2604

README.md

@@ -11,7 +11,7 @@ datasets:
 
 *XS Size, Excess Quality (not for now)* 
 
-Train status: 1xRTX5090 / we need more gold / [support us please..](https://huggingface.co/AiArtLab/sdxs-1b#donations)
+Train status: 1xRTX4090 / we need more gold / [support us please..](https://huggingface.co/AiArtLab/sdxs-1b#donations)
 
 <img src="media/gandr.jpg" height="256"/>
 
@@ -325,7 +325,7 @@ cd sdxs-1b
 pip install -r requirements.txt -U
 mkdir datasets
 cd datasets
-hf download babkasotona/ds1234_noanime_704_vae8x16x  --repo-type dataset --local-dir ds1234_noanime_704_vae8x16x
+#hf download babkasotona/ds1234_noanime_704_vae8x16x  --repo-type dataset --local-dir ds1234_noanime_704_vae8x16x
 cd ..
 nohup accelerate launch train.py &
 ```
@@ -337,9 +337,9 @@ nohup accelerate launch train.py &
 - **[CaptionEmporium](https://huggingface.co/CaptionEmporium)**
 
 ## Donations
-Thanks for your support! Every $5 equals a full day of RTX 5090 rental and makes the waifu 0.01% better!
+Thanks for your support! Every $5 equals a full day of RTX 4090 rental and makes the waifu 0.01% better!
 
- - Euro / dollars / other: Sorry, we can’t accept payments via Patreon or Ko-fi. Please register on [vast.ai](https://cloud.vast.ai/billing/) (this is the GPU provider we use for training the model), top up any amount, and send: Transfer Money / User / Email: vadim-kulibaba@yandex.ru.
+ - Euro / dollars: sorry, we can’t accept payments via Patreon or Ko-fi. Please register on [vast.ai](https://cloud.vast.ai/billing/) (this is the GPU provider we use for training the model), top up any amount, and transfer: Transfer Money / User / Email: vadim-kulibaba@yandex.ru.
  - RUB: [donate in rub](https://www.tbank.ru/cf/90ensBQqpJj)
  - DOGE: DEw2DR8C7BnF8GgcrfTzUjSnGkuMeJhg83
  - BTC: 3JHv9Hb8kEW8zMAccdgCdZGfrHeMhH1rpN

pipeline_sdxs.py

@@ -14,7 +14,7 @@ class SdxsPipelineOutput(BaseOutput):
     prompt: Optional[Union[str, List[str]]] = None
 
 class SdxsPipeline(DiffusionPipeline):
-    MAX_TEXT_TOKENS = 248
+    MAX_TEXT_TOKENS = 250
     
     def __init__(self, vae, text_encoder, processor, tokenizer, unet, scheduler):
         super().__init__()
@@ -154,9 +154,8 @@ class SdxsPipeline(DiffusionPipeline):
         final_mask = torch.cat([torch.ones((batch_size, 1), device=device, dtype=inputs.attention_mask.dtype), inputs.attention_mask], dim=1)
 
         return final_embeddings.to(dtype=dtype), final_mask.to(dtype=torch.int64)
+
     @torch.no_grad()
-    
-    
     def encode_text_and_image_naive(self, text: Union[str, List[str]], image: Optional[Union[Image.Image, List[Image.Image], str, List[str]]] = None, scale = 0.5) -> Tuple[torch.Tensor, torch.Tensor]:
         # 1. Получаем текстовый эмбеддинг
         text_embeds, text_mask = self.encode_text(text)
@@ -276,7 +275,7 @@ class SdxsPipeline(DiffusionPipeline):
         height: int = 1408,
         width: int = 1024,
         num_inference_steps: int = 40,
-        guidance_scale: float = 5.0,
+        guidance_scale: float = 4.0,
         generator: Optional[torch.Generator] = None,
         seed: Optional[int] = None,
         output_type: str = "pil",

test.ipynb

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:f2ef4cc2b7ae0fb40a388690c323f3921b448465471d00bdfeb71ce50ca01961
-size 17615512
+oid sha256:5f490d6558ed482b844dcac5723d174fe4bd5a13f4767641cff9c306ee05f385
+size 17750930

train.py

@@ -32,7 +32,7 @@ os.environ["NCCL_IB_DISABLE"] = "1" # comment this on H100!
 os.environ["PYTORCH_CUDA_ALLOC_CONF"] = "expandable_segments:True"
 
 # --------------------------- Параметры ---------------------------
-ds_path = "datasets/ds1234_noanime_704_vae8x16x" #alchemist_704_vae8x16x_imgpool"
+ds_path = "datasets/alchemist_704_vae8x16x_imgpool" #ds1234_noanime_704_vae8x16x" #"
 project = "unet"
 
 gpu_mem_gb = torch.cuda.get_device_properties(0).total_memory / 1e9
@@ -62,7 +62,7 @@ use_wandb = True
 use_comet_ml = False
 save_model = True
 use_decay = True
-fbp = False
+fbp = True
 torch_compile = False
 unet_gradient = True
 loss_normalize = False
@@ -83,13 +83,13 @@ torch.backends.cuda.enable_math_sdp(False) # Отключаем медленны
 save_barrier = 1.25
 warmup_percent = 0.0025
 betta2 = 0.997
-eps = 1e-7
+eps = 1e-6
 clip_grad_norm = 1.0
 limit = 0
 checkpoints_folder = ""
 gradient_accumulation_steps = 1
 dtype = torch.float32
-mixed_precision = "no"
+mixed_precision = "bf16"
 
 # Параметры для диффузии
 n_diffusion_steps = 40
@@ -461,11 +461,11 @@ 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=0.01
+            params, lr=base_learning_rate, betas=(0.9, betta2), eps=eps, weight_decay=0.001
         )
     elif name == "adam":
         return torch.optim.AdamW(
-            params, lr=base_learning_rate, betas=(0.9, betta2), eps=eps, weight_decay=0.01
+            params, lr=base_learning_rate, betas=(0.9, betta2), eps=eps, weight_decay=0.001
         )
     elif name == "muon_adam8bit":
         return MuonAdamW8bit(
@@ -799,6 +799,8 @@ for epoch in range(start_epoch, start_epoch + num_epochs):
     
                 # 3. Время t, bias = -0.5 (Фокус на Деталях ~300) bias = 0.5 (Фокус на структуре) bias = 0 (колокол/ равномерно)
                 bias = 0.2
+                if lvl>3:
+                    bias = -0.2
                 t = torch.sigmoid(torch.randn(latents.shape[0], device=latents.device, dtype=latents.dtype) + bias)
                 
                 # интерполяция между x0 и шумом

unet/diffusion_pytorch_model.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:998fdcb3b5d267ed5ed987ee9649935ca646ca85b8f81e153fb395a14f575d5b
+oid sha256:656fb4709a2c0a80f31787fd508c946b569a63d4b700aa8817bdc1e77ac4b163
 size 3210307232