Wonder 4 routes experts per image patch, so blur, noise and scratches in one photo get separate fixes
Synopsis
Real photographs often carry blur, noise, compression blocks and scratches at once, and the corrections conflict, while existing unified models describe the whole image with one global description and cannot say where each correction belongs. Wonder 4 places a set of structurally identical small expert networks in every block of its backbone, lets a router local to that block pick experts patch by patch, and replaces the language-model rule of equal expert use with a minimum share of use per expert. On three test settings over DIV2K and Flickr2K it reaches the best LPIPS, DISTS and NIQE against five published methods and three open-weight pipelines, though its PSNR is not the highest.
Interpretation
Image quality enhancement no longer depends on one degradation description for the whole image; each patch decides which repair experts to use. Earlier unified restoration models condition on a global degradation description, whether an inferred code, a prompt, text from a vision-language model or a task label, and one description cannot say which region each correction belongs to; obtaining it also needs a second network. On DIV2K validation under the NTIRE 2018 mixed degradation and under bicubic downsampling, and on Flickr2K, Wonder 4 reaches the best LPIPS, DISTS and NIQE against five published methods and three open-weight pipelines; the largest LPIPS and DISTS reductions relative to the strongest published baseline appear on Flickr2K.
The router makes a separate choice in every block for every patch, so different regions of one image can take different expert combinations. Standard sparse mixture of experts inherits the language-model rule of equal use, which applies every correction everywhere, including where it is not needed, so stars in a night sky would be erased as physical damage. Routing maps show early blocks using almost the same expert everywhere, while later blocks follow image structure, for example one expert covering most patches on a person and only a small part of the sky; the within-image routing mutual information averaged over blocks is clearly above zero, while a router that ignores location or decides once per image gives zero.
Each expert is only required to reach a minimum share of use rather than matching the others. Under equal use, an expert needed on only a few patches is pushed onto patches where it should not be used, and every expert learns the same average correction. The bias follows the auxiliary-loss-free load balancing of DeepSeek-V3 and acts only on selection scores, not on the probabilities; an expert below the floor gets higher selection scores until it is selected, and one above the floor is left unchanged, while the training objective adds a routing entropy penalty and a cosine-similarity penalty between expert weights to keep experts from learning the same correction.
Perspective
The demonstrated scope is super-resolution-style evaluation on DIV2K validation under the NTIRE 2018 mixed degradation and bicubic downsampling settings and on Flickr2K, plus qualitative examples of real photographs carrying several defects at once; no DIV2K or Flickr2K image was used to train Wonder 4, while several baselines such as Real-ESRGAN include both datasets in their training data.
The authors state that the full version of the report will be released soon, so the quantitative comparison visible here is limited to the three test settings; Wonder 4 also does not have the highest PSNR, which the authors explain as the usual trade-off between added detail and PSNR.
