Demo feedback: video upload, fix washed-out points, smoother scrub, rename folder

- Accept a video upload (gr.Video); its first MAX_FRAMES frames are decoded and
used (via io_utils.load_frame_paths). Images or video both work.
- Fix washed-out viewer colors: glTF COLOR_0 is linear, so convert our sRGB point
colors to linear before writing the .glb (the viewer re-applies display gamma).
- Smoother frame scrubbing: bind the slider's .release (reload on let-go, not on
every drag tick) and shrink the viewer point cloud to <=60k points per frame.
- Rename the download subfolder colored_points/ -> points/.
Verified on a real run incl. a video input: first-N sampling, linearized GLB
colors, <=60k pts, zip has tracks/ + points/; UI builds on gradio 6.19.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

README.md

@@ -22,9 +22,10 @@ tags:
 
 # Face Anything — Gradio demo
 
-4D face reconstruction and tracking from **any image sequence** (up to 40 frames),
-in a single feed-forward pass. The model jointly predicts depth and **canonical
-facial coordinates**, from which the demo derives:
+4D face reconstruction and tracking from **any image sequence** — upload up to 40
+images **or a video** (its first 40 frames are used) — in a single feed-forward
+pass. The model jointly predicts depth and **canonical facial coordinates**, from
+which the demo derives:
 
 - **Canonical 2D video** — per-frame canonical facial-coordinate map
 - **Depth 2D video** — per-frame depth map (JET)
@@ -32,7 +33,7 @@ facial coordinates**, from which the demo derives:
 - **3D point cloud with colorful tracks** — viewable/orbitable in the 3D viewer
   (rendered as a `.glb` on a white background), with a frame slider to scrub the
   sequence and a downloadable `.zip` containing both the track-colored point
-  clouds (`tracks/`) and the plain colored point clouds (`colored_points/`)
+  clouds (`tracks/`) and the plain colored point clouds (`points/`)
 - *(bonus)* a 2D point-track overlay video
 
 Reconstruction always uses the model's **predicted camera poses** (a multi-view

app.py

@@ -197,15 +197,28 @@ def _sniff_ext(path):
         return ".png"
 
 
-def _prepare_inputs(files, max_frames, workdir):
-    """Normalize uploaded images into a clean, ordered folder of frames.
+def _prepare_inputs(files, video, max_frames, workdir):
+    """Normalize the upload (an image set OR a video) into an ordered list of
+    frame paths.
 
-    Robust to Gradio temp files that lack a usable extension: we don't glob by
-    extension. Files are natural-sorted by their original name (temporal order),
+    A video takes precedence: it is decoded and its first ``max_frames`` frames
+    are used. For images we don't glob by extension (Gradio temp files often lack
+    one): files are natural-sorted by their original name (temporal order),
     capped, then copied as ``frame_XXXX.<ext>`` with a content-sniffed extension.
     """
+    from faceanything.io_utils import load_frame_paths
+
+    if video:
+        vpath = video if isinstance(video, str) else _to_entry(video)[0]
+        if vpath and os.path.exists(vpath):
+            paths, _ = load_frame_paths(
+                vpath, max_frames=int(max_frames), stride=1,
+                work_dir=os.path.join(workdir, "video_frames"))
+            if paths:
+                return paths
+
     if not files:
-        raise gr.Error("Please upload at least one image.")
+        raise gr.Error("Please upload images or a video.")
 
     entries = []
     for f in files:
@@ -233,11 +246,22 @@ def _prepare_inputs(files, max_frames, workdir):
     return out
 
 
-def _points_to_glb(path, points, colors, max_points=200000):
+def _srgb_to_linear(cols_u8):
+    """sRGB uint8 (0-255) -> linear uint8. glTF COLOR_0 vertex colors are
+    interpreted as *linear* and the viewer re-applies the display gamma, so our
+    sRGB image colors must be linearized first or the points render washed-out."""
+    c = np.asarray(cols_u8, np.float32) / 255.0
+    lin = np.where(c <= 0.04045, c / 12.92, ((c + 0.055) / 1.055) ** 2.4)
+    return np.clip(lin * 255.0, 0, 255).astype(np.uint8)
+
+
+def _points_to_glb(path, points, colors, max_points=60000):
     """Write a colored point cloud as a ``.glb`` — the format gradio's Model3D
     actually renders as points (a vertex-only ``.ply`` is treated as an empty
     solid mesh, so nothing shows). ``points`` must already be in glTF axes
-    (X-right, Y-up, Z-back). Colors are (N,3/4) uint8."""
+    (X-right, Y-up, Z-back). Colors are (N,3/4) uint8 in sRGB.
+
+    Kept small (``max_points``) so each frame loads quickly when scrubbing."""
     import trimesh
 
     pts = np.asarray(points, np.float32)
@@ -252,8 +276,9 @@ def _points_to_glb(path, points, colors, max_points=200000):
         pts, cols = pts[idx], cols[idx]
     if cols.dtype != np.uint8:
         cols = np.clip(cols, 0, 255).astype(np.uint8)
+    rgb = _srgb_to_linear(cols[:, :3])
     rgba = np.concatenate(
-        [cols[:, :3], np.full((cols.shape[0], 1), 255, np.uint8)], axis=1)
+        [rgb, np.full((rgb.shape[0], 1), 255, np.uint8)], axis=1)
     scene = trimesh.Scene()
     scene.add_geometry(trimesh.points.PointCloud(vertices=pts, colors=rgba))
     scene.export(path)
@@ -263,6 +288,7 @@ def _points_to_glb(path, points, colors, max_points=200000):
 @GPU(duration=GPU_DURATION)
 def run(
     files,
+    video,
     mode,
     process_res,
     remove_bg,
@@ -303,7 +329,8 @@ def run(
 
     try:
         _say(0.02, "Preparing inputs…")
-        frame_paths = _prepare_inputs(files, min(int(max_frames), MAX_IMAGES), workdir)
+        frame_paths = _prepare_inputs(
+            files, video, min(int(max_frames), MAX_IMAGES), workdir)
         n_in = len(frame_paths)
         if n_in == 0:
             raise gr.Error("No valid images found in the upload.")
@@ -427,7 +454,7 @@ def run(
 
             # Downloadable .ply (repo coords): two colorings in the same zip.
             tracks_dir = os.path.join(workdir, "pointclouds", "tracks")
-            points_dir = os.path.join(workdir, "pointclouds", "colored_points")
+            points_dir = os.path.join(workdir, "pointclouds", "points")
             view_dir = os.path.join(workdir, "view_glb")
             for d in (tracks_dir, points_dir, view_dir):
                 os.makedirs(d, exist_ok=True)
@@ -544,6 +571,9 @@ def build_demo():
                     label="Preview", columns=6, height=180, show_label=True,
                     object_fit="contain",
                 )
+                video = gr.Video(
+                    label=f"…or upload a video (its first {MAX_IMAGES} frames are used)",
+                )
                 mode = gr.Radio(
                     choices=["Joint", "One-by-one"],
                     value="Joint",
@@ -593,7 +623,7 @@ def build_demo():
                     0, 1, value=0, step=1, visible=False,
                     label="Track point-cloud frame")
                 tracks_zip = gr.File(
-                    label="Download point clouds (.zip: tracks/ + colored_points/)")
+                    label="Download point clouds (.zip: tracks/ + points/)")
                 with gr.Tab("Canonical (2D)"):
                     canonical_vid = gr.Video(label="Canonical facial-coordinate map")
                 with gr.Tab("Depth (2D)"):
@@ -609,15 +639,17 @@ def build_demo():
 
         run_btn.click(
             run,
-            inputs=[files, mode, process_res, remove_bg,
+            inputs=[files, video, mode, process_res, remove_bg,
                     conf_percentile, n_tracks, track_k, track_threshold,
                     fps, max_frames],
             outputs=[model3d, canonical_vid, depth_vid, normals_vid, tracks2d_vid,
                      tracks_zip, view_state, frame_slider, status],
             concurrency_limit=1,
         )
-        frame_slider.change(pick_frame, inputs=[frame_slider, view_state],
-                            outputs=model3d)
+        # .release (not .change) so the viewer reloads only when you let go of the
+        # slider — smoother than reloading a glb on every drag tick.
+        frame_slider.release(pick_frame, inputs=[frame_slider, view_state],
+                             outputs=model3d)
     return demo