LumaFlux: Lifting 8-Bit Worlds to HDR Reality with Physically-Guided Diffusion Transformers

Abstract

The rapid adoption of HDR-capable devices has created a pressing need to convert 8-bit Standard Dynamic Range content into perceptually and physically accurate 10-bit HDR. Existing inverse tone-mapping (ITM) methods often rely on fixed tone-mapping operators that struggle to generalize to real-world degradations, stylistic variations, and camera pipelines — producing clipped highlights, desaturated colors, or unstable tone reproduction.

LumaFlux is the first physically and perceptually guided diffusion transformer for SDR→HDR reconstruction, built by adapting a large pretrained DiT without retraining its weights. It introduces (1) a Physically-Guided Adaptation module injecting luminance, spatial, and frequency cues into attention through gated low-rank residuals; (2) a Perceptual Cross-Modulation layer stabilizing chroma and texture via FiLM conditioning on SigLIP features; (3) an HDR Residual Coupler fusing both signals under a timestep- and layer-adaptive schedule; and (4) a Rational-Quadratic Spline decoder reconstructing smooth, interpretable tone fields for highlight and exposure expansion. We further curate the first large-scale SDR–HDR training corpus and establish Luma-Eval, a new evaluation benchmark with HDR references and expert-graded SDR.

Method

SDR formation is a lossy chain — tone mapping, BT.2020→709 gamut compression, quantization, and codec noise. Inverting it needs both physical priors (luminance and color constraints) and perceptual priors (semantics and realism). LumaFlux gets the latter for free from a frozen Flux backbone and injects the former through four lightweight, zero-initialized modules scheduled by a shared timestep–layer conditioner Ψ(t, ℓ): early layers and timesteps receive strong global tone gains, late ones refine highlight detail.

Architecture. The SDR input splits into physical (T_phys) and perceptual (T_perc) streams. In each Luma-MMDiT block, PGA injects luminance- and spectrum-aware low-rank updates into attention, PCM applies FiLM to normalized features, and the HDR Residual Coupler fuses both cues. A frozen VAE decoder with the RQS tone-field head reconstructs HDR in PQ/BT.2020.

Architectural paradigms. Left: direct LoRA/full fine-tuning of a DiT overfits small HDR datasets and hallucinates texture. Right: LumaFlux inserts lightweight, physically interpretable modules into the frozen MM-DiT, preserving the pretrained generative prior while enabling accurate ITM with few trainable parameters.

Dataset & Benchmark

We unify HIDROVQA (411 PGC videos), CHUG (428 UGC videos), and LIVE-TMHDR (40 studio HDR videos with expert-graded SDR) under PQ-encoded BT.2020 at a 1,000-nit mastering peak. Each HDR frame is paired with SDR variants from a composite degradation chain: 8 tone-mapping operators × CRF {23, 31, 39} compression — ≈318k pairs including 54k expert-graded references. Luma-Eval adds 20 held-out HDR sources (10 PGC + 10 UGC) evaluated under both perceptually optimized and degradation-heavy SDR conditions.

Results

Metrics in PU21 space for luminance (PSNR, SSIM), HDR domain (HDR-VDP3), and perceptual color (ΔE_ITP, lower is better). Best in blue.

Quantitative comparison across benchmarks (Table 1 of the paper).

Component ablations on Luma-Eval (Table 3 of the paper): PGA delivers the largest gain; the monotone RQS restores perceptual smoothness and broadens effective dynamic range.

BibTeX

@article{saini2026lumaflux,
  title   = {LumaFlux: Lifting 8-Bit Worlds to HDR Reality with
             Physically-Guided Diffusion Transformers},
  author  = {Saini, Shreshth and Gedik, Hakan and Birkbeck, Neil and
             Wang, Yilin and Adsumilli, Balu and Bovik, Alan C.},
  journal = {arXiv preprint arXiv:2604.02787},
  year    = {2026}
}