Searcharxiv⌕ Search

arXiv subjects

Jonathan Guo

Publications and source records attributed to Jonathan Guo.

2 recordsLinked to original sources

FDIR: Fidelity-Anchored Decompressed Image Restoration for Human and Machine Perception

Evaluating restored images requires considering pixel fidelity, human perception, and machine preference. However, existing compressed image restoration methods rarely satisfy all three criteria at once. Regression models preserve pixel fidelity but often produce overly smooth results. Conversely, generative models create realistic textures but tend to drift from the ground truth. To address this trade-off, we propose FDIR (Fidelity-Anchored Decompressed Image Restoration), a framework that uses generative priors while preserving strong pixel fidelity. FDIR works in two stages. First, a quality-guided one-step flow model (QO-Flow) adapts a pretrained latent flow backbone with LoRA to generate a structural prior in a single forward pass. Second, rather than outputting this generated image directly, a flow-conditioned detail refinement network (FCDR) uses it to predict a pixel-space residual on the decompressed input. This design brings in perceptual details while keeping the final output faithful to the original image. We evaluate FDIR on LIVE1 and DIV2K under severe compression at quality factors 1 and 5. Across all four settings, FDIR achieves the highest PSNR among the seven compared methods and beats the strongest perceptual baseline, CODiff, by 1.3 to 2.1 dB. It also ranks second on LPIPS in all four settings and on FID and DISTS on DIV2K-Val, right behind CODiff. For downstream detection, segmentation, and retrieval at quality factors 20 and 30, FDIR ranks second on every individual metric and achieves the best average rank across all ten metrics. Overall, FDIR is the only method to maintain a top-two average rank across fidelity, perception, and machine preference.

eess.IV↗

Conditional Gaussian Limits and Character Laws for Randomized Symmetric-Group Representations

Let $ρ_n$ be a real orthogonal representation of the symmetric group $S_n$ of dimension $d_n\to\infty$, let $σ_n$ be a uniform random permutation, and let $O_n$ be an independent Haar-orthogonal matrix. We study the linear statistics $\operatorname{Tr}(A_nO_nρ_n(σ_n)O_n^T)$ for deterministic real test matrices $A_n$ normalized only in Frobenius norm, which allows tests that isolate individual entries. Conditionally on $σ_n$, the statistic has mean $s_nχ_n(σ_n)$, where $χ_n$ is the character of $ρ_n$ and $s_n=d_n^{-1}\operatorname{Tr} A_n\to s$. We show that the centered statistic has conditional Gaussian limit $N(0,1-s^2)$ for every such test if and only if the trivial and sign constituents of $ρ_n$ occupy $o(d_n)$ dimensions, and that this limit is asymptotically independent of every convergent $σ_n$-measurable observable. The key input is a tensor-square estimate, based on the Frobenius-Schur identity and the minimum degree of non-linear irreducible representations of $S_n$, which shows that this single condition controls both normalized traces needed in the Haar-conjugation argument. For the uncentered statistic, convergence in distribution of $χ_n(σ_n)$ to a law $Ξ$ is necessary and sufficient for every test statistic to converge to $sΞ+\sqrt{1-s^2}G$, with $G$ standard Gaussian and independent of $Ξ$. For every irreducible sequence with $d_n\to\infty$, finitely many entries scaled by $\sqrt{d_n}$ have independent standard Gaussian limits. Finally, we compute $Ξ$ explicitly for diagrams with a fixed first-row tail, their transposes, and diagrams with sublinear largest hook, answering in these cases a question of Tsou.

math.PR↗