Settings reference

MAGIC ROTO: Select Objects (image workflow)

Setting / input	Default	What it does	Tips
`text`	empty	What to find by name. Comma-separate several terms; `term:N` keeps up to N of that term	Leave it empty to use only clicks and boxes. Without `:N`, each term keeps its single best match
`threshold`	0.4	Confidence needed for a text match	Raise it if wrong things are found. Lower it if the object is missed
`refine_iterations`	2	Extra SAM decoder passes that tighten each mask	0 = raw detector mask. 2–3 is a good balance
`positive_coords` / `negative_coords`	from Points Editor	Include and exclude clicks. All clicks together make one object	Add red (exclude) points to cut away parts SAM3 grabbed by mistake
`bboxes`	from Points Editor	Each box becomes its own object	Draw boxes snug around the object
`editor_mask`	Points Editor `bbox_mask`	Tells the node the editor's canvas size, so clicks land exactly on the image	Keep it connected

Outputs: masks (one per object, in the order points → boxes → text), count, and labels (e.g. points, box1, text1).

No settings. Image + matte → RGBA image. Feed it to Save Image to get transparent PNGs.

Setting	Default in workflow	What it does	When to change it
`detection_threshold`	0.4	How confident SAM3 must be to accept an object	Raise it (0.5–0.6) if it picks wrong things. Lower it (0.3) if it misses the subject
`max_objects`	1	How many objects to keep	Raise it for several people or a prompt with several items
`detect_interval`	1	Re-detect every N frames	Keep it at 1 for clips with cuts. 2–5 is faster on single continuous shots

Text prompt (the What to cut out node): a short noun phrase. Separate several subjects with commas; word:N caps the count for that word.

Setting	Default	What it does	Tips
`erode_px`	8	Shrinks the mask inward. Whatever stays is "definitely object"	Increase it if the edge of the mask is wrong or halo-y
`dilate_px`	20	Grows the mask outward. Beyond it is "definitely background"	5–10 for hard edges (products, cars) · 20–40 for hair, fur, glow, motion blur

Setting	Default	What it does	Tips
`model_name`	vitmatte-small-composition-1k	Which VitMatte model in `models/vitmatte/` to use	Any VitMatte folder with a `config.json` shows up here
`max_side`	1280	Long side the frame is scaled to for VitMatte	Lower it (960) if you get out-of-memory errors. Raise it (1920) for 4K detail if you have VRAM
`white_point`	0.99	Alpha above this becomes solid 1.0	Use 0.95 if the interior still looks semi-transparent

Setting	Default	What it does	Tips
`strength`	0.4	How much of the previous frame is blended in (0 = off)	0.2–0.4 for moving subjects · 0.5–0.7 for locked-off, slow shots
`motion_threshold`	0.25	Alpha changes larger than this count as real motion and are not blended	Lower it if you see trails on fast motion
`trimap` (optional input)	connected	Limits smoothing to the edge band	Keep it connected so the solid interior is never touched

Setting	Default	What it does
`color`	`#00FF00`	Background colour for the preview cut-out. Try `#000000` or `#FFFFFF` to check edges against dark and light backgrounds