EquivFlow：可逆曝光等变编码修复 VAE 的 HDR 视频重建暗部坍缩
Dark Matters: VAE-Aware HDR Encoding for SDR-to-HDR Video Reconstruction
做视频生成或 HDR 后期的开发者该看看这篇：它指出类 log 编码会把暗部色彩在 VAE 潜空间里压没，换成 EquivFlow 这种可逆编码就能保住暗部色度，还不用动预训练 VAE，比直接改模型省事。
arXiv 论文提出 EquivFlow，一种紧凑的可逆曝光等变编码，用于解决 SDR-to-HDR 视频重建中的暗部色彩丢失问题。
作者发现常用的类 log 编码会在生成前抑制低亮度色度差异，导致冻结 VAE 潜空间出现 dark-side collapse。EquivFlow 将曝光变化映射为加性平移，同时保留尺度无关的色度分量，且不需要修改预训练 VAE。
在多个预训练图像与视频 VAE 的编码层测试中，其曝光结构和深暗区色度传递均保持稳定。集成进 SDR-to-HDR 视频扩散模型后，EquivFlow 在各测试设置下带来一致的色彩保真度提升，重建保真度保持竞争力。
Dark Matters: VAE-Aware HDR Encoding for SDR-to-HDR Video Reconstruction
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.
This report is categorized under AITOP's "Papers" desk. Within our cross-regional analytical framework, it directly maps to AI developer toolchains, coding agents, and engineering workflows. Key entities and technical topics under continuous tracking include HDR, VAE, SDR-to-HDR, 潜空间.
Source attribution: curated from arXiv cs.AI (@Donghoon Cha, Bomseok Park, Hansl von Kwon, Ku DoHyung, Hanwool Sul).