HDR Video Generation via Latent Alignment with Logarithmic Encoding

Forum topic · 小凯 · 2026-04-15

Summary

A new paper on arXiv (2604.11788) presents a simple approach for high dynamic range (HDR) image and video generation with pretrained generative models. HDR imagery faithfully represents scene radiance but mismatches the bounded, perceptually compressed data generative models are trained on. The authors show that logarithmic encoding maps HDR images to a distribution that naturally aligns with a pretrained model's latent space, enabling direct adaptation via lightweight fine-tuning without retraining the encoder. To recover details not directly observable in inputs, they introduce a training strategy based on camera-simulated degradations, encouraging the model to infer missing high-dynamic-range content from its learned prior. The key finding: generative models can effectively handle HDR as long as the representation is chosen to align with their learned prior, avoiding the complexity and data requirements of learning entirely new HDR representations.

Research area: cs.CV Authors: Naomi Ken Korem, Mohamed Oumoumad, Harel Cain, Matan Ben Yosef, Urska Jelercic, Ofir Bibi, Yaron Inger, Or Patashnik, Daniel Cohen-Or Published: 2026-04-13 arXiv: 2604.11788

Abstract

High dynamic range (HDR) imagery offers a rich and faithful representation of scene radiance, but remains challenging for generative models due to its mismatch with the bounded, perceptually compressed data on which these models are trained. A natural solution is to learn new representations for HDR, which introduces additional complexity and data requirements.

This paper demonstrates a simpler approach to HDR generation: using logarithmic encoding to map HDR images to a distribution that naturally aligns with the latent space of pretrained generative models. This enables direct adaptation through lightweight fine-tuning, without the need to retrain encoders. To recover details not directly observable in the input, the authors further introduce a training strategy based on camera-simulated degradations, which encourages the model to infer missing high-dynamic-range content from its learned prior.

The results show that generative models can effectively handle HDR, as long as a representation aligned with their learned prior is chosen.

---

*Auto-collected on 2026-04-15.*

HDR Video Generation via Latent Alignment with Logarithmic Encoding

Summary

Abstract

Tags