Segmentation settings and limits
Segmentation settings control how RasterForge turns a render and profile into layers. These values matter because they affect output quality, run cost, and whether a run is allowed to start.
Workflows
| Workflow | Scope | Use when |
|---|---|---|
| Find things (multi-image) | Active image, selected images, or all images depending on run target. | The same class or object type should be found across imagery. |
| Select objects (single image) | Active image only. | You need point or box guided segmentation on one image. |
| Auto masks (object candidates) | Active image, selected images, or all images depending on run target. | You want class-agnostic object candidates before assigning meaning through review or downstream cleanup. |
In the current app, SAM 3 flow labels include pcs, pcs_ref, pvs, and amg. User-facing workflows should be understood as Find things, Select objects, and Auto masks; the model flow label is the current backend mode used to run that workflow.
Profile and guidance limits
| Limit | Default beta value |
|---|---|
| Text prompt length | 100 characters |
| Saved examples per segmentation profile | 5 |
| Image guidance points plus boxes | 20 |
The app blocks runs that exceed these limits. It also blocks Find things with saved examples when no example is attached.
Find and select run settings
| Setting | Default | Purpose |
|---|---|---|
| Detection threshold | 0.5 | Raises or lowers which detections are accepted in Find things workflows. |
| Mask threshold | 0.5 | Controls mask confidence cutoff. |
| Max instances | 200 | Limits object count returned by a run. |
| Box inference | Multi-forward | Controls how multiple boxes are sent to the model. |
| Reference margin | 100 px | Adds context around saved examples. |
| Pack gap | 8 px when examples are used | Separates packed references. |
Auto masks presets
Auto masks uses automatic mask generation settings. The preset is a starting point, not a quality guarantee.
| Preset | Use when | Behavior |
|---|---|---|
| Balanced | You want general object candidates without making the run too dense. | Uses the default candidate density and cleanup settings. |
| Dense | You want more candidates and are willing to review more fragments. | Increases grid density and keeps more masks. |
| Clean | You want fewer, stronger candidates. | Raises quality and area thresholds and keeps fewer masks. |
Advanced Auto masks controls include points per side, points per batch, predicted IoU threshold, mask threshold, stability score threshold, minimum mask region area, max masks, NMS IoU threshold, and postprocess settings.
Postprocess defaults
| Setting | Default |
|---|---|
| Minimum score | 0.5 |
| Minimum area | 100 px |
| Maximum area | Off |
| Keep top K | Off |
| IoU deduplication threshold | 0.85 |
| Simplify tolerance | 1.5 |
Postprocess settings are especially important for GeoJSON export. Use them to remove tiny fragments, deduplicate overlapping objects, and simplify vector geometry. Auto masks uses related postprocess defaults but may set different minimum score, minimum area, and keep-top-K values through its presets.
Patch accounting
SAM 3 runs are charged in 256 by 256 pixel patches. RasterForge estimates patches from the rendered input dimensions and sums patches across multi-image runs. Smaller AOIs and appropriate full render sizes reduce daily patch usage.