Fix TripoSG NaN cascade and remove stale autocast wrapper

- inference_utils.py: replace empty next_points guard — instead of
filling next_logits with all-(-10000) sentinel (which cascades to
all-NaN after the loop), trilinearly upsample the coarse grid to the
target resolution so real SDF values are preserved through refinement
- app.py: remove torch.autocast wrapper around run_triposg(); autocast
was previously identified (commit 9757b8d) as causing NaN / no-isosurface
issues and must not re-wrap the inference call

app.py

@@ -880,9 +880,8 @@ def generate_shape(input_image, remove_background, num_steps, guidance_scale,
         img.save(img_path)
 
         progress(0.5, desc="Generating shape (SDF diffusion)...")
-        with torch.autocast(device_type="cuda", dtype=torch.float16):
-            from scripts.inference_triposg import run_triposg
-            mesh = run_triposg(
+        from scripts.inference_triposg import run_triposg
+        mesh = run_triposg(
                 pipe=pipe,
                 image_input=img_path,
                 rmbg_net=rmbg_net if remove_background else None,

patches/triposg/triposg/inference_utils.py

@@ -431,8 +431,13 @@ def flash_extract_geometry(
 
         next_points = torch.stack(nidx, dim=1)
         if next_points.shape[0] == 0:
-            # No active voxels at this refinement level — skip and carry forward
-            grid_logits = next_logits.unsqueeze(0)
+            # No active voxels at this refinement level — upsample coarse grid instead of
+            # filling with all-sentinel values (which cascade to all-NaN at extract time)
+            print(f"[TripoSG] skip level {octree_depth_now}: empty near-surface, upsampling coarse grid")
+            tgt = tuple(int(x) for x in grid_size)
+            grid_logits = F.interpolate(
+                grid_logits.float().unsqueeze(1), size=tgt, mode='trilinear', align_corners=False
+            ).squeeze(1).to(dtype)
             continue
         next_points = (next_points * torch.tensor(resolution, dtype=torch.float32, device=device) +
                         torch.tensor(bbox_min, dtype=torch.float32, device=device))