Sources, credits and test scope
Saved results from an independent proof of concept on October 22, 2026. The photograph, oil painting and abstract painting were cropped to 2048 × 2048 pixels from the distributed JPEGs without upscaling. The watercolor uses an existing 1254 × 1254 test image. These are not claimed to be lossless scans of the original works.
- Photograph: R. Hagerty/USFWS, Monarch Butterfly. The official distribution page identifies it as Public Domain.
- Oil painting: Vincent van Gogh, Irises, 1890. The Metropolitan Museum of Art Open Access, Public Domain / CC0.
- Abstract painting: Piet Mondrian, Composition with Red, Yellow, and Blue, 1927. Cleveland Museum of Art, CC0.
- Watercolor: existing test material (birds and plants wallpaper, 1254 × 1254). The museums’ CC0 terms are not applied to this image.
Methods compared
- CODiff: author weights and CaVE, fp16, seed 20261022, no VAE or latent tiling. The SD2.1 base uses a community mirror pinned to a fixed revision. AdaIN uses the degraded input as its color reference. Byte identity with the original distribution and full reproduction of the paper are not claimed.
- SwinIR colorCAR: RGB float32, tile 504 / overlap 28. Separate Q10/20/40 weights; Q60 input uses Q40 weights as an extrapolation test. Clean controls also use Q40 weights.
- WaveMamba-MAP: port with a parallel scan, tile 512 / overlap 64, with tile-local frequency maps and scans. Exact parity with the official CUDA kernel and reproduction of its speed were not measured. Licensing of the derived code and weights remains unconfirmed. Neither models nor inference code are distributed here.
Findings and limitations
JPEG artifacts were partly reduced, but the criteria for automatic acceptance with faithful preservation of fine hairs, individual brushstrokes and signatures were not met. Sufficient banding removal was not demonstrated for WaveMamba under these synthetic quantization conditions.
JPEG Q10/20/40/60 are synthetic recompressions with 4:2:0 subsampling. RGB5/RGB4 are synthetic quantization without dithering. Clean controls were also processed. There is one source image per material type. This does not measure performance on customer images, print proofs, artist acceptance, unknown-quality estimation, labor savings or a general success rate. The original was not supplied as an additional model reference or used for output post-processing; clean controls use the unchanged original as their declared input.
Auxiliary metrics were calculated on saved RGB8 pixels in native coordinates. LPIPS uses the AlexNet 0.1 family on limited regions and does not establish the correctness of signatures or individual brushstrokes. APSNR and DBI were not measured. Passing metric and software checks does not establish acceptable visual quality.
The English and Japanese pages share identical image files and metric values. Saved PNG sheets retain their Japanese headers. Their columns, from left to right, are Original / Degraded input (or clean control) / Restoration output. Full views are reduced display previews; detail views use native 100% or nearest-neighbor 200% display.
This repository contains only static HTML, images and public comparison records. It excludes inference code, weights, environments and internal logs. It does not grant a blanket license covering all images or pretrained models.