Fix: Remove global gamma, bake BiRefNet model, optimize Docker for startup

Dockerfile

@@ -62,8 +62,11 @@ RUN wget -P assets/example_image https://raw.githubusercontent.com/trellis3d/tre
     && wget -P assets/example_image https://raw.githubusercontent.com/trellis3d/trellis/main/assets/example_image/cherries_1.png \
     && wget -P assets/example_image https://raw.githubusercontent.com/trellis3d/trellis/main/assets/example_image/dog_1.png \
     && wget -P assets/example_image https://raw.githubusercontent.com/trellis3d/trellis/main/assets/example_image/hotdog_1.png \
-    && wget -P assets/example_image https://raw.githubusercontent.com/trellis3d/trellis/main/assets/example_image/mushroom_1.png \
-    && wget -P assets/example_image https://raw.githubusercontent.com/trellis3d/trellis/main/assets/example_image/soccer_1.png
+    && wget -P assets/example_image https://raw.githubusercontent.com/trellis3d/trellis/main/assets/example_image/mushroom_1.png \    && wget -P assets/example_image https://raw.githubusercontent.com/trellis3d/trellis/main/assets/example_image/soccer_1.png
+
+# Bake BiRefNet model into the image (prevents slow runtime download)
+COPY download_models.py .
+RUN python3 download_models.py
 
 RUN wget -P assets/example_multi_image https://raw.githubusercontent.com/trellis3d/trellis/main/assets/example_multi_image/chair_1.png \
     && wget -P assets/example_multi_image https://raw.githubusercontent.com/trellis3d/trellis/main/assets/example_multi_image/chair_2.png \

download_models.py

@@ -0,0 +1,22 @@
+
+import os
+import torch
+from transformers import AutoModelForImageSegmentation
+
+def download_birefnet():
+    print("⬇️ Downloading BiRefNet model for baking...")
+    try:
+        # This will download the model to the huggingface cache directory
+        # which is persisted in the Docker image if done during build
+        birefnet = AutoModelForImageSegmentation.from_pretrained(
+            'ZhengPeng7/BiRefNet', 
+            trust_remote_code=True
+        )
+        print("✅ BiRefNet model downloaded successfully.")
+    except Exception as e:
+        print(f"❌ Failed to download BiRefNet model: {e}")
+        # Investigate if we need a workaround or specific cache dir
+        raise e
+
+if __name__ == "__main__":
+    download_birefnet()

trellis/utils/postprocessing_utils.py

@@ -337,11 +337,8 @@ def bake_texture(
         mask = texture_weights > 0
         texture[mask] /= texture_weights[mask][:, None]
         
-        # 🔧 Apply gamma correction (Linear → sRGB)
-        # Fixes "dark model" issue by converting from linear color space to sRGB
-        texture_linear = texture.reshape(texture_size, texture_size, 3).cpu().numpy()
-        texture_srgb = np.power(np.clip(texture_linear, 0, 1), 1.0/2.2)  # Gamma 2.2 correction
-        texture = np.clip(texture_srgb * 255, 0, 255).astype(np.uint8)
+        # 🔧 Gamma correction REMOVED as per user request (relying on SH boost instead)
+        texture = np.clip(texture.reshape(texture_size, texture_size, 3).cpu().numpy() * 255, 0, 255).astype(np.uint8)
 
         # inpaint
         mask = (texture_weights == 0).cpu().numpy().astype(np.uint8).reshape(texture_size, texture_size)
@@ -391,11 +388,8 @@ def bake_texture(
                 pbar.set_postfix({'loss': loss.item()})
                 pbar.update()
         
-        # 🔧 Apply gamma correction (Linear → sRGB)
-        # Fixes "dark model" issue by converting from linear color space to sRGB
-        texture_linear = texture[0].flip(0).detach().cpu().numpy()
-        texture_srgb = np.power(np.clip(texture_linear, 0, 1), 1.0/2.2)  # Gamma 2.2 correction
-        texture = np.clip(texture_srgb * 255, 0, 255).astype(np.uint8)
+        # 🔧 Gamma correction REMOVED as per user request (relying on SH boost instead)
+        texture = np.clip(texture[0].flip(0).detach().cpu().numpy() * 255, 0, 255).astype(np.uint8)
         mask = 1 - utils3d.torch.rasterize_triangle_faces(
             rastctx, (uvs * 2 - 1)[None], faces, texture_size, texture_size
         )['mask'][0].detach().cpu().numpy().astype(np.uint8)