Skip to content

SarabResolver

SarabResolver resolves your footage's depth, confidence, sky and camera (position, rotation, lens) per frame with Depth Anything 3, so effects like SarabOrtex shaders and SarabBrush3D strokes can live in the shot's real 3D space — matched to the camera and occluded by real objects.

Adding it to your shot

Apply SarabResolver from the SarabFX category (SarabFX/3D) onto the clip you shot. Pick a model, click Download Model the first time (a progress bar shows the transfer), then choose an Output.

Models

Model Gives Size Licence
DA3 Small depth + confidence + camera 105 MB Apache-2.0
DA3 Base depth + confidence + camera 149 MB Apache-2.0
DA3 Large depth + confidence + camera, higher quality 1.4 GB CC BY-NC 4.0 — non-commercial
DA3 Giant + 3D Gaussians export 4.9 GB CC BY-NC 4.0 — non-commercial
DA3 Mono Large depth + sky (no camera) 1.3 GB Apache-2.0
DA3 Metric Large depth in metres + sky (no camera) 1.3 GB Apache-2.0

Models are stored in ~/Library/Application Support/SarabFX/models/depth-anything.

Outputs

Output Use it for
Depth map occlusion and depth-driven effects (feed it to a shader channel)
Confidence map masking where the depth is uncertain
Sky mask replacing or grading the sky
Source + camera data pass the footage through with the camera encoded for SarabBrush3D and SarabOrtex
Source with depth in alpha footage in RGB, depth in alpha — one clip carries both

Depth Mapping: Normalise per frame for a look, Fixed range (Near..Far) for anything that must be stable across frames (occlusion), Raw values for the model's own units.

Camera

  • Static — each frame on its own: depth and lens, camera at the origin.
  • Tracked — frames are solved in overlapping windows so the camera path is continuous across the clip. Needs a depth+camera model.

Resolve Clip Now solves every frame once with a progress bar; afterwards playback is served from the cache, which persists on disk. Start it from the first frame. The first frame's camera becomes the origin of the 3D world.

Keeping the camera steady

Setting Default What it does
Lock Lens on Uses one focal length for the whole clip, so inserted objects don't "breathe". Turn it off only for shots that really zoom.
Camera Smoothing 4 frames Averages the solved camera over neighbouring frames. This removes the model's frame-to-frame jitter, which a steady plate doesn't have.
Hold When Still on When the footage doesn't move between two frames (a locked-off camera, or a pan that has come to rest), the camera stays put. Without it the model's pose wanders on identical frames and inserted objects slide away while the shot stands still. Slow moves and forward walks are still tracked.
Restore Plate Shake on Measures the footage's own shake, the way a stabiliser does, and adds it back to the smoothed camera. Parallax is taken into account, so a handheld shot keeps its real shake and inserted objects stay locked to the picture.

When these settings change, the clip has to be resolved again, so click Resolve Clip Now.

Objects crossing the foreground. When something passes close in front of the lens — a pillar, a person, a door frame — the depth model has little to go on for those frames. SarabResolver keeps the world it already has: the scale of each new window is taken from the far part of the picture, a frame whose solved camera moves more than the footage itself shows (rotation or travel) is moved by the footage's measured flow instead, and the track picks the scene up again when it comes back into view. Only a real cut — the picture changes completely between two frames — starts a new world, and the log says so (cut at frame N).

Using it with SarabOrtex

Set Output to Source with depth in alpha and connect SarabOrtex after it. The shader then receives the footage, its depth and the camera in one input. It can composite itself into the shot, hide behind real objects, and cast light and shadow on real surfaces. See Shaders in real footage.

Exports

  • Export Camera — a JSON file with the per-frame camera (extrinsics, intrinsics, and the SarabFX position/rotation/FOV).
  • Export Gaussians — with DA3 Giant, the current frame's 3D Gaussians as a PLY.

Performance

Inference runs at about 300 ms per frame with DA3 Small on an M3 Max; resolved frames are cached, so scrub or play the clip once (or use Resolve Clip Now) and it stays instant.