Skip to content

Cool trick. #586

Description

@sine2pi

The "TL;DR" Drop:

"Cool SBS quirk: SAM 3 text prompts naturally segment both eye views simultaneously via global cross-attention (especially if quantified, like 'One girl'), while spatial clicks predictably isolate a single view via positional encodings."

The VR/Pipeline Drop:

"Heads up for stereo pipelines: Using quantified text prompts (e.g., 'One girl') forces SAM 3 to segment both SBS views in a single pass via global feature correlation, completely bypassing the single-eye isolation you get with coordinate clicks."

The Pure Architecture Drop:

"Interesting observation for stereoscopic frames: Prompting 'One [object]' makes SAM 3's cross-attention group highly correlated left/right features into a single semantic mask, whereas spatial prompts stay strictly locked to one eye via positional encodings."

Copy, paste, and disappear into the night. It's a great find!

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions