SAM 3 in RasterForge
SAM 3 is the current promptable segmentation engine behind RasterForge Segment runs. RasterForge wraps the model in project workflows for imports, render profiles, segmentation profiles, run queues, layer review, object editing, and export.
Prompt types
RasterForge can use text prompts, positive points, negative points, boxes, saved examples, and profile defaults. These inputs define what SAM 3 should segment in the rendered image.
Current user workflows map to model modes:
- Find things: search for repeated targets using text, boxes, and optionally saved examples.
- Select objects: use points or boxes to select targets on the active image.
- Auto masks: generate class-agnostic object candidates with automatic mask generation settings.
Satellite and aerial considerations
Remote sensing scenes often have repeated texture, seasonal variation, clouds, shadows, mixed pixels, compression artifacts, and multiple objects that share similar color. A good workflow usually combines a targeted render, narrow prompt, representative examples, careful thresholds, and manual review.
Auto masks can be useful for broad discovery or cleanup workflows, but it does not know the project class by itself. Treat it as candidate generation until a user reviews, edits, names, or exports the useful layer.
Single image before broader runs
Always test a segmentation profile or Auto masks preset on a representative image before applying it to selected images or a full project. This keeps a weak prompt, bad render, poor threshold, or overly dense candidate setting from multiplying across many layers.