LumaFlux converts 8-bit SDR (BT.709) images and video into 10-bit HDR (PQ, BT.2020) by adapting a frozen FLUX.1-dev MM-DiT. This repository holds the trained adapter weights only: 71,315,893 parameters, 0.57% of the backbone. The frozen FLUX.1-dev backbone and SigLIP encoder are downloaded separately at load time.

Conversion is prompt-free. Inference integrates a rectified-flow bridge from the SDR latent in 8 steps, and video is handled by the same image model with no temporal layers, optical flow, or video fine-tuning.

Checkpoints

File	Training corpus	Steps	Intended use
`lumaflux-main.safetensors`	Mixed UGC + PGC (314,396 pairs from 2,092 videos)	100,000	Default. General SDR-to-HDR conversion, video, and the Luma-Eval benchmark.
`lumaflux-hdrtv1k.safetensors`	HDRTV1K training split	50,000	In-domain HDRTV1K comparisons and compression-robustness studies.

Both files carry the identical architecture and are interchangeable at load time; they differ only in training data and schedule.

Results

Luma-Eval (our common run-by-us protocol; 100 held-out frames, synthetic tone-mapping and codec degradations). lumaflux-main:

Method	PU21-PSNR ↑	PU21-SSIM ↑	ΔE_ITP ↓	HDR-VDP-3 ↑	FR-HIDRO ↓
BT.2446c inverse	23.74	0.8291	61.60	5.611	0.753
HDRTVNet++	23.03	0.8192	66.67	5.417	0.703
FMNet	22.86	0.8176	67.92	5.412	0.724
LumaFlux	24.23	0.8294	56.54	5.812	0.631

HDRTV1K (117 published test pairs, evaluated under the authors' protocol):

Model	PSNR ↑	SSIM ↑	SR-SIM ↑	ΔE_ITP ↓	HDR-VDP-3 ↑
`lumaflux-hdrtv1k` (in-domain)	33.34	0.9427	0.9941	14.69	7.962
`lumaflux-main` (zero-shot)	27.52	0.9307	0.9837	29.86	7.618

The zero-shot row is reported for transparency: the main model never sees the HDRTV1K training split, so it is not comparable to methods trained on it.

Video. Sharing one bridge-noise realization across frames and smoothing the applied tone-curve parameters with an EMA reduces signed excess PU21 flicker by 53.3%, with pooled temporal energy within 3.3% of the reference, without any temporal training.

Cost. 8 solver steps, 5.28 s per 1080p frame and 27.1 GB peak memory on a single GH200. Forty steps cost 4.7x more without improving quality.

Usage

bash

pip install "lumaflux @ git+https://github.com/shreshthsaini/LumaFlux.git"
huggingface-cli download shreshthsaini/LumaFlux lumaflux-main.safetensors --local-dir weights

bash

# Image or video in, 10-bit PQ/BT.2020 HDR out
python -m lumaflux.inference.cli \
  --config configs/model/flux_dev.yaml \
  --adapters weights/lumaflux-main.safetensors \
  --input input_sdr.mp4 --output output_hdr.mp4 --steps 8

Loading the adapters also pulls black-forest-labs/FLUX.1-dev (gated; accept its license first) and google/siglip-so400m-patch14-384. Full instructions, data preparation, training, and the evaluation suite are in the GitHub repository.

Training

Both checkpoints were trained with a frozen backbone: only PGA (gated low-rank attention residuals), PCM (SigLIP FiLM conditioning), the HDR Residual Coupler, the timestep-layer conditioner, and the monotone RQS tone head receive gradients. Optimization used AdamW at learning rate 2e-4, global batch 16, 512px crops, bf16, and a rectified-flow velocity objective with decoded-reconstruction and spline-smoothness terms. The main model trained for about 36 hours on four GH200 GPUs (~144 GPU-hours).

Training data combines CHUG user-generated HDR, LIVE-TMHDR (including expert-graded SDR), HDRTV1K, and the Netflix Open Content Sol Levante HDR10 master. SDR inputs are either genuine expert grades or generated with a frozen degradation chain of eight tone-mapping operators, BT.2020-to-BT.709 gamut conversion, and three CRF settings.

License and intended use

These adapters only function on top of FLUX.1-dev and are therefore released under the FLUX.1-dev Non-Commercial License; research use only. The accompanying source code is Apache-2.0. Users must separately accept the FLUX.1-dev license.

Intended for research on inverse tone mapping, archival SDR remastering, and HDR video pipelines. Not validated for medical, forensic, or safety-critical imaging.

Limitations

The frozen VAE was not trained as an HDR codec, so the tone head calibrates its output but cannot recover detail absent from heavily clipped SDR.
Temporal stabilization is inference-time only and does not model long-range motion explicitly.
On the selected qualitative scenes the outputs measure slightly more saturated than their references (mean ITP chroma 60.65 versus 49.80).
Correct display metadata and EOTF settings are required at playback.

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}
}