Update app.py

app.py

@@ -100,11 +100,13 @@ def image_to_3d(
     slat_guidance_strength: float,
     slat_sampling_steps: int,
     req: gr.Request,
-) -> Tuple[dict, str, str]:
+) -> Tuple[dict, str]:
     """
-    Convert an image to a 3D model and save full-quality GLB.
+    Convert an image to a 3D model with memory management.
     """
     user_dir = os.path.join(TMP_DIR, str(req.session_hash))
+    
+    # Generate base outputs
     outputs = pipeline.run(
         image,
         seed=seed,
@@ -120,29 +122,82 @@ def image_to_3d(
         },
     )
 
-    # Generate video preview
-    video = render_utils.render_video(outputs['gaussian'][0], num_frames=120)['color']
-    video_geo = render_utils.render_video(outputs['mesh'][0], num_frames=120)['normal']
+    # Clear CUDA cache after model generation
+    torch.cuda.empty_cache()
+    
+    # Generate video preview in smaller batches
+    video = []
+    video_geo = []
+    batch_size = 30  # Process 30 frames at a time
+    num_frames = 120
+    
+    for i in range(0, num_frames, batch_size):
+        end_idx = min(i + batch_size, num_frames)
+        curr_frames = end_idx - i
+        
+        # Generate color frames
+        batch_frames = render_utils.render_video(
+            outputs['gaussian'][0],
+            num_frames=curr_frames,
+            start_frame=i
+        )['color']
+        video.extend(batch_frames)
+        
+        # Generate geometry frames
+        batch_geo = render_utils.render_video(
+            outputs['mesh'][0],
+            num_frames=curr_frames,
+            start_frame=i
+        )['normal']
+        video_geo.extend(batch_geo)
+        
+        # Clear cache after each batch
+        torch.cuda.empty_cache()
+
+    # Combine and save video
     video = [np.concatenate([video[i], video_geo[i]], axis=1) for i in range(len(video))]
     trial_id = str(uuid.uuid4())
     video_path = os.path.join(user_dir, f"{trial_id}.mp4")
     imageio.mimsave(video_path, video, fps=15)
+    
+    # Clear memory
+    del video
+    del video_geo
+    torch.cuda.empty_cache()
+    
+    # Pack state and return
+    state = pack_state(outputs['gaussian'][0], outputs['mesh'][0], trial_id)
+    return state, video_path
 
-    # Create full quality GLB while we have the data in memory
-    full_glb_path = os.path.join(user_dir, f"{trial_id}_full.glb")
+@spaces.GPU
+def export_full_quality_glb(
+    state: dict,
+    req: gr.Request,
+) -> Tuple[str, str]:
+    """
+    Export a full-quality GLB file with memory management.
+    """
+    user_dir = os.path.join(TMP_DIR, str(req.session_hash))
+    gs, mesh, trial_id = unpack_state(state)
+    
+    # Clear cache before starting
+    torch.cuda.empty_cache()
+    
     glb = postprocessing_utils.to_glb(
-        outputs['gaussian'][0],
-        outputs['mesh'][0],
+        gs,
+        mesh,
         simplify=0.0,  # No simplification
         fill_holes=True,
         fill_holes_max_size=0.04,
         texture_size=2048,  # Maximum texture resolution
-        verbose=False
+        verbose=True  # Show progress
     )
-    glb.export(full_glb_path)
+    glb_path = os.path.join(user_dir, f"{trial_id}_full.glb")
+    glb.export(glb_path)
     
-    state = pack_state(outputs['gaussian'][0], outputs['mesh'][0], trial_id)
-    return state, video_path, full_glb_path
+    # Clear cache after finishing
+    torch.cuda.empty_cache()
+    return glb_path, glb_path
 
 @spaces.GPU
 def extract_glb(
@@ -152,7 +207,7 @@ def extract_glb(
     req: gr.Request,
 ) -> Tuple[str, str]:
     """
-    Extract a reduced-quality GLB file from the 3D model.
+    Extract a GLB file from the 3D model.
     """
     user_dir = os.path.join(TMP_DIR, str(req.session_hash))
     gs, mesh, trial_id = unpack_state(state)
@@ -165,8 +220,9 @@ with gr.Blocks(delete_cache=(600, 600)) as demo:
     gr.Markdown("""
     ## Image to 3D Asset with [TRELLIS](https://trellis3d.github.io/)
     * Upload an image and click "Generate" to create a 3D asset. If the image has alpha channel, it be used as the mask. Otherwise, we use `rembg` to remove the background.
-    * The full-quality GLB will be available immediately after generation.
-    * You can also create a reduced size version using the GLB Extraction Settings.
+    * After generation:
+        * Click "Download Full-Quality GLB" for maximum quality
+        * Or use GLB Extraction Settings for a reduced size version
     """)
     
     with gr.Row():
@@ -233,12 +289,21 @@ with gr.Blocks(delete_cache=(600, 600)) as demo:
     ).then(
         image_to_3d,
         inputs=[image_prompt, seed, ss_guidance_strength, ss_sampling_steps, slat_guidance_strength, slat_sampling_steps],
-        outputs=[output_buf, video_output, download_full],
+        outputs=[output_buf, video_output],
     ).then(
         lambda: [gr.Button(interactive=True), gr.Button(interactive=True), gr.Button(interactive=False)],
         outputs=[download_full, extract_glb_btn, download_reduced],
     )
 
+    download_full.click(
+        export_full_quality_glb,
+        inputs=[output_buf],
+        outputs=[model_output, download_full],
+    ).then(
+        lambda: gr.Button(interactive=True),
+        outputs=[download_full],
+    )
+
     extract_glb_btn.click(
         extract_glb,
         inputs=[output_buf, mesh_simplify, texture_size],