image_urlsfile-list
Input images to reconstruct. Accepts JPEG, PNG, WEBP and the other formats Pillow decodes. Images are padded to a square and resized so the longest side is 518 pixels.
video_urlfile
Optional input video. Frames are sampled every `frame_sampling_rate`-th frame (the first and last frames are always included) and appended after `image_urls`.
frame_sampling_rateintegerdefault: 24
Sample every n-th frame of `video_url`. The first and last frames of the video are always included. Ignored when no video is given.
alpha_blend_ontoselectdefault: "white"
How to handle images with an alpha channel. 'white'/'black' composite onto that background, 'mean' composites onto the ImageNet mean RGB, and 'keep' discards alpha and keeps the original pixel values.
Options: keep, white, black, mean
export_prediction_databooleandefault: true
Return one JSON file per frame with the raw model outputs (camera pose encoding, depth, depth confidence, mask).
export_depth_mapsbooleandefault: true
Return one 16-bit grayscale PNG depth raster per frame. Pixel values map linearly from `depth_range` onto 0-65535.
export_point_cloudbooleandefault: true
Return a GLB scene holding the fused coloured point cloud plus a cone mesh per estimated camera.
array_encodingselectdefault: "base64"
How arrays are encoded inside the prediction JSON files. 'base64' emits {data, shape, dtype} objects; 'list' emits nested JSON lists, which are far larger.
Options: base64, list
exclude_keysstring-list
Keys to omit from each prediction JSON file.
confidence_percentilenumberdefault: 50
Drop this percentage of the lowest-confidence points before building the point cloud. 0 keeps every point.
max_pointsintegerdefault: 250000
Upper bound on the number of points written to the GLB. Points above the default 250,000-point budget are evenly subsampled to limit GLB serialization and upload latency.
enable_safety_checkerbooleandefault: true
Enable safety checking of the input images and video.
Genbaz
