Repository navigation
Better background removal for image-to-3D
Fixes the 3.37.3 regressions where generated models came out very bright and often grew a flat block: the uniform-background rescue (added to stop thin limbs being amputated by the matte) was re-matting the whole frame, so drop shadows, vignettes and backdrop gradients were kept as foreground — TRELLIS.2 reconstructed them as slabs and their bright pixels washed out the texture bakes.
- Shading rejection: pixels that are just a darker/brighter version of the background color (shadows, vignettes, lighting falloff — including deep/clipped shadows) stay background; only a real chroma difference marks geometry.
- Connectivity: rescued regions must connect to the detected subject — a saliency-missed arm touches the body; a watermark or gradient blob does not.
- Safety valve: a rescue that would flood the frame bails out entirely.
The thin-limb rescue benefit is preserved (the T-pose goblin still generates with both arms), now without slabs or washed-out colors.
Full Changelog: 3.37.3...3.37.4
Result
| before | after |
|---|---|
![]() |
![]() |

