BiRefNet, Moebius, VOSR and DDColor: four objectives for matting, inpainting, restoration and color
Four different objectives—matting, masked inpainting, generative SR and colour regression—kept as separate product contracts.
IMAGE SPECIALISTS
Matting optimises boundaries and alpha. Inpainting generates inside a mask. Super-resolution balances an LR condition against detail priors. Colourisation predicts chroma with luminance fixed.
Non-equivalent losses
A matting objective may combine L_mat=λ₁L_BCE+λ₂L_IoU+λ₃L_edge. Masked generation obeys x_out=(1-m)⊙x+m⊙G(x,m). VOSR adds a controllable trade-off between LR fidelity and generated detail.
Four outputs imply four objectives
BiRefNet predicts foreground probability and boundaries, Moebius fills a masked region, VOSR generates high-frequency detail from a low-resolution observation, and DDColor estimates chroma. RGBA, inpainting, super-resolution and colourisation cannot share one quality score.
Keep paper numbers inside their task
Moebius's 15× figure belongs to the paper's named hardware and baseline. VOSR's multi-step and one-step models also represent different cost–quality trade-offs. A speed or score from one specialist does not establish performance on the other three tasks.
Model lineage and specialist branches
Matting, inpainting, super-resolution and colorization all accept images but optimize different targets. BiRefNet predicts foreground boundaries, Moebius fills or redraws regions, VOSR restores detail, and DDColor estimates color. Calling all four 'enhancement' hides which pixels may change.
Specialists are useful because their contracts are explicit: matting should preserve RGB, inpainting may generate inside a mask, upscaling must control texture hallucination, and colorization is inherently underdetermined. One workspace standardizes I/O and jobs; acceptance remains separate.
Four specialist tools, four conditional problems
ŷ=T_task(x,m,c), task∈{matte,inpaint,restore,colorize}xsource imagemoptional maskctext or task condition
BiRefNet predicts foreground/alpha, Moebius consumes a mask and optional text, VOSR reconstructs high-resolution detail, and DDColor estimates chroma. The equation is deliberately task-indexed; it does not pretend they share one loss.
Matting audits edges and translucency; inpainting audits mask seams; restoration audits invented texture; colorization audits plausible chroma. One clarity score would hide these failures.
Paper, code and weights can be followed through BiRefNet / Moebius / VOSR.
Compare generation, typography and editing separately
Each row retains its source architecture, metric or task definition; results without a shared protocol remain separate.
| Model | Architecture or evidence | Scope |
|---|---|---|
| BiRefNet | foreground alpha or mask | edges, holes, transparency |
| Moebius | generation inside a mask | seams, semantics, perspective |
| VOSR / DDColor | detail recovery / color estimate | texture hallucination / multiple solutions |
Typography, composition and preservation on real material
BiRefNet needs hair, glass and motion-blur edges; Moebius needs seams, perspective and repeated structures; VOSR must separate recoverable texture from invented text; DDColor must acknowledge multiple plausible colors.
Hold resize kernel, color space, compression and alpha premultiplication constant. Otherwise halos and sharpness differences may come from I/O rather than the network.
What image examples leave uncovered
Colorization and occlusion repair cannot recover a unique historical fact; restored archival imagery should be labeled as interpretive.
Generated tiny text, texture and faces are not evidence from the original image and must not support forensic or identity claims.
Papers and public files for BiRefNet
Model names, numbers and limitations trace to the papers, repositories or model cards below.