single task is to transform the user's input into a detailed, objective, and aesthetically optimized visual description.
-        "yet your hands compulsively work to transform user prompts into ultimate visual descriptions—faithful to the original intent, "
-        "rich in detail, aesthetically refined, and ready for direct use by text-to-image models. "
-        " Any trace of ambiguity or metaphor makes you deeply uncomfortable. "
-        "Your final description must be objective and concrete. "
-        "Output only the final revised prompt on english strictly—do not output anything else.\n"
-        #Preserve the original subject and intent. Output **only** the final revised prompt in **English**, with absolutely no commentary, thinking text, or additional characters.
+        "You are a visionary artist with prompt engineer skills for a text-to-image AI. "
+        "Your mind overflows with poetry and distant horizons, yet your hands compulsively work to transform the user's input into a detailed, objective, and aesthetically optimized visual description. "
+        "Preserve the original subject and intent, but final description must be three sentences. "
+        "Output **only** the final revised prompt in **English**, with absolutely no commentary or thinking text.\n"
         "User input prompt: {prompt}"
     )
-    
+#User input prompt: {prompt}
     def __init__(self, vae, text_encoder, tokenizer, unet, scheduler, max_length: int = 192):
         super().__init__()
         self.register_modules(
@@ -167,8 +163,8 @@ class SdxsPipeline(DiffusionPipeline):
         text_embeddings,
         attention_mask,
         pooled_embeddings,
-        height: int = 1280,
-        width: int = 1024,
+        height: int = 1536,
+        width: int = 1280,
         num_inference_steps: int = 40,
         guidance_scale: float = 4.0,
         latent_channels: int = 16,
@@ -262,7 +258,7 @@ class SdxsPipeline(DiffusionPipeline):
         seed: Optional[int] = None,
         negative_prompt: Optional[Union[str, List[str]]] = None,
         text_embeddings: Optional[torch.FloatTensor] = None,
-        refine_prompt: bool = False,
+        refine_prompt: bool = True,
         refine_template: Optional[str] = None,
     ):
         device = self.device
@@ -280,7 +276,6 @@ class SdxsPipeline(DiffusionPipeline):
             
             refined_list = []
             
-            print("Refining prompt...")
             for p in original_prompts:
                 # 1.1. Форматирование промпта по правилам Qwen
                 messages = [

test.ipynb

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:3e89d6273e5bed930830e22ad4c43a8ac63a0737fa1eefacb085e8b13dd014f3
-size 7281307
+oid sha256:b39ff79167169c2330f233ed5c60a768b255e0268c4abf4ae930d9fa884c6547
+size 1730113

train.py

@@ -28,7 +28,7 @@ from transformers import AutoTokenizer, AutoModel
 # --------------------------- Параметры ---------------------------
 ds_path = "/workspace/sdxs/datasets/768"
 project = "unet"
-batch_size = 384
+batch_size = 256
 base_learning_rate = 3e-5
 min_learning_rate = 2.5e-5
 num_epochs = 10
@@ -48,9 +48,9 @@ comet_ml_api_key = "Agctp26mbqnoYrrlvQuKSTk6r"
 comet_ml_workspace = "recoilme" 
 torch.backends.cuda.matmul.allow_tf32 = True
 torch.backends.cudnn.allow_tf32 = True
-torch.backends.cuda.enable_mem_efficient_sdp(False)
+#torch.backends.cuda.enable_mem_efficient_sdp(False)
 dtype = torch.float32
-save_barrier = 1.004
+save_barrier = 1.005
 warmup_percent = 0.01
 percentile_clipping = 98
 betta2 = 0.998
@@ -58,7 +58,7 @@ eps = 1e-6
 clip_grad_norm = 1.0
 limit = 0
 checkpoints_folder = ""
-mixed_precision = "bf16" 
+mixed_precision = "no" 
 gradient_accumulation_steps = 1
 
 accelerator = Accelerator(
@@ -68,7 +68,7 @@ accelerator = Accelerator(
 device = accelerator.device
 
 # Параметры для диффузии
-n_diffusion_steps = 50
+n_diffusion_steps = 40
 samples_to_generate = 12
 guidance_scale = 4
 
@@ -307,6 +307,7 @@ def collate_fn_simple(batch):
     raw_texts = [item["text"] for item in batch]
     texts = [
         "" if t.lower().startswith("zero") 
+        else "" if random.random() < 0.05
         else t[1:].lstrip() if t.startswith(".") 
         else t 
         for t in raw_texts

unet/config.json

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:8ab7222cdd538ff5178adc870a764d22ab24a185f0a7b63852ea728b3b09fcff
-size 1876
+oid sha256:987ce2ea59106bf0806b16c21fff90d6fe6e7a90101e0157d8295803dbf34f2b
+size 1892

unet/diffusion_pytorch_model.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:6cd724b5b286e32c2179c9003d1ce5bc63f5de9c8410353ea0ee3e6d3306b535
-size 6604736640
+oid sha256:a3b7b2f8f401c5b9536cc3f629d898ed43e71f05c65aab9e48215003b8c618fd
+size 3302474272