Update to guidance 1.5 (better samples)

README.md

@@ -23,7 +23,7 @@ from a caption using **classifier-free guidance**.
 ![samples](sample.png)
 
 *One row per class (t-shirt, trouser, pullover, dress, coat, sandal, shirt, sneaker, bag, ankle
-boot); each image is generated from pure noise, DDIM 50 steps, guidance 3.0.*
+boot); each image is generated from pure noise, DDIM 50 steps, guidance 1.5.*
 
 ## What it does
 
@@ -83,7 +83,8 @@ save_image(denormalize(imgs), "out.png", nrow=len(FASHION_CLASSES))
 - **Conditioning:** frozen CLIP (`openai/clip-vit-base-patch32`) text embeddings via FiLM +
   cross-attention; 15% caption dropout for classifier-free guidance
 - **Trained:** 15 epochs on Apple Silicon (MPS), final loss ≈ 0.043
-- **Sampling:** DDIM, 50 steps, guidance scale 3.0
+- **Sampling:** DDIM, 50 steps, guidance scale 1.5 (low guidance — high CFG artifacts on this
+  small, short-trained model)
 
 > These are the **raw** weights, not EMA — for a short (15-epoch) run the EMA average still lags
 > the live weights, so the raw model produces the cleaner samples.