SAM 3.1 · Video selection
Keep your
selection moving.
Choose an object on one frame. Keep its mask through the whole clip.
Sogni Labs · November 8, 2026
Download the selection preview · Download the lossless mask
The title goes behind the horn.
A character turns. A hand crosses a product. A subject walks through a crowd. A selection made on a still frame needs to follow that motion if you want to put a title behind it or apply an effect to just that object.
SAM 3.1 video selection takes an original clip and one selection prompt, then returns a black-and-white mask for every frame. In this example, the prompt is simply unicorn horn, selected on frame 127. The mask follows the horn in both directions from that frame.
Point at the instance you mean.
Name a visible object with text, click positive and negative points, or give it a box. Select on the clearest frame in the clip; it does not have to be the first. When a text prompt finds several matches, choose one ranked instance with objectIndex. Tracking follows that selection rather than adding every similar object it sees.
A mask that fits your source.
The result is a lossless MP4 at the source’s original dimensions, frame count and frame rate. White is selected; black is outside. Bring the mask and the original clip into your compositor for occlusion, selective effects or other moving selections.
Clips can be up to 30 seconds and 900 frames, with either edge up to 2560 pixels and an upload of 100 MB or less. Use a constant-frame-rate SDR export with square pixels and rotation applied. The output is a hard selection mask; fine fur and hair need a separate matting step.
One video job.
Call sam3_video_segment_fp16 through the JavaScript or Python SDK, or use the Creative Agent CLI:
sogni-agent --segment-video clip.mp4 --segment-text "unicorn horn" \
+ --segment-frame 127 -o mask.mp4The server reads the uploaded video before charging. Pricing follows its frame count and pixel area. A 1920 × 1080 source with 255 frames at 24 fps estimates at about 10.34 Spark before applicable multipliers; request a live estimate for your own clip.
That 10.625-second clip took 38.27 seconds end to end on an RTX 4090 in one qualification run. A 30-second, 720-frame clip at the same dimensions took 106.12 seconds. Supernet upload, queue and delivery times vary.
Explore SAM 3.1 Video on Sogni →
JavaScript, Python and CLI examples →
