跳到正文
研踪PaperTrail
← 研踪 PaperTrail

arXiv:2611.18690 · v2

Perceptual Weight Space Programming for Diffusion Super-Resolution

首次提交 2026-11-03 19:21
版本更新 2026-11-19 13:27
雷达观测 2026-11-20 19:58
arXiv ↗PDF ↗

摘要

Diffusion models for image super-resolution can recover scene content while producing distracting artifacts or unbalanced colors. We study whether a large language model (LLM) can improve perceptual quality by editing a small subset of existing U-Net weights without a clean reference image. The proposed framework provides the editor with architecture code, selected weights, activation summaries, and sampling previews. It permits up to 1,020 candidate trials while limiting accepted changes to three layers and fewer than 0.1\% of model parameters. The evaluation protocol uses ResShift, SinSR, and SUPIR and compares GPT-6 Astra with a specified Qwen3-VL model under matched access. Feedback metrics guide editing, while separate quality metrics and blinded human comparisons evaluate the final weights. We also report an executed color calibration experiment using real photographs and two small residual networks. This experiment demonstrates correction of a fixed color shift and reveals that an improving feedback score can accompany worse performance on unseen scenes. A separate lightweight super-resolution experiment isolates modest changes in luminance detail while preserving color. The findings motivate restricted updates and independent evaluation of perceptual weight editing.

作者备注:The authors have identified issues in the experimental evaluation that may affect the reported conclusions. The authors have identified methodological issues that require further investigation before the findings can be considered reliable

AI 论文解读

基于摘要 · 非全文审阅

此版本尚未生成 AI 解读。解读由后台任务生成并按论文版本保存,登录后共享已有结果。

版本差异

本站首次收录为 v2,暂无更早版本可对照。

观测到的版本

相邻记录可能跳过中间版本;首次发现 v2 及以上也会标为版本更新。

v2 · 2026-11-19 13:27 · 最新观测版本

Perceptual Weight Space Programming for Diffusion Super-Resolution

Diffusion models for image super-resolution can recover scene content while producing distracting artifacts or unbalanced colors. We study whether a large language model (LLM) can improve perceptual quality by editing a small subset of existing U-Net weights without a clean reference image. The proposed framework provides the editor with architecture code, selected weights, activation summaries, and sampling previews. It permits up to 1,020 candidate trials while limiting accepted changes to three layers and fewer than 0.1\% of model parameters. The evaluation protocol uses ResShift, SinSR, and SUPIR and compares GPT-6 Astra with a specified Qwen3-VL model under matched access. Feedback metrics guide editing, while separate quality metrics and blinded human comparisons evaluate the final weights. We also report an executed color calibration experiment using real photographs and two small residual networks. This experiment demonstrates correction of a fixed color shift and reveals that an improving feedback score can accompany worse performance on unseen scenes. A separate lightweight super-resolution experiment isolates modest changes in luminance detail while preserving color. The findings motivate restricted updates and independent evaluation of perceptual weight editing.

查看此版本解读 →查看此版本原文 ↗