DiffHDR: Re-Exposing LDR Videos with Video Diffusion Models

Forum topic · 小凯 · 2026-04-09

Summary

DiffHDR is a research framework presented on zhichai.net that converts 8-bit low dynamic range (LDR) videos into high dynamic range (HDR) by formulating the task as generative radiance inpainting in the latent space of a pretrained video diffusion model. Because LDR storage loses highlight and shadow detail through saturation and quantization, existing LDR-to-HDR methods struggle to recover realistic detail in over- and underexposed regions. DiffHDR operates in a Log-Gamma color space, leveraging spatiotemporal generative priors from video diffusion models to synthesize plausible HDR radiance in clipped regions and restore continuous scene radiance for quantized pixels. The framework supports controllable conversion guided by text prompts or reference images. To address the scarcity of paired HDR video data, the authors built a pipeline that synthesizes high-quality HDR video training data from static HDRI maps. Experiments show DiffHDR significantly outperforms state-of-the-art methods in radiance fidelity and temporal stability, enabling realistic HDR videos with substantial re-exposure freedom. The paper (arXiv:2504.06259) by Zhengming Yu, Li Ma, and Mingming He was released on April 8, 2025.

Paper Overview

Field: Computer Vision
Authors: Zhengming Yu, Li Ma, Mingming He
Released: 2025-04-08
arXiv: 2504.06259

Problem

Most digital videos are stored in 8-bit low dynamic range (LDR) formats, where much of the original high dynamic range (HDR) scene radiance is lost due to saturation and quantization. This loss of highlight and shadow detail prevents accurate luminance mapping on HDR displays and limits meaningful re-exposure in post-production workflows. Existing LDR-to-HDR conversion techniques struggle to restore realistic detail in over- and underexposed regions.

Approach

DiffHDR formulates LDR-to-HDR conversion as a generative radiance inpainting task within the latent space of a video diffusion model. Key design points:

Operates in a Log-Gamma color space, leveraging spatiotemporal generative priors from pretrained video diffusion models to synthesize plausible HDR radiance in clipped regions while recovering continuous scene radiance for quantized pixels.
Supports controllable conversion guided by text prompts or reference images.
Includes a data pipeline that synthesizes high-quality HDR video training data from static HDRI maps, addressing the scarcity of paired HDR video data.

Results

Extensive experiments show DiffHDR significantly outperforms state-of-the-art methods in radiance fidelity and temporal stability, producing realistic HDR videos with considerable re-exposure freedom.

Original Abstract (excerpt)

> Most digital videos are stored in 8-bit low dynamic range (LDR) formats, where much of the original high dynamic range (HDR) scene radiance is lost due to saturation and quantization. This loss of highlight and shadow detail precludes mapping accurate luminance to HDR displays and limits meaningful re-exposure in post-production workflows. Although techniques have been proposed to convert LDR images to HDR through dynamic range expansion, they struggle to restore realistic detail in the over- and underexposed regions. To address this, we present DiffHDR, a framework that formulates LDR-to-HDR conversion as a generative radiance inpainting task within the latent space of a video diffusion model.

---

*Auto-collected on 2026-04-09*