UT Austin LIVEYouTube

LumaFlux: Structure-Aware SDR-to-HDR Conversion with Monotone Luminance Flows

A frozen flow transformer that reads measurements of the SDR frame in place of a text prompt

Shreshth Saini, Hakan Gedik, Neil Birkbeck, Yilin Wang, Balu Adsumilli, Alan C. Bovik

The University of Texas at Austin · Google / YouTube

Paper: arXiv:2604.02787

November 2025

What the SDR pipeline throws away

Three frames, each split between HDR and SDR

Each frame is split, HDR upper left and SDR lower right. Shadow detail, sky color and specular highlights go missing in SDR.

  • Above 100 nit, light clips to white.
  • BT.2020 color squeezes into BT.709, and 10 bits become 8.
  • Many HDR frames share one SDR frame, so the inverse has to invent.
xsdr = Γ709OETF( M709 · ΓPQEOTF(xhdr) / Lmax ) + ε

The forward map decodes PQ (the perceptual quantizer curve), applies the BT.2020 to BT.709 gamut matrix M709, encodes with the 709 curve, then quantizes and compresses (ε).

LumaFluxarXiv:2604.02787

SDR input next to LumaFlux output

Two scenes, SDR input against LumaFlux output, with enlarged crops

Left of each white line is the SDR input, right is LumaFlux; the boxed regions are enlarged at right. HDR shown tone-mapped.

LumaFluxarXiv:2604.02787

Where earlier methods fall short

Analytic operators

Reinhard, BT.2446

One curve for every scene. Clipped highlights stay clipped.

Learned converters

HDRTVNet, FMNet, KUNet

Excellent in domain. They drift when the tone mapping changes.

Generative priors

LEDiff, X2HDR

A real structure prior, but color and tone drift.

We wanted that prior, pinned to what the SDR frame measures.

LumaFluxarXiv:2604.02787

Part 1 of 2

How LumaFlux works

What has to change in a frozen SDR image model before it can output calibrated HDR?

LumaFlux starts the flow at the SDR frame

Rectified flow paths

Rectified flow straightens the paths between paired samples, so a few Euler steps follow them. Figure from Liu et al., Flow Straight and Fast.

z1 = zSDR + 0.05 ε

The encoded SDR frame plus a small perturbation, ε ~ N(0, I).

  • Train on the straight path from the HDR latent to z1.
  • Sample in 8 Euler steps, with no strength knob.
  • 40 steps cost 4.7× the time and calibrate no better.
LumaFluxarXiv:2604.02787

LumaFlux in one picture

Direct MM-DiT adaptation compared with LumaFlux

Left: adapting the MM-DiT directly, driven by a text prompt. Right: LumaFlux, with the MM-DiT and VAE frozen.

  • Frozen FLUX.1-dev MM-DiT, VAE and SigLIP. No text encoder.
  • PGA, Physically-Guided Adaptation, a LoRA gated by measurements.
  • PCM, Perceptual Cross-Modulation, and a residual coupler add scene context.
  • RQS, a monotone rational-quadratic spline, calibrates luma.
0.57%of 12.48B weights trained, 71.3M parameters
LumaFluxarXiv:2604.02787

Measurements of the SDR frame steer every frozen block

LumaFlux architecture

All 57 MM-DiT blocks are wrapped (19 dual-stream, 38 single-stream, d = 3072, 24 heads). Flames mark the trainable parts.

Per token
luminance, gradient, saturation

Per frame
luminance statistics, 8 FFT bands

Per block and step
six gains from the conditioner Ψ(t, l)

LumaFluxarXiv:2604.02787

A monotone spline turns the VAE output into HDR luma

Ŷ = RQS(Yout; ξ, η, s)
  • A rational-quadratic spline (RQS) with 8 bins, 25 parameters per frame.
  • Widths, heights and slopes are positive, so luma keeps its ordering.
  • Removing it costs 9.69 dB. A linear spline costs 0.45 dB.
decoded luma Yout (PQ code, normalized to 1,000 nit) calibrated luma Ŷ identity knots ξ, η, s

8 widths ξ, 8 heights η and 9 knot slopes s, from a 21.9K-parameter head that starts at the identity (dashed).

LumaFluxarXiv:2604.02787

Part 2 of 2

Data, benchmark, results

Does it hold up when every method gets the same inputs and the same scoring code?

A training corpus of user-generated and professional HDR

314,396SDR-HDR pairs
2,092source videos
171,384 / 143,012UGC / PGC examples
HDR sourceWhat it adds
CHUG (Saini et al., ICIP 2025)856 user-generated references, 846 for training and 10 reserved
LIVE-TMHDR40 references with their 120 expert-graded SDR clips
HDRTV1K1,235 training pairs
Sol LevanteHDR10 master, Netflix Open Content (CC BY)
  • Missing SDR sides are synthesized with eight tone-mapping operators at three compression levels.
  • A manifest regenerates everything in about four CPU hours.
Corpus diversity

(a) PCA of ten frame statistics. The corpus hull holds all 100 LIVE-TMHDR frames and 116 of the 117 HDRTV1K test frames. (b) CIE xy footprint.

LumaFluxarXiv:2604.02787

LumaEval scores every method the same way

TrackPairsWhat is scoredCheckpoint
Synthetic (primary)10020 held-out CHUG and LIVE-TMHDR contents, eight tone-mapping variants, 78 video-variant groupsmixed-corpus
Expert100LIVE-TMHDR pairs whose SDR is an expert grademixed-corpus
Native117HDRTV1K test pairsmixed-corpus
Legacy HDRTV1K117same pairs, five-metric suite, published baseline valuesHDRTV1K-specific
Fixed-QP468intra-only HEVC on the SDR input at QP 27, 32, 37, 42both
Temporal5 clipsBonfire and four CHUG clips, PU21 flicker energymixed-corpus
Text-to-HDR6 clipsLTX-Video clips, 121 frames each, intrinsic metrics onlymixed-corpus
Outputs are PNG16 PQ BT.2020 at 1,000 nit. Every row is scored with identical inputs, outputs and metric code.

Degradations

Eight tone-mapping variants, from BT.2390 to YouTube LogC.

Metrics

PU21-PSNR and SSIM, ΔEITP color error, HDR-LPIPS, HDR-VDP-3, FR-HIDROVQA.

FR-HIDROVQA is a full-reference distance on HIDRO-VQA features (Saini et al., WACV 2024).

Baselines

Two analytic operators, four CNNs, PromptIR, LEDiff, X2HDR, and a VAE + RQS control.

LumaFluxarXiv:2604.02787

Best in every column of the synthetic track

MethodPU21-PSNR ↑PU21-SSIM ↑ΔEITP ↓HDR-LPIPS ↓HDR-VDP-3 ↑FR-HIDRO ↓
BT.2446c23.740.829161.600.3075.6110.753
Reinhard operator23.620.828661.760.2955.7260.753
HDRTVNet++23.030.819266.670.3285.4170.703
FMNet22.860.817667.920.3305.4120.724
ITM-LUT22.920.814267.410.3155.4960.761
VAE + RQS (control)18.540.7647111.670.3813.6730.924
LumaFlux (ours)24.230.829456.540.2695.8120.631
Macro-averages over the same 100 frame IDs and 78 video-variant groups. PU21-PSNR-Y is in the backup copy. Bold best, underline second best. FR-HIDRO is the HIDRO-VQA feature distance (Saini et al., WACV 2024).
24.23 dBPU21-PSNR, 0.49 dB above the strongest baseline
56.54ΔEITP, 5.06 lower than the strongest baseline
0.631FR-HIDROVQA, from 0.703 for HDRTVNet++

One checkpoint, no per-track tuning. Without diffusion transport the control collapses to 18.54 dB.

LumaFluxarXiv:2604.02787

On human-graded SDR, LumaFlux trails only on ΔEITP and HDR-VDP-3

MethodPU21-PSNR ↑PU21-SSIM ↑ΔEITP ↓HDR-LPIPS ↓HDR-VDP-3 ↑FR-HIDRO ↓
BT.2446c33.160.844237.830.1487.3970.660
Reinhard operator31.480.827043.590.1557.1730.662
HDRTVNet++31.530.812544.920.1477.1440.671
ITM-LUT31.460.808545.490.1427.0900.639
LEDiff18.300.6480103.150.4964.7230.898
X2HDR23.780.757851.280.3665.5320.758
LumaFlux (ours)33.310.848040.190.1386.9190.544
The 100 LIVE-TMHDR pairs whose SDR side is an expert grade, with the preregistered per-content exposure calibration. Bold best, underline second best. The full table, with every baseline and PU21-PSNR-Y, is in the backup.

FR-HIDROVQA 0.544, 14.9% below the best baseline. BT.2446c keeps the lower color error, 37.83 against 40.19, and the higher HDR-VDP-3.

LumaFluxarXiv:2604.02787

Three scenes, every method

The widest gap is on Sunset, 28.22 dB against 24.67 for BT.2446c.

Per-method crops for Nature, Monument and Sunset

Full-frame PU21-PSNR above each crop, best in bold. LumaFlux 30.83 dB on Nature (BT.2446c 30.70) and 26.13 on Monument (HDRTVNet++ 25.68). LEDiff (7.90, Nature) and X2HDR (14.36, Monument) are lowest.

LumaFluxarXiv:2604.02787

One region of one frame, every method

The same tower crop for the input, eight methods and the reference

Tower skyline, BT.2390 tone mapping at CRF 23. Same crop for every method, LumaFlux outlined. HDR shown tone-mapped.

LumaFlux keeps the facade structure and brings back the sky the curves and CNNs wash out.

LumaFluxarXiv:2604.02787

Where the color error sits

Per-pixel Delta E ITP maps for four methods on two scenes

Each panel splits into the preview, upper left, and per-pixel ΔEITP, lower right, on one shared scale. Darker is lower error.

  • Night city: the other three light up across sky and towers; LumaFlux stays dark.
  • Flame: LumaFlux still overshoots chroma in the core.
LumaFluxarXiv:2604.02787

Color volume on one city frame

Rendered frames and CIE xyY volumes for input, BT.2446c, HDRTVNet++, LumaFlux and reference

Frames above, CIE xyY point clouds below. Height is PQ-coded luminance; triangles are BT.709 and BT.2020. HDR shown tone-mapped.

BT.2446c and HDRTVNet++ under-fill the volume; LumaFlux recovers most of what the reference fills.

LumaFluxarXiv:2604.02787

Luminance along one row tracks the reference

City towers preview with a sampled row, and luminance along that row for five curves

City towers. The red line marks the sampled row; the plot shows luminance along it on a log nit axis.

0.058LumaFlux, mean absolute log deviation
0.166HDRTVNet++
0.190BT.2446c
LumaFluxarXiv:2604.02787

Removing the spline costs 9.69 dB

RowPGAGatePCMCoup.T-modRQSPSNR (Δ vs g)ΔEITP (Δ vs g)
(c)•••28.57 (−0.62)33.48 (+5.66)
(d)••••29.56 (+0.37)28.38 (+0.57)
(f) seed-0 repeat of (g)••••••28.67 (−0.52)35.00 (+7.18)
(g) full design, ours••••••29.19 (ref.)27.81 (ref.)
(h) linear spline•••••lin.28.74 (−0.45)32.16 (+4.35)
(i) no RQS•••••19.50 (−9.69)261.78 (+233.97)
(j) no T-mod•••••28.38 (−0.81)33.90 (+6.09)
Component ablations: 20k steps on a 55-video pool, the same 82 held-out frames. A dot marks a component that is on; Gate is the spectral gate, T-mod the timestep-and-layer modulation, lin. a linear spline. Higher PSNR and lower ΔEITP are better; deltas are relative to row (g). Rows (a), (b) and (e) are in the backup.
  • No RQS, row (i), falls 9.69 dB.
  • A linear spline, row (h), costs 0.45 dB.
  • The spectral gate, (c) to (d), lifts 28.57 to 29.56 dB.
  • No timestep-layer modulation, row (j), costs 0.81 dB.

Rows (f) and (g) are the same configuration, 0.52 dB apart, so smaller gaps are noise.

LumaFluxarXiv:2604.02787

Heavier compression barely moves LumaFlux

PSNR change from QP 27

PU21-PSNR change from QP 27 (QP is the HEVC quantization parameter) on the 117 HDRTV1K pairs. Flatter is better.

PU21-PSNR (dB)QP 27QP 32QP 37QP 42Δ
HDRTVNet++35.9035.3034.3433.13−2.77
FMNet35.5634.9734.0432.85−2.71
LumaFlux (mixed-corpus)27.2927.2727.0426.75−0.55
LumaFlux (in-domain)33.9433.6733.1232.33−1.61
Fixed-QP track, 117 HDRTV1K pairs. Δ is QP 42 minus QP 27; closer to zero is better. Mixed-corpus is the main checkpoint; in-domain is the HDRTV1K-specific one. Bold best, underline second per column. ΔEITP and HDR-VDP-3 are in the backup.

From QP 27 to 42 the mixed-corpus model loses 0.55 dB. HDRTVNet++ loses 2.77 and FMNet 2.71.

The CNNs still score higher at every QP on their home set.

LumaFluxarXiv:2604.02787

What it costs

5.28 sper 1080p frame, 8 steps, one GH200
27.06 GBpeak memory at 1080p
71.32Mtrainable of 12.48B parameters
Multi-resolution frontier

FR-HIDROVQA (lower is better) on 30 expert-track frames. LumaFlux 0.431, 0.477 and 0.567 at 540p, 720p and 1080p, in 2.64, 3.08 and 5.29 s.

MethodStepsParams (M)Trainable (M)s/frameTFLOPsMem. (GB)
HDRTVNet++10.590.590.8570.52.31
ITM-LUT*10.600.600.367–0.18
PromptIR + Reinhard135.5935.591.70810.917.86
LEDiff (5122)501370.23859.534.4491496.37
X2HDR (5122)3017020.02149.424.2441,02333.77
LumaFlux, 8 steps812484.7771.325.2822,06727.06
LumaFlux, 40 steps4012484.7771.3224.99310,10027.06
One GH200, batch 1, complete 1080p exports; LEDiff and X2HDR at their registered 5122 protocol (scaled by pixel count to 1080p: 35.19 and 33.57 s). *ITM-LUT runs a compiled kernel the tracer cannot observe. Lower is better for s/frame, TFLOPs and memory. Full table in the backup.

2.4 to 16.1× faster than X2HDR and LEDiff. HDRTVNet++ stays 173 to 201× faster than us.

LumaFluxarXiv:2604.02787

Video without temporal training

Per-frame mean PU21 luma

Per-frame mean PU21 luma, four clips, first 120 frames, relative to each reference mean. Only the stabilized run is drawn.

MethodBonfire02c3055e03920237Mean
BT.2446c7.8131.4110.4500.0630.1241.972
HDRTVNet++3.0050.2070.1470.4421.2571.012
FMNet3.8530.2020.0740.5791.1241.166
LumaFlux1.9920.6912.5000.9680.5691.344
LumaFlux + stabilization2.1090.4691.6760.8790.3621.099
Temporal-energy deviation |flickerpred − flickerref| in 106 PU21-energy units on five clips (02c3 to 0237 are CHUG content IDs). Smaller is closer to the reference; bold marks the closest method per column. The reference row is zero by definition.
  • Reuse one bridge noise ε for every frame.
  • Smooth the spline over time, θ̄f = 0.8 θ̄f−1 + 0.2 θf.
  • Signed excess flicker falls 53.3%. Raw flicker stays within 3.27% of the reference.

Nothing is trained for video, so there are no temporal layers, optical flow or fine-tuning.

LumaFluxarXiv:2604.02787

Lifting generated video to HDR

Two generated SDR frames and their lifted luminance maps

Two of six LTX-Video clips. Right column: gray is at or below 100 nit, color is above it.

  • Text-to-video gives SDR; LumaFlux lifts it frame by frame.
  • Peaks of 739 to 957 nit, 7.4 to 9.6× the SDR ceiling.
  • 5.82 s per frame, same checkpoint, no retraining.
LumaFluxarXiv:2604.02787

Limitations

  • The VAE is an SDR codec. The spline cannot restore what clipping removed.
  • Oversaturation. Mean ITP chroma is 60.65 against 49.80 for the references.
  • Native HDRTV1K. In-domain CNNs reach 38.11 dB where we reach 23.56.
  • Speed. 5.282 s per 1080p frame against 0.367 to 1.708 s single-pass.

Still ahead: a larger primary track and a controlled subjective study.

LumaFluxarXiv:2604.02787

Thank you

Code: github.com/shreshthsaini/LumaFlux

Project page: shreshthsaini.github.io/LumaFlux

Paper: arXiv:2604.02787

Related work of mine

  • LumaGuide, training-free HDR generation. Preprint, 2026
  • Seeing Beyond 8bits, HDR-UGC video quality. CVPR 2026
  • Rectified CFG++ for flow based models. NeurIPS 2025
  • BrightRate, quality of user-generated HDR video. WACV 2026, oral, best student paper
  • HIDRO-VQA, contrastive learning for HDR video quality. WACV 2024
UT Austin LIVE YouTube

Backup

Backup slides

Flow matching in four lines

The training recipe behind FLUX.1-dev, which LumaFlux keeps unchanged

xt = (1 − t) x0 + t x1

A straight path from a data sample x0 to a noise sample x1, with t in [0, 1].

ut = x1 − x0

The target velocity along that path is constant.

L = E ‖ vθ(xt, t) − (x1 − x0) ‖2

Train the network vθ by least squares on the velocity.

dx / dt = vθ(x, t)

Sample by integrating from t = 1 back to t = 0 with a few Euler steps. No noise is injected on the way.

z(k+1) = z(k) + (tk+1 − tk) Fθ(z(k), tk, c),   tk = 1 − k/K,   K = 8

The LumaFlux sampler, with z1 = zSDR + σ ε, ε ~ N(0, I), σ = 0.05. Training interpolates zt = (1 − t) zHDR + t z1 and predicts the velocity z1 − zHDR. A text-to-image model starts at pure noise and needs many more steps.

LumaFluxarXiv:2604.02787

Physical conditioning and PGA in detail

What the model measures

  • Linear-light BT.2020 luminance Y = m2020T xlin, with m2020 = [0.2627, 0.6780, 0.0593].
  • Per token: Tphys = Conv3×3([Y, log(1 + |∇Y|), sat]).
  • Per frame: g = MLPg([μY, σY, p95, p99]) in ℝ16.
  • Spectrum: r = log mean energy of the FFT of Y in Kr = 8 radial bands.
  • Gains per step and block: [αpga, βpga, αpcm, βpcm, nspec, λ] = Ψ(t, l). Sinusoidal flow-time features, learned block embeddings, 41,094 parameters.

PGA, a gated low-rank residual on the frozen value projection

Gphys,i = Diag( σ( Pv [Tphys,i ‖ g] ) ),   gFFT = softplus(Wr r)
Rv,it,l = ( αpgat,l AvBv + βpgat,l I ) Gphys,i ( I + nspect,l Diag(gFFT) )
token i uses  WV(0) + Rv,it,l
  • Rank 8, on the image-token value projection only. Bv and Wr start at zero, so training starts from the frozen model.
  • The physical gate is per token and per head; the spectral gate is per head at frame level.
  • Replacing Ψ with six learned constants costs 0.81 dB and 6.09 ΔEITP.
LumaFluxarXiv:2604.02787

PCM and the HDR Residual Coupler in detail

Perceptual Cross-Modulation

[γt,l, ζt,l] = αpcmt,l MLP( meani T′perc,i ) + βpcmt,l
PCM(hl) = hl + γt,l ⊙ LN(hl) + ζt,l
  • FiLM on the normalized state: one scale and one shift per frame, from the pooled 729 SigLIP tokens (width 1152, encoder frozen).
  • γ and ζ are zero-initialized, so step zero is exactly the frozen FLUX.

HDR Residual Coupler

zoutl = zresl + λt,l [ Wp(Tphys) + Wc( R( Cperc(Tperc) ) ) ]
  • Adds both signals to both streams at spatial resolution, after every block.
  • Wp and Wc are pointwise with width-64 bottlenecks and zero-initialized last layers; R resamples the 729 SigLIP tokens onto the latent grid.
zl−1residual stream
PCMFiLM before the block
Frozen MM-DiT block lPGA inside its attention
Couplerresidual after the block
zlrepeated for all 57 blocks

Order inside every frozen block. Nothing in the frozen weights moves.

LumaFluxarXiv:2604.02787

Training

L = LFM + λ1 ‖Ylin − Ylin*‖1 + λ2 ‖xlin − xlin*‖1 + λ3 Lsmooth

(λ1, λ2, λ3) = (1, 0.5, 0.1)

  • LFM regresses the bridge velocity. The two L1 terms compare decoded output and target in linear light (inverse PQ at the 1,000-nit peak), on luminance and on RGB.
  • Lsmooth regularizes adjacent knot slopes of the spline.
  • Decoded losses run every fourth optimizer step on the one-step estimate ẑ0 = zt − t v̂t; the sampler is never unrolled.
  • 5122 crops (1,024 tokens), 100,000 steps, global batch 16, learning rate 2e-4.
  • About 36 hours on four GH200 GPUs, one per node, 144 GPU-hours.
  • A second, HDRTV1K-specific checkpoint trains for 50,000 steps, adapters only. It appears only on the legacy HDRTV1K table and the fixed-QP table, labeled in-domain.
Convergence curves

HDR reconstruction loss against optimizer step. The full design converges to 29.2 dB; without the RQS decoder 19.5 dB; LoRA only 19.3 dB.

LumaFluxarXiv:2604.02787

Frame win rates on the synthetic track

MethodPSNRSSIMΔEITPLPIPSVDP-3HIDRO
BT.2446c14.219.116.57.213.26.5
Reinhard operator27.237.127.514.211.23.5
HDRTVNet++6.211.111.01.26.017.0
FMNet13.213.114.00.21.23.0
ITM-LUT3.20.13.07.22.21.0
VAE + RQS (control)7.04.12.01.04.01.0
LumaFlux (ours)29.215.126.069.262.068.0
Share of aligned synthetic-track frames (%) on which each method is best, exact ties split equally. Higher is better. Bold best, underline second best.
  • LumaFlux wins the most PSNR frames, 29.2%.
  • The fixed Reinhard operator wins the most SSIM and ΔEITP frames, 37.1% and 27.5%, against 15.1% and 26.0% for LumaFlux.
  • On the perceptual metrics LumaFlux is best on 69.2% of frames for HDR-LPIPS, 62.0% for HDR-VDP-3 and 68.0% for FR-HIDROVQA.
LumaFluxarXiv:2604.02787

Native HDRTV1K pairs, where the in-domain CNNs win

MethodPU21-PSNR ↑PU21-PSNR-Y ↑PU21-SSIM ↑ΔEITP ↓HDR-LPIPS ↓HDR-VDP-3 ↑FR-HIDRO ↓
BT.2446c25.0024.890.951139.640.0787.6320.514
Reinhard operator24.4924.320.934243.970.0837.6370.485
HDRTVNet++38.1140.060.98248.130.0438.2240.338
FMNet38.0040.050.98228.100.0478.1880.358
KUNet37.9939.850.98207.840.0388.2350.258
LEDiff10.9911.280.5245198.580.5933.9641.088
X2HDR18.6518.800.658492.200.4455.0710.874
PromptIR23.1823.020.861753.350.0997.5540.475
VAE + RQS (control)21.4723.300.921571.860.2016.1920.804
LumaFlux (ours, mixed-corpus)23.5627.710.931647.590.1706.8990.632
Native track: 117 HDRTV1K test pairs scored with the mixed-corpus checkpoint, which was never tuned on HDRTV1K. ITM-LUT has no result on this track.

CNNs trained on this set reach 38.11 dB. The mixed-corpus model, never tuned on HDRTV1K, reaches 23.56 dB, above X2HDR (18.65) and LEDiff (10.99). That is a distribution gap.

LumaFluxarXiv:2604.02787

The in-domain checkpoint on the legacy HDRTV1K suite

MethodPSNRSSIMSR-SIMΔEITPHDR-VDP-3
HDRTVNet family
AGCM+LE+HG37.210.96990.99689.118.569
AGCM-LE-HR++38.360.97350.99758.288.751
AGCM-LE++, joint38.600.97450.99737.678.696
HDCFM38.420.97320.99747.838.572
FMNet37.940.97470.99578.108.510
KUNet37.780.98680.99717.808.393
Deep SR-ITM37.100.96860.99509.248.233
ITM-LUT36.690.9811–10.198.912
Beyond-FMG35.060.9609–––
RealRep29.120.9412–––
LumaFlux (in-domain, ours)33.340.94270.994114.697.962
Baseline values as published under each author's protocol; "–" means not reported. The LumaFlux row is the HDRTV1K-specific checkpoint (50,000 steps, adapters only) under our 117-image PQ-code protocol. Higher is better except ΔEITP.

Two checkpoints, never in the same row

  • 33.34 dB belongs to the in-domain checkpoint, with adapters retrained on HDRTV1K for 50,000 steps.
  • The main mixed-corpus model scores 27.52 dB zero-shot on the same 117 images under this protocol.
  • The published best is 38.60 dB, so even in domain we stay below the specialized CNNs.

Each baseline keeps its author's protocol, so cross-row ranking is protocol-sensitive. This table and the native-track table use different metric suites and are not interchangeable.

LumaFluxarXiv:2604.02787

Where the highlights go

Headroom maps

Headroom maps of one LumaEval frame. Gray is at or below 100 nit; color marks pixels above the SDR ceiling. Percentages under the maps are for this frame.

MethodShare above 100 nitIoU with reference highlightsError inside highlights (stops)
Reference32.3%––
BT.2446c19.1%––
Reinhard29.7%0.563–
HDRTVNet++47.9%0.5380.81
LumaFlux37.5%0.5790.87
100 synthetic-track frames. Share closer to the reference and higher IoU are better; lower error inside the reference highlights is better.

The analytic curve under-expands, HDRTVNet++ pushes too much of the frame bright, and LumaFlux puts highlights where the reference has them.

LumaFluxarXiv:2604.02787

Best in every column of the synthetic track (full table)

MethodPU21-PSNR ↑PU21-PSNR-Y ↑PU21-SSIM ↑ΔEITP ↓HDR-LPIPS ↓HDR-VDP-3 ↑FR-HIDRO ↓
BT.2446c23.7424.220.829161.600.3075.6110.753
Reinhard operator23.6224.250.828661.760.2955.7260.753
HDRTVNet++23.0323.230.819266.670.3285.4170.703
FMNet22.8622.970.817667.920.3305.4120.724
ITM-LUT22.9223.060.814267.410.3155.4960.761
VAE + RQS (control)18.5419.410.7647111.670.3813.6730.924
LumaFlux (ours)24.2324.570.829456.540.2695.8120.631
Macro-averages over the same 100 frame IDs and 78 video-variant groups. Bold best, underline second best. FR-HIDRO is the HIDRO-VQA feature distance (Saini et al., WACV 2024).
24.23 dBPU21-PSNR, 0.49 dB above the strongest baseline
56.54ΔEITP, 5.06 lower than the strongest baseline
0.631FR-HIDROVQA, from 0.703 for HDRTVNet++

One checkpoint, no per-track tuning. Without diffusion transport the control collapses to 18.54 dB.

LumaFluxarXiv:2604.02787

On human-graded SDR, LumaFlux leads five of seven metrics (full table)

MethodPU21-PSNR ↑PU21-PSNR-Y ↑PU21-SSIM ↑ΔEITP ↓HDR-LPIPS ↓HDR-VDP-3 ↑FR-HIDRO ↓
BT.2446c33.1634.260.844237.830.1487.3970.660
Reinhard operator31.4832.240.827043.590.1557.1730.662
HDRTVNet++31.5332.050.812544.920.1477.1440.671
FMNet31.6432.110.813744.850.1477.1350.675
KUNet31.5131.960.807845.260.1437.0580.647
ITM-LUT31.4632.060.808545.490.1427.0900.639
LEDiff18.3018.990.6480103.150.4964.7230.898
X2HDR23.7824.690.757851.280.3665.5320.758
PromptIR29.2129.670.675653.110.1856.1410.690
VAE + RQS (control)26.0527.330.723191.700.2375.8140.863
LumaFlux (ours)33.3135.170.848040.190.1386.9190.544
The 100 LIVE-TMHDR pairs whose SDR side is an expert grade, with the preregistered per-content exposure calibration. KUNet and PromptIR (86 of the 100 synthetic IDs) and LEDiff and X2HDR (five synthetic variants) are scored here and on the native track.

FR-HIDROVQA 0.544, 14.9% below the best baseline. BT.2446c keeps the lower color error, 37.83 against 40.19, and the higher HDR-VDP-3.

LumaFluxarXiv:2604.02787

Removing the spline costs 9.69 dB (full table)

RowPGAGatePCMCoup.T-modRQSPSNR (Δ vs g)ΔEITP (Δ vs g)
(a) LoRA only•19.29 (−9.90)256.84 (+229.03)
(b) attention + MLP LoRA••28.36 (−0.83)36.35 (+8.53)
(c)•••28.57 (−0.62)33.48 (+5.66)
(d)••••29.56 (+0.37)28.38 (+0.57)
(e)•••••29.20 (+0.01)31.76 (+3.95)
(f) seed-0 repeat of (g)••••••28.67 (−0.52)35.00 (+7.18)
(g) full design, ours••••••29.19 (ref.)27.81 (ref.)
(h) linear spline•••••lin.28.74 (−0.45)32.16 (+4.35)
(i) no RQS•••••19.50 (−9.69)261.78 (+233.97)
(j) no T-mod•••••28.38 (−0.81)33.90 (+6.09)
Component ablations: 20k steps on a 55-video pool, the same 82 held-out frames. A dot marks a component that is on; Gate is the spectral gate, T-mod the timestep-and-layer modulation, lin. a linear spline. Higher PSNR and lower ΔEITP are better; deltas are relative to row (g).
  • No RQS, row (i), falls 9.69 dB.
  • A linear spline, row (h), costs 0.45 dB.
  • The spectral gate, (c) to (d), lifts 28.57 to 29.56 dB.
  • No timestep-layer modulation, row (j), costs 0.81 dB.

Rows (f) and (g) are the same configuration, 0.52 dB apart, so smaller gaps are noise.

LumaFluxarXiv:2604.02787

Heavier compression barely moves LumaFlux (full table)

PSNR change from QP 27

PU21-PSNR change from QP 27 (QP is the HEVC quantization parameter) on the 117 HDRTV1K pairs. Flatter is better.

From QP 27 to 42 the mixed-corpus model loses 0.55 dB. HDRTVNet++ loses 2.77 and FMNet 2.71.

The CNNs still score higher at every QP on their home set.

MethodQP 27QP 32QP 37QP 42Δ
PU21-PSNR (dB), higher is better
HDRTVNet++35.9035.3034.3433.13−2.77
FMNet35.5634.9734.0432.85−2.71
LumaFlux (mixed-corpus)27.2927.2727.0426.75−0.55
LumaFlux (in-domain)33.9433.6733.1232.33−1.61
ΔEITP, lower is better
HDRTVNet++10.4111.1812.4713.983.58
FMNet10.5311.3312.6214.183.65
LumaFlux (mixed-corpus)30.3230.5631.3032.081.76
LumaFlux (in-domain)13.6414.0915.0116.202.56
HDR-VDP-3, higher is better
HDRTVNet++8.4098.1677.7597.198−1.211
FMNet8.3298.0977.7047.156−1.173
LumaFlux (mixed-corpus)7.6157.5257.3086.948−0.667
LumaFlux (in-domain)7.9187.7917.5337.117−0.802
Fixed-QP track, 117 HDRTV1K pairs. Δ is QP 42 minus QP 27; closer to zero is better. Mixed-corpus is the main checkpoint; in-domain is the HDRTV1K-specific one. Bold best, underline second per column and block.
LumaFluxarXiv:2604.02787

What it costs (full table)

5.28 sper 1080p frame, 8 steps, one GH200
27.06 GBpeak memory at 1080p
71.32Mtrainable of 12.48B parameters
Multi-resolution frontier

FR-HIDROVQA (lower is better) on 30 expert-track frames. LumaFlux 0.431, 0.477 and 0.567 at 540p, 720p and 1080p, in 2.64, 3.08 and 5.29 s.

MethodStepsParams (M)Trainable (M)s/frameTFLOPsMem. (GB)
BT.2446c10.000.000.415––
HDRTVNet++10.590.590.8570.52.31
FMNet11.301.300.6401.52.16
KUNet12.162.160.5573.73.76
ITM-LUT*10.600.600.367–0.18
PromptIR + Reinhard135.5935.591.70810.917.86
LEDiff (5122)501370.23859.534.4491496.37
X2HDR (5122)3017020.02149.424.2441,02333.77
LumaFlux, 8 steps812484.7771.325.2822,06727.06
LumaFlux, 40 steps4012484.7771.3224.99310,10027.06
One GH200, batch 1, complete 1080p exports; LEDiff and X2HDR at their registered 5122 protocol (scaled by pixel count to 1080p: 35.19 and 33.57 s). *ITM-LUT runs a compiled kernel the tracer cannot observe. Lower is better for s/frame, TFLOPs and memory.

2.4 to 16.1× faster than X2HDR and LEDiff. HDRTVNet++ stays 173 to 201× faster than us.

LumaFluxarXiv:2604.02787

Video without temporal training (full table)

Per-frame mean PU21 luma

Per-frame mean PU21 luma, four clips, first 120 frames, relative to each reference mean. Only the stabilized run is drawn.

MethodBonfire02c3055e03920237Mean
Reference0.0000.0000.0000.0000.0000.000
BT.2446c7.8131.4110.4500.0630.1241.972
HDRTVNet++3.0050.2070.1470.4421.2571.012
FMNet3.8530.2020.0740.5791.1241.166
LumaFlux1.9920.6912.5000.9680.5691.344
LumaFlux + stabilization2.1090.4691.6760.8790.3621.099
Temporal-energy deviation |flickerpred − flickerref| in 106 PU21-energy units on five clips (02c3 to 0237 are CHUG content IDs). Smaller is closer to the reference; bold marks the closest method per column.
  • Reuse one bridge noise ε for every frame.
  • Smooth the spline over time, θ̄f = 0.8 θ̄f−1 + 0.2 θf.
  • Signed excess flicker falls 53.3%. Raw flicker stays within 3.27% of the reference.

Nothing is trained for video, so there are no temporal layers, optical flow or fine-tuning.

LumaFluxarXiv:2604.02787

Three scenes enlarged, every HDR method

Enlarged regions of Nature, Bright scene and Monument for each method and the reference

The red box on each SDR frame marks the enlarged region, all under one fixed display transform. HDR shown tone-mapped.

The flame row shows the chroma overshoot that LumaFlux still has.

LumaFluxarXiv:2604.02787

The same regions in log luminance

Log-luminance patches for every method on three regions

Every method on one log scale from 0.1 to 1,000 nit. SDR uses a 100-nit display white; LumaFlux outlined.

Selected examples, not an aggregate ranking.

LumaFluxarXiv:2604.02787

Luminance along one row, two scenes

Luminance along one row for City towers and Sunset forest

Mean absolute log deviation from the reference: BT.2446c, HDRTVNet++, LumaFlux = 0.190, 0.166, 0.058 (City towers); 0.272, 0.171, 0.078 (Sunset forest).

LumaFluxarXiv:2604.02787

Text-to-HDR keyframes over time

Campfire and sunlit forest clips at four time points, SDR and HDR lift

Each panel splits into the frame and its log-luminance map, 0.1 to 1,000 nit. SDR maps saturate at the 100-nit ceiling.

LumaFluxarXiv:2604.02787

Three more text-to-HDR lifts

Campfire, neon street and sunlit forest lifts with headroom maps

Each lift at its keyframe with the largest area above 100 nit. Gray is at or below 100 nit, color above.

Of the pixels above 100 nit, 90.4% to 98.3% lie among the brightest source pixels.

LumaFluxarXiv:2604.02787

Training evolution

Reconstructions and per-pixel error maps at 5k, 15k, 30k and 50k steps

Reconstruction at 5k to 50k steps; per-pixel ΔEITP against the HDR reference falls as training runs.

LumaFluxarXiv:2604.02787

Frozen-VAE round trips, PQ against PU21

Six HDR crops encoded and decoded by the frozen VAE under PQ and PU21 coding

Six LIVE-TMHDR crops through the frozen VAE. Lower right of each round trip is its per-pixel ΔEITP.

LumaFluxarXiv:2604.02787

One crop at three display exposures

HDRTVNet++, LumaFlux and reference at minus 4, 0 and plus 4 EV

One stored crop at −4, 0 and +4 EV of display exposure, with one shared display mapping.

Dropping exposure shows what each method put above the SDR range.

LumaFluxarXiv:2604.02787

Tone variants of one content stay close inside the frozen model

Within-content against between-content similarity for VAE tokens and velocity directions

Cosine similarity within one content's tone variants, against between contents, for VAE tokens and velocity at t = 0.25, 0.5, 0.75.

The within-content gap is positive in all four panels.

LumaFluxarXiv:2604.02787