meta/sam-3/video-to-video

What makes SAM 3 stand out

SAM 3 is a high-fidelity video-to-video segmentation and tracking model built to preserve structure, identity, and temporal coherence. SAM 3 converts raw footage into stable, frame-accurate masks driven by concise natural-language prompts, enabling targeted operations without regenerating full frames. With open-vocabulary understanding and frame-by-frame identity control, SAM 3 maintains consistent masks through motion, occlusion, and lighting changes. Its design emphasizes reliability and responsiveness for iterative workflows that demand precise boundaries, low drift on long shots, and consistent results suitable for downstream editorial and VFX pipelines. Key capabilities of SAM 3:

• Open-vocabulary segmentation from concise text prompts.
• Identity-stable tracking with persistent IDs across frames.
• Structure- and motion-aware boundaries that respect edges and articulation.
• Occlusion-tolerant propagation with minimal temporal jitter.
• Frame-consistent masks ready for compositing, grading, and clean plate work.
• Real-time responsiveness from SAM 3 for interactive review and rapid iteration.
• Fine-grained regional control to isolate subjects or restrict changes.

Prompting guide for SAM 3

When using SAM 3, start by supplying video_url and a focused text_prompt that names the subject to segment and track. Describe the target with clear nouns and a few attributes, and state what to exclude or preserve to help SAM 3 focus. Use consistent terminology across shots to help SAM 3 maintain continuity. For crowded scenes, begin with the primary subject, review the mask, then refine with small prompt adjustments so SAM 3 stays stable. SAM 3 supports open-vocabulary phrasing, but specificity improves stability, especially across fast motion or partial occlusion. Examples:

• red sedan car, foreground only, include wheels and windows
• person with yellow jacket, exclude hands and face, background unchanged
• soccer ball only, ignore players and crowd, maintain full silhouette
• brown dog, keep ears and tail, do not segment grass
• neon sign on the left wall, background element only Pro tips for SAM 3:
• Be explicit about scope: say what to isolate and what must not change.
• Use spatial cues like left, right, center, near camera, background only.
• Prefer a few strong attributes over long descriptive lists.
• Keep prompts consistent across cuts for SAM 3; re-run at clear scene boundaries.
• Higher-resolution, less-compressed footage yields cleaner edges and steadier masks in SAM 3.