arXiv preprint · arXiv:2604.02787

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

8-bit SDR to 10-bit HDR with a physically-guided diffusion transformer

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

The University of Texas at Austin · Google, Inc.

What starts here changes the world

What Happens When HDR Becomes SDR

The forward pipeline from HDRTVNet (ICCV '21): every step is lossy, and inverting it is ill-posed

HDRTVNet pipeline: Raw Data → Tone Mapping → Gamut Mapping → Transfer Function → Quantization. Top (blue) = SDR. Bottom (orange) = HDR.

Luminance Compression

PQ maps 0–10,000 cd/m². Gamma tops at 100 cd/m². SDR clips 99% of the luminance range. Highlights → flat white. Shadows → crushed black. 8-bit quantization → banding.

Color Gamut Compression

BT.2020 (UHDTV, dashed outer) covers 75.8% of CIE. BT.709 (HDTV, dotted inner) covers only 35.9%. Over half the color volume is discarded: saturated greens, deep reds, vivid cyans lost irreversibly.

Forward: x_sdr = Γ_OETF⁷⁰⁹(M_2020→709(Γ_EOTF^PQ(x_hdr)/L_max)) + ε : PQ decode → gamut compress → gamma encode → quantize. Inverting requires content-aware generative priors.

LumaFlux (arXiv:2604.02787)LIVE Lab, UT Austin

What starts here changes the world

What Gets Lost: 10-bit HDR vs 8-bit SDR (Beyond8Bits Dataset)

Split-view from our Beyond8Bits dataset: HDR (upper-left) vs SDR (lower-right) of the same frame

Refrigerator scene (top-left): SDR crushes shadow detail inside the fridge: metallic shelving and contents disappear to black. HDR reveals all internal structure.

Autumn lake (top-right): SDR desaturates the vivid fall foliage and loses the cloud detail in the sky. HDR preserves the full color volume and highlight gradations in the water reflection.

Crystal close-up (bottom): SDR clips the specular highlights on the gemstone to flat white. HDR preserves the translucent internal structure and surface reflections.

LumaFlux (arXiv:2604.02787)LIVE Lab, UT Austin

What starts here changes the world

Real HDR Videos in Yxy Space: What SDR Clips

Computed from real 10-bit PQ BT.2020 videos (Beyond8Bits). Red prism = Rec.709 SDR volume (≤100 nits). Orange dots = pixels outside SDR: lost in tone mapping.

Sunset scene: 15% pixels above SDR ceiling. Sun flare + specular highlights on grass reach ~1,300 nits.

Indoor/night scene: 27% pixels above SDR ceiling. Bright artificial lights reach ~2,600 nits: massive clipping.

Key insight: 15–27% of HDR pixel information is permanently destroyed by SDR conversion, and the loss is concentrated in the most perceptually important regions: highlights, specular reflections, light sources, and wide-gamut colors. This is what ITM must reconstruct.

LumaFlux (arXiv:2604.02787)LIVE Lab, UT Austin

What starts here changes the world

Why ITM Is Fundamentally Ill-Posed

Three irreversible losses make exact inversion impossible: any method must hallucinate missing information

1. Many-to-One Mapping

Forward tone mapping is surjective: multiple distinct HDR luminance values (e.g., 500, 2000, 8000 cd/m²) all map to the same clipped SDR code value (255). The inverse has infinite solutions: the mapping is non-invertible.

2. Color Gamut Loss

BT.2020 → BT.709 gamut mapping discards 53% of the color volume. Saturated greens, deep reds, and vivid cyans in the wide gamut are compressed into the sRGB triangle. The original chroma information cannot be recovered from the compressed representation.

3. Quantization Loss

10-bit PQ (1024 levels) → 8-bit gamma (256 levels) destroys fine gradations. In shadows, adjacent PQ code values are 4× further apart in 8-bit, causing visible banding/contouring. In highlights, the entire upper luminance range collapses to a few code values.

Any ITM method must hallucinate plausible HDR detail where SDR has none.

This requires content-aware generative priors: not fixed curves or local CNN features.

LumaFlux (arXiv:2604.02787)LIVE Lab, UT Austin

What starts here changes the world

Prior ITM Methods: Three Generations of Architectures

From hand-crafted curves to CNNs to diffusion: each generation has fundamental limits

Gen 1: Classical Tone Curves

BT.2446, Reinhard, Huo et al.: Fixed parametric mapping from SDR→HDR using inverse tone curves. Content-blind: the same curve is applied to sunsets, interiors, and neon signs. Cannot recover scene-specific highlights or local contrast.

Gen 2: CNN-Based Learning

HDRTVNet++ (ICCV'21): 3-branch modulation net (global, local, condition). HDCFM (MM'22): Hierarchical feature modulation with dynamic context. HDRTVDM (CVPR'23): Dynamic context transformation. Deep SR-ITM (ICCV'19): Joint super-resolution + ITM.

Limitation: Local receptive fields → cannot model global illumination. Overfit to specific TMOs used in training data.

Gen 3: Diffusion-Based

LEDiff: Latent diffusion conditioned on SDR. PromptIR: Text-guided restoration. FlashVSR: Video super-resolution with diffusion prior.

Limitation: Hue shifts, over-saturated tones, hallucinated details. No physical grounding: the model has no concept of luminance or color space.

LumaFlux (Ours)

A frozen FLUX.1-dev MM-DiT adapted with PGA, PCM, the HDR Residual Coupler, and the RQS tone-field decoder.

HDRTVNet++ (Chen et al., TMM'23): 3-step CNN pipeline: adaptive global color mapping → local enhancement → highlight refinement. All operations are local (conv + FC layers). No global scene understanding.

LEDiff (Wang et al., CVPR'25): Latent exposure fusion: encoder produces multi-exposure latents, learnable fusion in latent space, separate shadow/highlight denoisers. No physical priors, prone to hue shifts.

What LumaFlux does differently

PGA uses luminance, gradient, saturation, and spectral-band cues. PCM supplies FiLM conditioning from frozen SigLIP.

LumaFlux (arXiv:2604.02787)LIVE Lab, UT Austin

What starts here changes the world

Generative Class: Like-for-Like Native-Track Comparison

Quality

LumaFlux reaches 23.56 PU21-PSNR and 47.59 dE-ITP on the 117-pair native track.

Solver steps

LumaFlux uses 8 steps, compared with 50 for LEDiff and 30 for X2HDR.

Runtime

LumaFlux costs 2.5 seconds per megapixel, versus 17.0 for LEDiff and 16.2 for X2HDR.

Trainable parameters

LumaFlux uses 71 M trainable parameters, versus 860 M for LEDiff and 149 M for X2HDR.

Method	Steps	PU21-P	dE-ITP	s/MP	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

LumaFlux (arXiv:2604.02787)LIVE Lab, UT Austin

What starts here changes the world

Prior ITM Methods: Visual Failures

Visual comparison on real HDR dataset: from FastHDRNet (arXiv 2404.04483)

Desert sunset (top) + evening cityscape (bottom). Ground Truth vs 7 methods: Ada-3DLUT, HDRTVNet, JSI-GAN, CSRNet, HDRNet, HuoPhyEO, FastHDRNet.

Color shift & washout

Ada-3DLUT & HDRTVNet wash out the sun and lose sand gradients. HDRNet introduces severe blue color shift on the cityscape. HuoPhyEO produces unnatural yellow-green cast.

Halo & detail loss

JSI-GAN adds halo artifacts around the sun disc. CSRNet and JSI-GAN lose building detail and city lights in shadow regions. No method recovers both highlights and shadows.

The common pattern

LumaFlux combines physical cues in PGA with perceptual conditioning from frozen SigLIP in PCM.

LumaFlux (arXiv:2604.02787)LIVE Lab, UT Austin

What starts here changes the world

Background Flow Matching in 4 Equations

The generation framework behind FLUX.1-dev and LumaFlux

1. Interpolation Path

x_t = (1−t)x₀ + t·x₁

data → noise, straight line

2. Target Velocity

u_t = x₁ − x₀

Constant: network learns to predict this

3. Training Loss

L = E[‖v_θ(x_t,t) − (x₁−x₀)‖²]

Simple MSE: predicted vs true velocity

4. Inference: Solve ODE

dx/dt = v_θ(x,t)

t=1→0 via Euler. Deterministic, no noise.

Color: data x₀ · noise x₁ · interpolant xₜ · time t · network vθ

Background: Flow MatchingLIVE Lab, UT Austin

What starts here changes the world

Background Rectified Flow: Straight Paths vs. Noisy Diffusion

The published rectified-flow bridge for FLUX.1-dev

From Liu et al. (ICLR '23): (a) Linear interpolation has crossing paths. (b) Rectified flow straightens them via reflow. (c-d) After rectification, paths are straight ≈ optimal transport. Figure from official Rectified Flow repo.

Generative baseline: LEDiff

Uses 50 steps, costs 17.0 seconds per megapixel, and has 860 M trainable parameters on the published native-track comparison.

LumaFlux rectified-flow bridge

Starts from z₁ = VAE_enc(SDR) + 0.05·ε and integrates from t=1 to t=0 in 8 steps.

Published operating point: 8 steps, prompt-free, 5.28 seconds per 1080p frame, and 27.06 GB peak memory on a single GH200.

Background: Rectified FlowLIVE Lab, UT Austin

What starts here changes the world

Method Frozen FLUX.1-dev with Lightweight Adaptation

Frozen backbone

FLUX.1-dev MM-DiT and its VAE remain frozen. The SigLIP encoder used for perceptual conditioning also remains frozen.

Standard LoRA

W = W⁽⁰⁾ + AB

LumaFlux adds gated low-rank attention residuals while keeping the backbone fixed.

What the adaptation must capture

Physical cues: luminance, gradient, saturation, and spectral bands.
Perceptual conditioning from frozen SigLIP.
Timestep-and-layer modulation.
Display-referred HDR luminance calibration.

LumaFlux: Adapting the Frozen MM-DiT

Four published components adapt the frozen model:

PGA: Physically-Guided Adaptation

PCM: Perceptual Cross-Modulation

RQS tone-field decoder

HDR Residual Coupler

71,315,893 trainable parameters, 0.57% of the FLUX.1-dev backbone

BackgroundLIVE Lab, UT Austin

What starts here changes the world

Architecture Evolution: Frozen MM-DiT → Adapted MM-DiT

The LumaFlux path adds PGA, PCM, the HDR Residual Coupler, and the RQS tone-field decoder while keeping FLUX.1-dev frozen.

CNN: Local, no physical priors, no global context

MM-DiT: Frozen FLUX.1-dev backbone and VAE

Adapted MM-DiT: SigLIP, physical cues, HDR Residual Coupler, and RQS

LumaFlux (arXiv:2604.02787)LIVE Lab, UT Austin

What starts here changes the world

LumaFlux: Full Pipeline

PGA
Luminance + gradient + saturation + FFT → gated LoRA on W_V

PCM
Frozen SigLIP → cross-attn → FiLM: γ⊙LN(h)+ζ

HDR Residual Coupler
Fuses physical and perceptual paths under timestep-and-layer modulation

RQS tone-field decoder
Monotone rational-quadratic spline for display-referred HDR luminance

LumaFlux (arXiv:2604.02787)LIVE Lab, UT Austin

What starts here changes the world

Frozen MM-DiT Block → Adapted MM-DiT Block

The colored adaptation paths contain 71,315,893 trainable parameters, 0.57% of the frozen backbone

The adapted block adds PGA, PCM, and the HDR Residual Coupler.

// Frozen MM-DiT block
z → LN(z)
Q = zWQ, K = zWK
V = zWV
h = Attn(Q,K,V)
z = z + h + MLP(LN(z))
// + LumaFlux adaptation:
V = z(WV0 + Rvt,ℓ) ← PGA
h = γt,ℓ⊙LN(h) + ζt,ℓ ← PCM
z += λ(WpTphys + WcC(Tperc)) ← Coupler

HDR Residual Coupler

Fuses the physical and perceptual paths under timestep-and-layer modulation.

LumaFlux (arXiv:2604.02787)LIVE Lab, UT Austin

What starts here changes the world

PGA: Physically-Guided Adaptation: Step by Step

Step 1: Extract physical features

Physical cues = luminance + gradient + saturation + spectral bands

T_phys = Conv_3×3([Y, log(1+|∇Y|), sat])

G_phys = gated low-rank attention residuals

Step 2: Physical & spectral gating

G_phys = diag(σ(P_v[T_phys ‖ g]))

g_spectral = modulation from spectral-band cues

Step 3: Modulated LoRA

R_v^t,ℓ = (α·AB + β·I) · G_phys · (I + n_spec·diag(g_FFT))

W_V ← W_V⁽⁰⁾ + R_v^t,ℓ

Three multiplicative modulations:

Timestep-and-layer modulation: adapts the residuals across the solver trajectory and model depth.

G_phys: Per-head luminance-aware gating. Sigmoid: 0=suppress, 1=pass.

g_FFT: Frequency-aware gating. Prevents over-expansion in smooth (low-freq) regions.

Key property: Unlike standard LoRA (static), PGA is input-dependent, timestep-adaptive, and frequency-aware. The same image region gets different adaptation at different diffusion timesteps.

LumaFlux (arXiv:2604.02787)LIVE Lab, UT Austin

What starts here changes the world

PCM (Perceptual) & RQS (Decoder)

Perceptual Cross-Modulation

T_perc = φ_SigLIP(x_sdr) (frozen)

[γ^t,ℓ, ζ^t,ℓ] = α_pcm · MLP(C(T_perc)) + β_pcm

PCM(h) = γ ⊙ LN(h) + ζ

FiLM: scale & shift per dimension after layer norm. Enforces color constancy: e.g., brighten sky highlights without shifting skin tones.

RQS Tone-Field Decoder

Ŷ(α) = [η_k(s_k+1)α² + d_ks_kα] / [s_k + (s_k+1+s_k−2d_k)α(1−α)]

Monotone rational-quadratic spline calibration.
Monotonic: preserves luminance ordering.
Differentiable & invertible.
Operates in YUV BT.2020: separates luminance from chroma.

Rectified-flow bridge

z₁ = VAE_enc(SDR) + 0.05·ε

Inference

Integrate t=1 to t=0 in K=8 steps

No text prompt and no sampling-strength knob.

LumaFlux (arXiv:2604.02787)LIVE Lab, UT Austin

What starts here changes the world

Training: Data & Configuration

314,396 SDR-HDR pairs from 2,092 source videos

Source	Published scope	Role
CHUG	856 reference contents	User-generated HDR
LIVE-TMHDR	40 references + 120 SDR clips	Expert-graded SDR
HDRTV1K	1,235 training pairs	Native pairs
Sol Levante	HDR10 master	Open content

Frozen degradation chain

8 tone-mapping operators, BT.2020-to-BT.709 gamut conversion, and 3 CRF settings. Inputs also include genuine expert grades.

Training Configuration

Backbone: Frozen FLUX.1-dev and VAE
Trainable: 71,315,893 parameters, 0.57%
Main model: 100,000 optimizer steps, global batch 16, lr 2e-4, 512px crops, bf16
Training: Four GH200 GPUs, about 36 hours, approximately 144 GPU-hours
Inference: 8 steps, prompt-free, 5.28 s per 1080p frame

Evaluation Benchmarks

HDRTV1K
117 published test pairs

Generative class
117 native-track pairs

Luma-Eval
Three tracks
Seven metrics

PU21-PSNR, PU21-PSNR-Y, PU21-SSIM, dE-ITP, HDR-LPIPS, HDR-VDP-3, and FR-HIDROVQA

LumaFlux (arXiv:2604.02787)LIVE Lab, UT Austin

What starts here changes the world

Results: Best on Every Luma-Eval Metric

Method	PU21-P↑	PU21-S↑	dE-ITP↓	LPIPS↓	VDP-3↑	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	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

+0.49dB over strongest
baseline

-5.06dE-ITP over
strongest baseline

0.631FR-HIDROVQA
best

8solver steps
prompt-free

LumaFlux is best in every column. The Luma-Eval synthetic track uses 100 held-out frames and identical inputs, output encoding, and metric implementations for every method.

LumaFlux (arXiv:2604.02787)LIVE Lab, UT Austin

What starts here changes the world

Qualitative Example 1: Church Interior

SDR Input (left half)

The source is an 8-bit SDR BT.709 image.

LumaFlux HDR Output (right half)

The output is 10-bit HDR in PQ BT.2020, produced prompt-free in 8 solver steps.

Published limitation

The frozen VAE is not an HDR codec. The tone head cannot invent detail absent from heavily clipped SDR.

LumaFlux (arXiv:2604.02787)LIVE Lab, UT Austin

What starts here changes the world

Qualitative Example 2: Shanghai Cityscape

SDR input (left) vs LumaFlux HDR output (right): split view with zoomed crop below

SDR input (left half)

The source is an 8-bit SDR BT.709 image.

LumaFlux output (right half)

The output is 10-bit HDR in PQ BT.2020, produced prompt-free in 8 solver steps.

Published limitation

Mean ITP chroma is 60.65 for LumaFlux versus 49.80 for the references, indicating slight oversaturation.

LumaFlux (arXiv:2604.02787)LIVE Lab, UT Austin

What starts here changes the world

Efficiency: The 8-Step Operating Point

Published measurements on a single GH200

Measure	Published value	Context
Solver steps	8	Prompt-free
1080p runtime	5.28 s	Single GH200
Peak memory	27.06 GB	Single GH200
40-step cost	4.7x	No quality gain
Trainable parameters	71,315,893	0.57% of backbone
Adapter file	285 MB	Backbone separate

Eight steps is the published operating point.

Published efficiency summary

Steps

8

Runtime

5.28 s

Peak memory

27.06 GB

Trainable

0.57%

Comparison with 40 steps

Forty steps cost 4.7 times more with no quality gain. LumaFlux therefore operates at 8 steps.

LumaFlux (arXiv:2604.02787)LIVE Lab, UT Austin

What starts here changes the world

Training-Free Video Stabilization

The same image model is used for video, with no temporal layers, optical flow, video fine-tuning, or temporal loss

53.3%less signed excess
PU21 flicker

3.3%from reference raw
temporal energy

0temporal
training

1shared image
model

Shared bridge noise

The same bridge-noise realization is shared across frames.

RQS-parameter EMA

An EMA is applied to the RQS parameters used across frames.

No temporal model

No temporal layers, optical flow, video fine-tuning, or temporal loss are used.

LumaFlux (arXiv:2604.02787)LIVE Lab, UT Austin

What starts here changes the world

What We Learn & Remaining Challenges

Key insights

Eight steps are enough. Forty steps cost 4.7 times more with no quality gain.
Prompt-free conversion. There is no text prompt and no sampling-strength knob.
Parameter-efficient adaptation. 71,315,893 trainable parameters are 0.57% of the FLUX.1-dev backbone.
Compression robustness. The mixed-corpus model loses only 0.55 dB from QP 27 to 42.

Remaining challenges

Frozen VAE limitation. The tone head cannot invent detail absent from heavily clipped SDR.
Temporal limitation. Stabilization is inference-time only, with no explicit long-range motion model.
Speed. Inference takes 5.28 seconds per 1080p frame, versus 0.28 to 0.86 seconds for feed-forward CNN converters.
Color. Mean ITP chroma is 60.65 for LumaFlux versus 49.80 for the references on the qualitative scenes.

LumaFlux (arXiv:2604.02787)LIVE Lab, UT Austin

What starts here changes the world

Key Takeaways

1

8-bit SDR to 10-bit HDR. Prompt-free BT.709 to PQ BT.2020 conversion in 8 solver steps.

2

71.3M trainable parameters. Only 0.57% of the frozen FLUX.1-dev backbone.

3

Best on every Luma-Eval metric. The margin over the strongest baseline is +0.49 dB and -5.06 dE-ITP.

4

Training-free video stabilization. Shared noise and an RQS-parameter EMA cut signed excess PU21 flicker by 53.3%.

Code and adapter weights are released. The weights require FLUX.1-dev and use its Non-Commercial License.

LumaFlux (arXiv:2604.02787)LIVE Lab, UT Austin