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

LumaFlux teaser showing SDR inputs and corresponding HDR results — LumaFlux converts 8-bit SDR (BT.709) images and video into 10-bit HDR (PQ, BT.2020).

Overview

Abstract

LumaFlux converts 8-bit SDR content in BT.709 into 10-bit HDR in PQ and BT.2020. It adapts a frozen FLUX.1-dev MM-DiT using physical image cues and frozen perceptual features, while the backbone, VAE, and SigLIP encoder remain frozen. The resulting model is prompt-free and uses 71.3M trainable parameters, 0.57% of the backbone.

Physically-Guided Adaptation injects luminance, gradient, saturation, and spectral-band cues through gated low-rank attention residuals. Perceptual Cross-Modulation supplies FiLM conditioning from SigLIP, and the HDR Residual Coupler combines both paths under timestep-and-layer modulation. A monotone RQS tone-field decoder calibrates the frozen VAE output into display-referred HDR luminance. Inference follows an 8-step rectified-flow bridge with no text prompt or sampling-strength knob. For video, the same image model uses shared bridge noise and an EMA over applied RQS parameters, with no temporal training.

8solver steps

71.3Mtrainable parameters

0.57%of the backbone

Prompt-freeSDR to HDR bridge

Approach

Method

LumaFlux keeps the FLUX.1-dev MM-DiT and VAE frozen, and also keeps the SigLIP encoder frozen. It learns a compact physical and perceptual adaptation around that backbone, then calibrates the decoded result into HDR luminance.

Overview of the LumaFlux architecture, including physical and perceptual adaptation paths and the HDR decoder — LumaFlux combines physical cues, frozen SigLIP features, a frozen FLUX.1-dev backbone, and an RQS tone-field decoder.

PGA

Physically-Guided Adaptation

Gated low-rank attention residuals receive luminance, gradient, saturation, and spectral-band cues.

PCM

Perceptual Cross-Modulation

FiLM conditioning carries perceptual features from a frozen SigLIP encoder into the adaptation.

COUPLER

HDR Residual Coupler

The coupler fuses the physical and perceptual paths under timestep-and-layer modulation.

RQS

RQS tone-field decoder

A monotone rational-quadratic spline calibrates the frozen VAE decode into display-referred HDR luminance.

Comparison of SDR to HDR adaptation paradigms and the LumaFlux physically-guided design — The LumaFlux paradigm adapts a frozen generative backbone through physically-guided and perceptual paths.

Prompt-free bridge. Inference starts at z₁ = VAE_enc(SDR) + 0.05·ε and integrates from t = 1 to t = 0 in K = 8 steps. There is no text prompt and no sampling-strength knob.

Training-free video stabilization. Video uses the same image model with one shared bridge-noise realization across frames and an EMA on the applied RQS parameters. It adds no temporal layers, optical flow, video fine-tuning, or temporal loss.

Evaluation

Results

Luma-Eval covers synthetic tone-mapping and codec degradations, expert-graded SDR from LIVE-TMHDR, and native paired SDR/HDR. Every method is scored on identical inputs, output encoding, and metric implementations.

Luma-Eval

Run-by-us synthetic-track protocol on 100 held-out frames. Higher is better for PU21-PSNR, PU21-SSIM, and HDR-VDP-3. Lower is better for dE-ITP, HDR-LPIPS, and FR-HIDRO.

Luma-Eval comparison
Method	PU21-PSNR ↑	PU21-SSIM ↑	dE-ITP ↓	HDR-LPIPS ↓	HDR-VDP-3 ↑	FR-HIDRO ↓
BT.2446c inverse	23.74	0.8291	61.60	0.307	5.611	0.753
Reinhard inverse	23.62	0.8286	61.76	0.295	5.726	0.753
HDRTVNet++	23.03	0.8192	66.67	0.328	5.417	0.703
FMNet	22.86	0.8176	67.92	0.330	5.412	0.724
KUNet	22.27	0.8111	70.40	0.325	5.445	0.693
ITM-LUT	22.92	0.8142	67.41	0.315	5.496	0.761
VAE+RQS (no-diffusion control)	18.54	0.7647	111.67	0.381	3.673	0.924
LumaFlux	24.23	0.8294	56.54	0.269	5.812	0.631

LumaFlux is best in every column. Its margin over the strongest baseline is +0.49 dB in PU21-PSNR and -5.06 in dE-ITP. FR-HIDROVQA falls from 0.703 for HDRTVNet++ to 0.631.

HDRTV1K

Results on 117 published test pairs using the authors' protocol. The in-domain checkpoint and zero-shot main checkpoint are distinct.

In-domain versus zero-shot HDRTV1K evaluation
Model	PSNR	SSIM	SR-SIM	dE-ITP	HDR-VDP-3
LumaFlux in-domain (`lumaflux-hdrtv1k`)	33.34	0.9427	0.9941	14.69	7.962
LumaFlux zero-shot (`lumaflux-main`)	27.52	0.9307	0.9837	29.86	7.618

The 33.34 dB result comes from the HDRTV1K-trained checkpoint. The mixed-corpus main model was never trained on HDRTV1K, so its 27.52 dB result is zero-shot. Published baseline numbers on HDRTV1K retain their authors' own protocols and are not directly rankable against these results.

Generative-class comparison

Like-for-like native-track comparison on 117 pairs.

Generative inverse tone-mapping methods
Method	Steps	PU21-PSNR	dE-ITP	s/megapixel	Trainable params
LEDiff	50	10.99	198.58	17.0	860 M
X2HDR	30	18.65	92.20	16.2	149 M
LumaFlux	8	23.56	47.59	2.5	71 M

Against the closest diffusion competitor, LumaFlux gains +4.9 dB, costs about 6x less per output pixel, uses 2x fewer trainable parameters, and requires fewer steps.

53.3%

Less signed excess flicker

Shared noise and RQS-parameter EMA reduce signed excess PU21 flicker by 53.3%. Pooled raw temporal energy remains within 3.3% of the reference, with zero temporal training.

0.55 dB

Compression robustness

Across the HDRTV1K fixed QP 27 to 42 sweep, the mixed-corpus LumaFlux model loses 0.55 dB. HDRTVNet++ loses 2.78 dB and FMNet loses 2.71 dB.

Release

Getting the weights

Download the adapter checkpoints from the LumaFlux Hugging Face repository. Each file contains adapters only. The frozen FLUX.1-dev backbone and SigLIP are downloaded separately.

lumaflux-main.safetensors

Main checkpoint

Trained for 100k steps on the mixed UGC+PGC corpus of 314,396 pairs. Use it by default, for video, and for Luma-Eval.

lumaflux-hdrtv1k.safetensors

HDRTV1K checkpoint

Trained for 50k steps on the HDRTV1K train split. Use it for in-domain HDRTV1K comparisons.

Two-command quickstart

huggingface-cli download shreshthsaini/LumaFlux lumaflux-main.safetensors --local-dir .
python -m lumaflux.inference.cli --config configs/model/flux_dev.yaml --adapters lumaflux-main.safetensors --input input_sdr.mp4 --output output_hdr.mp4 --num-steps 8

Gated backbone: FLUX.1-dev is gated on Hugging Face. Accept its license separately before running inference.

Scope

Limitations

The frozen VAE is not an HDR codec. The tone head calibrates its output but cannot invent detail absent from heavily clipped SDR.
Temporal stabilization is inference-time only, with no explicit long-range motion model.
Qualitative scenes show slight oversaturation: mean ITP chroma is 60.65 for LumaFlux versus 49.80 for the references.
Inference is slower than feed-forward CNN converters, at 5.28 s versus 0.28 to 0.86 s per 1080p frame, though it is cheaper per pixel than other diffusion ITM methods.
HDRTV4K ground truth could not be obtained through an open channel, so that benchmark is absent.
A controlled subjective study is future work.

Citation

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}
}