UniRestorer++ scales universal image restoration with multi-granularity degradation modeling, beating dense and MoE baselines
Jingbo Lin · hf · 2026-11-05
UniRestorer++, released on Hugging Face, scales universal image restoration via multi-granularity degradation modeling and adaptive routing, improving generalization across diverse degradation types.
Accompanying charts show parameter-PSNR scaling curves on CDD-50 (in-distribution and out-of-distribution degradation parameters) and FoundIR (real-world): across 1.41M–3.47M parameters, its hierarchical model scaling rises near-log-linearly (fitted R² of 0.955–0.986), consistently above dense and vanilla MoE baselines of the same size, and avoids the drop-off dense models show at larger scales.
More from Multimodal
- Recraft V4.1 turns flat 2D cartoon characters into consistent 3D renders — aziz4ai · 2026-11-05
- Voyager agent picks AI video models per shot, claims 2-3x fewer generations at half price — rohanpaul_ai · 2026-11-05
- Simulated 'ancient surveillance' videos go viral: full MiniMax H3 workflow revealed — Hailuo_AI · 2026-11-05
- Canva explores character-specific feature for CRAFT model, shows new extensions — pbaylies · 2026-11-05
- LiveFoley: real-time streaming video-to-audio generation with latency below perceptual thresholds — Yusheng Dai · 2026-11-05
- 29 speech-to-speech model launches in one year — when does s2s overtake cascades? — rdesh26 · 2026-11-05