Sr Diffusion

LuSIR is a vision-only x4 super-resolution research project trained without a pretrained text-to-image diffusion model. GitHub: The repository stores selected research checkpoints, configs, metrics, and sample grids. It does not redistribute training datasets. ## Current Selected Detail Artifact The latest public stable detail-branch checkpoint remains: It is a deterministic 3.02M-parameter image-space detail branch on top of the frozen dual-context LSDIR Stage 2 step 98000 condition encoder and frozen Stage 1 decoder. The run completed micro-steps, exactly three epochs, and selected step by . Selected ordinary val100 result: Exploratory strict-bicubic DIV2K five-center-crop result: The strict-bicubic result is not a formal SOTA benchmark. It uses five 512x512 center crops, PIL bicubic x4 degradation, full-image RGB PSNR, and no border shave. Formal full-image clean-bicubic benchmark, reported as Y PSNR / Y SSIM: | Dataset | Dual-context base | Detail v1d | | --- | ---: | ---: | | DIV2K validation | 29.9575 / 0.82887 | | | Set5 | 31.6621 / 0.88952 | | | Set14 | 28.2441 / 0.77340 | | | Urban100 | 25.4816 / 0.76473 | | This uses public x4 LR pairs, MATLAB-compatible BT.601 Y, a four-pixel border shave, and MATLAB-style SSIM. V1d improves its frozen base on all four datasets. These clean-bicubic fidelity results are not a claim of classical-SR SOTA or a substitute for real-degradation and perceptual evaluation. For scale, the official SwinIR classical x4 checkpoint reaches on the same DIV2K evaluator, Y PSNR ahead of detail v1d. The next clean-fidelity priority is therefore the Stage 2/base reconstruction path rather than a larger detail branch. A clean-bicubic Stage 2 continuation improved its task-specific val100 proxy only gradually and plateaued around . Learning-rate probes did not change that conclusion: LR collapsed, while a from-init run matched the original LR within evaluation noise. These val100 values are not directly comparable with the formal full-image Y-channel benchmark above. The signed-high-frequency residual diffusion path was evaluated and rejected: longer noise-MSE training collapsed residual magnitude and seed diversity toward zero. The next separate generative research path keeps the deterministic base and validated learned mask frozen, then tests a small bounded mask-weighted patch perceptual/adversarial head with fidelity and artifact guardrails. ## Latest Masked Detail Research Candidate The learned-mask-gated v2 candidate is: It combines the frozen 460K-parameter detail-mask predictor step 3250 with the 3.02M-parameter detail branch and a soft-mask floor of . On ordinary val100, selected step 38000 improves the frozen base by aggregate PSNR, mean PSNR, and SSIM, with wins. The score plateaued after step 38000 and fixed grids were nearly indistinguishable from nearby checkpoints. It modestly improves metrics over v1d but does not visibly recover the missing fine texture that motivated the experiment. It is therefore a reproducible research option, not the public default. On the same formal 219-image clean-bicubic benchmark, masked v2 reaches , , , and on DIV2K, Set5, Set14, and Urban100. It improves v1d on all four datasets, but the overall gain is only Y PSNR and Y SSIM. ## Download From a LuSIR GitHub clone: Other useful presets include: The public Colab default is now the T4-friendly guarded-detail Stage 2 v2 checkpoint with tile batch size 1. Residual refiner v2 remains available as the conservative deterministic option. Detail v1d and masked detail v2 remain research options in the Colab WebUI with single-image and tiled inference. ## Runtime Paths Stage numbers describe training order. Stage 3 and Stage 4 are alternative diffusion checkpoints, not modules executed sequentially. ## License - Checkpoints, generated samples, metrics, and other non-code artifacts: CC BY-NC 4.0. - Source code: PolyForm Noncommercial License 1.0.0. Commercial use is not permitted without separate written permission.