UniRestorer++ reports log-linear PSNR scaling to 2.9B params

Universal image restoration, where one model handles rain, haze, noise, blur and more, has a known ceiling: make the backbone bigger and performance tends to plateau. UniRestorer++, posted to Hugging Face by Zhilu Zhang and co-authors, proposes a different scaling path and claims it holds up to 2.9 billion parameters.

The paper's central claim is a scaling law. The authors scale the model from 154 million to 2.9 billion parameters and report "approximately log-linear PSNR trends over the evaluated model-size range." On their new CDD-50 benchmark, which covers 50 degradation categories ranging from single-degradation to penta-degradation, average in-distribution PSNR moves from 26.54 dB at the smallest configuration to 27.05 dB at the largest. The gap widens on real-world data: FoundIR composites climb from 26.22 to 28.17 dB as the model scales.

The architecture is not a vanilla mixture of experts. It routes each input twice. First down a tree of degradation groups to pick a path, then along that path to pick the granularity. "Degradation-estimation-based routing selects a root-to-leaf path within the hierarchy, while granularity-estimation-based routing selects an appropriate level along that path," the authors write, which they argue keeps fine specialization from collapsing when the degradation estimator makes a mistake.

The largest variant packs 510 LoRA experts. Training the 2.9B model took about 23 days on 64 NVIDIA V100 GPUs and 5 million paired samples. The paper reports "stronger scaling gains on out-of-distribution and real-world data than vanilla MoE scaling," along with zero-shot generalization and few-shot transfer capability on real-world images. CDD-50 was built by the same team, so the headline numbers wait on a third party running the same ladder.

It lands alongside a steady run of open computer-vision work we have been tracking, including FactEmbed and NEO-VLA in the past two days.