Existing LDR image generators can be adapted to make HDR images by encoding HDR data in a perceptually uniform space.
The problem: HDR images are stored in linear RGB, whose intensity and color statistics differ a lot from sRGB LDR images, and large-scale HDR training data is scarce.
The fix: Convert HDR inputs to perceptually uniform encodings such as PU21 or PQ. LDR-pretrained VAEs reconstruct these with fidelity comparable to LDR data, while linear RGB causes severe degradation.
The method: Freeze the VAE and finetune only the denoiser with low-rank adaptation. One approach covers both text-to-HDR synthesis and single-image RAW-to-HDR reconstruction.
The claims: Better perceptual fidelity, text-image alignment and effective dynamic range than previous techniques. It is a NeurIPS 2026 spotlight.
Sign in to comment.
TL;DR of the paper, generated automatically.
Existing LDR image generators can be adapted to make HDR images by encoding HDR data in a perceptually uniform space.
No retraining from scratch needed.