arXiv cs.AI· Donghoon Cha, Bomseok Park, Hansl von Kwon, Ku DoHyung, Hanwool Sul·· 10 小时前AI 评分27
EquivFlow：面向 SDR-to-HDR 视频重建的 VAE 感知 HDR 编码
Dark Matters: VAE-Aware HDR Encoding for SDR-to-HDR Video Reconstruction
AI 导读
研究者提出 EquivFlow，一种紧凑、可逆的曝光等变编码，将曝光变化映射为加性偏移并保留尺度无关的色度分量，且无需修改 VAE。该工作发现常见对数类编码会在生成前压制低亮度色差，导致冻结 VAE 潜空间出现暗部塌陷。
正文 · 原文
Abstract:SDR-to-HDR reconstruction is an ill-posed problem in which clipped highlights, crushed shadows, and quantization artifacts must be recovered while preserving the original content. Latent diffusion models provide powerful generative priors for restoring missing detail, but their effectiveness is constrained by how HDR signals pass through the pretrained VAE. We find that common log-like encodings suppress low-luminance chromatic differences before generation, producing a dark-side collapse in the frozen-VAE latent. We introduce EquivFlow, a compact, invertible exposure-equivariant encoding that maps exposure changes to additive shifts while retaining a scale-independent chromatic component, without modifying the VAE. Encoding-level tests across pretrained image and video VAEs show that its exposure structure and deep-dark chroma transport persist across the tested hosts. Integrated into SDR-to-HDR video diffusion, EquivFlow delivers consistent chromatic-fidelity gains across the tested settings while retaining competitive reconstruction fidelity.
| Comments: | 13 pages, 4 figures |
| Subjects: | Computer Vision and Pattern Recognition (cs.CV); Artificial Intelligence (cs.AI) |
| Cite as: | arXiv:2611.13504 [cs.CV] |
| (or arXiv:2611.13504v1 [cs.CV] for this version) | |
| https://doi.org/10.48550/arXiv.2611.13504 arXiv-issued DOI via DataCite (pending registration) |
Submission history
From: Donghoon Cha [view email]
[v1]
Tue, 10 Nov 2026 03:16:31 UTC (2,728 KB)
来源：arXiv cs.AI · arxiv.org