Update app.py

app.py

@@ -476,17 +476,10 @@ def render_single_frame(
         blank = np.ones((frame_height, frame_width, 3), dtype=np.uint8) * 200
         return blank
     
-    # Use vertices as-is (rotation should be applied before calling this function)
-    # But if fixed_center is None, we need to compute from unrotated vertices
-    # If fixed_center is provided, it should be from rotated vertices
-    verts_used = verts.copy()
-    
-    # Compute center for camera positioning
-    mesh_center = verts_used.mean(axis=0)
-    if fixed_center is not None:
-        camera_target = fixed_center
-    else:
-        camera_target = mesh_center
+    # IMPORTANT: Rotate mesh 180 degrees around X-axis (like visualize.py)
+    # This fixes the coordinate system so we view from the front
+    rot_matrix = trimesh.transformations.rotation_matrix(np.radians(180), [1, 0, 0])
+    verts_rotated = np.dot(verts, rot_matrix[:3, :3].T)
     
     # Create scene
     scene = pyrender.Scene(bg_color=bg_color, ambient_light=[0.4, 0.4, 0.4])
@@ -499,10 +492,19 @@ def render_single_frame(
         baseColorFactor=color
     )
     
-    # Create mesh (exactly like compare_vqvae.py)
-    mesh = trimesh.Trimesh(vertices=verts_used, faces=faces)
+    # Create mesh with rotated vertices
+    mesh = trimesh.Trimesh(vertices=verts_rotated, faces=faces)
+    # Fix normals to ensure proper face rendering
+    # This is critical for proper face/lip visibility - ensures all faces point outward
+    mesh.fix_normals()
+    # Recompute vertex normals for smooth shading
+    mesh.vertex_normals
     mesh_render = pyrender.Mesh.from_trimesh(mesh, material=material, smooth=True)
     scene.add(mesh_render)
+    
+    # Compute center for camera positioning (using rotated vertices)
+    mesh_center = verts_rotated.mean(axis=0)
+    camera_target = fixed_center if fixed_center is not None else mesh_center
 
     # Camera setup
     camera = pyrender.IntrinsicsCamera(
@@ -511,13 +513,22 @@ def render_single_frame(
         znear=0.1, zfar=20.0
     )
 
-    # Camera pose: positioned BEHIND the subject (at +Z), looking at -Z (toward face)
-    # This matches compare_vqvae.py exactly
+    # Camera pose: positioned in front of the subject, looking at them
+    # After 180-degree rotation, Z points down
+    # Camera should be at negative Z (in front) looking at positive Z
+    # This matches visualize.py exactly
     camera_pose = np.eye(4)
     camera_pose[0, 3] = camera_target[0]                    # Center X
-    camera_pose[1, 3] = camera_target[1]                    # Center Y  
-    camera_pose[2, 3] = camera_target[2] + camera_distance  # BEHIND (positive Z)
-    camera_pose[:3, :3] = np.eye(3)  # Identity = look at -Z
+    camera_pose[1, 3] = camera_target[1]                    # Center Y (body center)
+    camera_pose[2, 3] = camera_target[2] - camera_distance  # In front, distance controls zoom
+    
+    # Camera orientation: flip to look at subject (SOKE-style)
+    # This rotation makes camera look toward +Z (at the subject)
+    camera_pose[:3, :3] = np.array([
+        [1,  0,  0],
+        [0, -1,  0],
+        [0,  0, -1]
+    ])
 
     scene.add(camera, pose=camera_pose)
 
@@ -619,17 +630,16 @@ def render_video(
     # Apply rotation to all frames
     # (T, V, 3) dot (3, 3) -> (T, V, 3)
     verts = np.matmul(verts, rot_matrix[:3, :3].T)
-    # Apply orientation fix: rotate 180 degrees around X-axis (like compare_vqvae.py)
-    verts = verts.copy()
-    verts[..., 1:] *= -1
-    
     # Trim last few frames to remove end-of-sequence artifacts
     T_total = verts.shape[0]
     trim_amount = min(8, int(T_total * 0.15))
     T = max(5, T_total - trim_amount)
 
-    # Compute fixed camera target from first frame (using rotated vertices)
-    fixed_center = verts[0].mean(axis=0)
+    # Compute fixed camera target from first frame (rotation happens inside render_single_frame)
+    # We need to compute it from rotated vertices for consistency
+    rot_matrix = trimesh.transformations.rotation_matrix(np.radians(180), [1, 0, 0])
+    verts_rotated_first = np.dot(verts[0], rot_matrix[:3, :3].T)
+    fixed_center = verts_rotated_first.mean(axis=0)
 
     frames = []
     for t in range(T):
@@ -668,12 +678,6 @@ def render_comparison_video(
     if not ensure_pyrender():
         raise RuntimeError("PyRender not available")
 
-    # Apply orientation fix: rotate 180 degrees around X-axis (like compare_vqvae.py)
-    verts1 = verts1.copy()
-    verts2 = verts2.copy()
-    verts1[..., 1:] *= -1
-    verts2[..., 1:] *= -1
-
     # Match lengths and trim
     T_total = min(verts1.shape[0], verts2.shape[0])
     trim_amount = min(8, int(T_total * 0.15))
@@ -682,9 +686,13 @@ def render_comparison_video(
     verts1 = verts1[:T]
     verts2 = verts2[:T]
 
-    # Compute fixed camera targets
-    fixed_center1 = verts1[0].mean(axis=0)
-    fixed_center2 = verts2[0].mean(axis=0)
+    # Compute fixed camera targets (rotation happens inside render_single_frame)
+    # We need to compute from rotated vertices for consistency
+    rot_matrix = trimesh.transformations.rotation_matrix(np.radians(180), [1, 0, 0])
+    verts1_rotated_first = np.dot(verts1[0], rot_matrix[:3, :3].T)
+    verts2_rotated_first = np.dot(verts2[0], rot_matrix[:3, :3].T)
+    fixed_center1 = verts1_rotated_first.mean(axis=0)
+    fixed_center2 = verts2_rotated_first.mean(axis=0)
 
     labels = [label1, label2]