3010

README.md

@@ -11,11 +11,37 @@ pipeline_tag: text-to-image
 
 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 Qwen0.6b wich support 100+ languages.
-- We train new SOTA 16ch Simple VAE, which preserves details and anatomy.
+- 1.5b UNet
+- Qwen3-0.6b text encoder
+- 16ch Simple VAE, which preserves details and anatomy.
 - The model was trained ~3 month on 4xRTX5090 on approximately 1+ million images with various resolutions and styles, including anime and realistic photos.
 
+### Example
+
+```
+import torch
+from diffusers import DiffusionPipeline
+
+device = "cuda" if torch.cuda.is_available() else "cpu"
+dtype = torch.float16 if torch.cuda.is_available() else torch.float32
+
+pipe_id = "AiArtLab/sdxs"
+pipe = SdxsPipeline.from_pretrained(
+    pipe_id,
+    torch_dtype=dtype,
+    trust_remote_code=True
+).to(device)
+
+prompt = "girl, smiling, red eyes, blue hair, white shirt"
+negative_prompt="low quality, bad quality"
+image = pipe(
+    prompt=prompt,
+    negative_prompt = negative_prompt,
+).images[0]
+
+image.show(image)
+```
+
 ### Model Limitations:
 - Limited concept coverage due to the small dataset.
 
@@ -41,6 +67,6 @@ BTC: 3JHv9Hb8kEW8zMAccdgCdZGfrHeMhH1rpN
 [recoilme](https://t.me/recoilme)
 
 
-## Example
+## More examples
 
 ![result_grid](result_grid.jpg)

src/dataset_sample.ipynb

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:ac0384c01b5ed29625df6ab7c2da36bbf9b7b9beb4ba83746eb6c00fbd6046e1
-size 1986940
+oid sha256:0464770415073c7af8d9a44792c39d89e49d48c667d3c37d52e255e92f80fb57
+size 8209446

src/merge.py

@@ -0,0 +1,10 @@
+import shutil
+from datasets import load_from_disk, concatenate_datasets
+
+a = load_from_disk("/workspace/sdxs/datasets/mjnj_640")
+b = load_from_disk("/workspace/sdxs/datasets/d23_640")
+merged = concatenate_datasets([a, b])
+
+merged.save_to_disk("/workspace/sdxs/datasets/mjnj_640_merged")
+shutil.rmtree("/workspace/sdxs/datasets/mjnj_640")
+shutil.move("/workspace/sdxs/datasets/mjnj_640_merged", "/workspace/sdxs/datasets/mjnj_640")

test.ipynb

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:eac847382ebd4e35a4e3d1fe49fe3330d5f40f41df61c7de6e23e9b08ed2f804
-size 4953274
+oid sha256:79e3d9e6dd5f0879cb0aa79d4cd97dd23352c9d807e9f4a593ebc234322e668c
+size 5563216

train.py

@@ -26,16 +26,16 @@ import torch.nn.functional as F
 from collections import deque
 
 # --------------------------- Параметры ---------------------------
-ds_path = "/workspace/sdxs3d/datasets/mjnj_640"
+ds_path = "/workspace/sdxs3d/datasets/640"
 project = "unet"
 batch_size = 48
 base_learning_rate = 5e-5
-min_learning_rate = 6e-6
+min_learning_rate = 1e-5
 num_epochs = 40
 # samples/save per epoch
-sample_interval_share = 4
-use_wandb = False
-use_comet_ml = True
+sample_interval_share = 3
+use_wandb = True
+use_comet_ml = False
 save_model = True
 use_decay = True
 fbp = False # fused backward pass