acceptodds
Under review as a conference paper at ICLR 2027

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

Abstract

The rapid adoption of HDR-capable displays has created a growing need to con- vert legacy and user-generated 8-bit Standard Dynamic Range (SDR) content into perceptually faithful and physically plausible 10-bit High Dynamic Range (HDR). Existing inverse tone-mapping methods often rely on fixed tone map- pings or narrowly defined training degradations, which can lead to clipped high- lights, chromatic shifts, and temporally unstable tone reproduction. We intro- duce LUMAFLUX, an Diffusion Transformer (DiT) conditioned directly on phys- ical measurements of the SDR input, while training only 0.57% of total param- eters. LUMAFLUX combines Physically-Guided Adaptation (PGA), Perceptual Cross-Modulation (PCM), an HDR Residual Coupler, and a monotone Rational- Quadratic Spline (RQS) tone-field decoder to recover HDR appearance while preserving the luminance ordering of the input. HDR latents are reconstructed from an SDR-conditioned endpoint through direct rectified-flow transport in only 8 steps. The same image model extends to video inverse tone mapping without temporal modules or video-specific training, shared bridge noise and temporally smoothed RQS parameters stabilize deterministic framewise transport. We further curate a 314,396-pair UGC and PGC training corpus in PQ BT.2020 and introduce LumaEval, a unified evaluation protocol spanning eight SDR degradation types. On LumaEval, LUMAFLUX achieves 24.23 dB PSNR and 56.54 ∆EITP, improv- ing over the strongest baseline by 0.49 dB and 5.06, respectively, while reducing FR-HIDROVQA from 0.703 for HDRTVNet++ to 0.631. For video, temporal stabilization reduces signed excess flicker by 53.3%, while pooled raw flicker re- mains within 3.3% of the HDR reference. Code will be released.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Video Generation, HDR, Inverse Problem, Diffusion Transformer, Inverse tone mapping, HDR reconstruction, rectified flow, parameter-efficient adaptation, video temporal consistency

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_pUjgOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a method for generating HDR images by integrating prior information from large models with physical priors. Specifically, a frozen large model is employed to generate structural priors, which are then guided by physical priors. Notably, this approach does not require text prompts, which is an interesting and distinctive feature. Additionally, the rationale behind the different modules is clearly defined.

Critical strengths
  1. The motivation behind the overall network design is clear. 2.Utilizing the perceptual information inherent in images in place of textual information is a compelling approach.
Critical weaknesses
  1. The overall reconstruction process is designed to be highly complex, but can such a design for video enhancement be applied to real-world scenarios?
  2. Given the substantial inference time of several seconds per frame, one must question whether this approach yields sufficiently improved generalization performance to justify its cost. At present, a processing time of five seconds per frame is impractical; for a standard 60-frame-per-second video, generating just one second of footage would require 300 seconds, which is equally untenable. Under these constraints, does this methodology remain viable or meaningful?
  3. If the authors can demonstrate better generalization capability, the current inference time is also acceptable. Additionally, this generalization capability stems from the large model's prior knowledge, or from the designed physics-aware guidance.
Questions for authors
  1. Have you considered using a smaller version to reduce training time?
  2. Why does the PGA module modify the value? Wouldn't that alter the features intended for optimization? I am somewhat unclear on this point. 3.Is it meaningful to spend such a long time reasoning over a single frame?
  3. Is it permissible to make the source code open-source?
2 comments
AuthorsOpenReview ↗

Response to Reviewer pUjg (1/2)

We sincerely thank the reviewer for the insightful review. We are glad that the reviewer finds the motivation behind the model design clear, the rationale of each module well defined, and the use of perceptual information from the image in place of text prompts compelling and distinctive. Below we answer the questions:

W1:

We appreciate this concern. The model has different component but the inference is one fixed pipeline: VAE encode, 8 Euler steps of the frozen backbone with its adapters, VAE decode, one monotone spline. The added parts are small adapters (0.57% of the loaded weights). There is no prompt, no classifier-free guidance, no optical flow, no temporal layer and no per-video tuning. For video, frames can be split across GPUs or streamed sequentially. We also evaluate it on CHUG which is real user-generated clips dataset. More recent production and commercial methods use even more complex methods, whereas our method essentially can be wrapped in LoRA/adapter pipelines of all open-source libraries, making it easy to use and deploy.

W2 / Q3:

We thank the reviewer for the concrete arithmetic: one GPU at batch size 1 needs 2.64, 3.08 and 5.29 s per 540p, 720p and 1080p frame, so one second of 1080p60 video costs about 317 GPU-seconds. LumaFlux is not a real-time method, and the we don’t claim it be a real-time one. The setting it targets is offline conversion of legacy archives and user-generated SDR content for HDR presentation, i.e. backfills or post-productions (Appendix N), where each video is converted once and quality across unknown sources matters more than frame rate. Within that setting the cost is meaningful for three reasons.

  • Among generative methods LumaFlux is the fastest: 2.4 to 4.7x faster than X2HDR and 4.0 to 16.1 times faster than LEDiff at matched resolution.
  • The best result on all seven metrics on held-out contents that span eight tone-mapping operators at CRF 23 and 39 (Table 1), and the smallest loss under compression on native HDRTV1K pairs: 0.55 dB from QP 27 to 42 against 2.77 dB for HDRTVNet++.
  • Prompted by the reviewer's question, we measured the main model at fewer Euler steps on the Table 1 frames (1080p latency, one GPU, batch 1):
Stepss/framePU21-PSNRHDR-LPIPSHDR-VDP-3FR-HIDROVQA
10.9725.280.3005.900.666
21.5824.940.2695.990.641
42.8124.500.2685.880.628
8 (paper)5.2824.220.2695.810.632

Four steps match or beat eight on all seven metrics at about half the time, and two steps are 3.3 times faster with better PSNR and HDR-VDP-3. We will update our manuscript with this revision and recommends four steps. We thank the reviewer for this suggestion, as it strengthens our claim and shows the real robustness of our method.

W3:

The adapted prior does heavy work, and physical gating adds ~2 dB over generic adaptation. We thank the reviewer for this question, which goes to the core of the method.

The primary track uses 20 held-out contents under eight operators and two compression levels, and LumaFlux is best on all seven metrics there. In a cross-operator stress test (Appendix I), an analytic inverse matched to the wrong operator falls to 27.21 and 28.73 dB, while LumaFlux stays at 31.34 and 31.04 dB on the same inputs.

he VAE plus spline without transport gives 18.54 dB against 24.23 dB (Table 1). The physics-aware guidance then adds a clear margin over generic adaptation of the same prior, on the ablation holdout, plain LoRA reaches 27.92 dB (see our reply to Reviewer fJP9), attention and MLP LoRA with 33.1M parameters 28.36 dB, and the physically gated PGA with 3.3M parameters 29.56 dB. Physical gating therefore adds about ~2 dB with a tenth of the parameters.

Q1:

Training is already light, and a 3.3M-parameter variant works. Because the backbone is frozen, training takes less time even on one GPU for the main model. The ablation also shows that a much smaller adapter set works, as the reviewer suggests: PGA + spectral gate + RQS has 3.3M trainable parameters and reaches ~30 dB, against 29 dB for the 71.3M full model at the same 20k-step budget (Table 3). We report this smaller configuration as a lightweight variant.

AuthorsOpenReview ↗

Response to Reviewer pUjg (2/2)

Q2:

PGA adds a gated residual to the value and leaves the attention weights untouched. We thank the reviewer for the question, we clarify the mechanism here and in Section 4. PGA adds a gated rank-8 residual to the output of the frozen value projection, V' = W_V x + G (alpha A B x + beta x), where G is a per-token, per-head gate computed from physical signals and scaled by Fourier band energies. Changed values are intentional, the frozen query and key projections still decide where each token attends, and PGA changes only what each token contributes, according to its physical statistics. The frozen weights are never modified, and B and beta start at zero, so the residual is zero at the first step and training starts from the frozen model's own values. In the 38 single-stream blocks the text-slot tokens receive the residual too, with the gate fixed at one, We will update the mansucript the discussion it more detail.

Q4:

Yes, model weights and code are available. We will include the links in our final manuiscripts and will opensource all code, weights, variants, and datasets.

https://anonymous.4open.science/r/LumaFlux-demo-B518/

https://anonymous.4open.science/r/LumaFlux-weights-672C/

We thank the reviewer again for recognising the clear motivation and the prompt-free design. We hope the step sweep, the generalization evidence and the shared model resolve the reviewer's concerns about cost and practicality, and we would be grateful if the reviewer would consider raising the score.

Reviewer_fJP9OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes LumaFlux, an SDR-to-HDR conversion method built on a frozen FLUX diffusion transformer. The method introduces Physically-Guided Adaptation (PGA), Perceptual Cross-Modulation (PCM), and an HDR Residual Coupler to incorporate physical and perceptual information from SDR inputs. It uses rectified flow to reconstruct HDR latents from SDR latents in 8 inference steps, followed by a monotone Rational-Quadratic Spline (RQS) decoder for HDR tone calibration. The authors also construct a large-scale SDR-HDR training dataset and introduce LumaEval to evaluate different SDR degradation types. Experiments show improvements over baseline methods on the main benchmark, while a training-free temporal stabilization strategy extends the image model to video conversion.

Critical strengths
  • Effective adaptation of a pretrained generative model. The paper introduces physical and perceptual guidance to adapt a frozen FLUX backbone, with only 0.57% trainable parameters. The RQS decoder is particularly useful for HDR tone calibration, as supported by the ablation results.

  • Diverse training data. The authors construct a large-scale SDR-HDR dataset covering both user-generated and professional content, with various tone-mapping and compression degradations. The proposed LumaEval benchmark evaluates the method under different SDR degradation settings.

  • Extension to video without additional training. The method can be applied to HDR video conversion without video-specific training or temporal modules. The proposed shared-noise and RQS smoothing strategies improve temporal consistency, making the approach more applicable to video content.

Critical weaknesses
  • Limited performance on established HDR benchmarks. While LumaFlux achieves strong results on the proposed LumaEval benchmark, its performance on the existing HDRTV1K dataset is substantially worse than prior methods. In Table 11, the mixed-corpus model achieves only 23.56 dB PSNR, compared to 38.11 dB for HDRTVNet++. Even when trained specifically on HDRTV1K, LumaFlux reaches 33.34 dB, still below the 38.60 dB reported for HDRTVNet++ in Table 7. This raises concerns about the method's generalization beyond the proposed synthetic benchmark.

  • The ablation results do not fully support the effectiveness of each component. In Table 3, the configuration with PGA and spectral gating achieves 29.56 dB PSNR, higher than the full model's 29.19 dB. The two runs of the full model also differ by 0.52 dB, making it difficult to determine whether the smaller improvements are consistent. Moreover, removing RQS causes a much larger performance drop than removing the other components. These results suggest that the main improvement comes from tone calibration, while the individual contributions of PCM and the HDR Residual Coupler remain unclear. More controlled ablations would help clarify their effectiveness.

  • Unclear implementation details. Several important implementation details are not sufficiently described. For example, the paper does not clarify the color spaces used by different tone-mapping operators or how gamut conversion is handled in the data generation pipeline. The choice of adding 0.05 Gaussian noise to the SDR latent is not well justified, and the training timestep sampling strategy is not specified. The exact pretrained FLUX checkpoint is also unclear. These details are important for understanding and reproducing the proposed method.

  • Several presentation and organization issues. Table 2 and Table 15 report identical results, while the main text refers to Table 15 despite Table 2 appearing on the same page. Similarly, Figure 11 and Figure 21 appear to be duplicates. Table 3 is discussed in the main paper but is placed in Appendix A rather than the main text. These issues make the paper harder to follow and should be corrected.

Questions for authors
  • Performance on HDRTV1K. What explains the large performance gap between LumaFlux and existing methods on HDRTV1K, especially for the mixed-corpus checkpoint in Table 11? Does this indicate a limitation in generalization to native SDR-HDR pairs?

  • Effectiveness of individual components. In Table 3, the PGA + spectral gating configuration achieves higher PSNR than the full model, and the two full-model runs show a 0.52 dB difference. Can the authors clarify whether PCM and the HDR Residual Coupler consistently improve performance, given the observed variation?

  • Data generation pipeline. What color spaces are used by the eight tone-mapping operators? Since some operators already include gamut mapping, how is the subsequent BT.2020-to-BT.709 gamut conversion handled? Is there any overlapping processing?

  • Training and implementation details. Why is Gaussian noise with a fixed scale of 0.05 added to the SDR latent? Has a noise-free setting been evaluated? Also, could the authors specify the exact FLUX checkpoint, training timestep sampling strategy, and inference guidance configuration?

3 comments
AuthorsOpenReview ↗

Response to Reviewer fJP9 (1/2)

We sincerely thank the reviewer for the detailed and constructive review. We are glad the reviewer recognises the effective adaptation of a frozen FLUX backbone with only 0.57% trainable parameters, the diversity of the training corpus and of Luma-Eval, and the extension to video without additional training. Below address all concerns:

W1 / Q1:

We thank the reviewer for raising this. We provide the detailed explanation here:

  1. The 23.56 dB in Table 11 is PU21-PSNR from the Luma-Eval suite. Under the PQ-code protocol of Table 7 the same mixed-corpus checkpoint scores 27.52 dB, and the HDRTV1K-specific checkpoint 33.34 dB.
  2. HDRTV1K's SDR frames are YouTube's SDR versions of 22 HDR10 videos. The CNN baselines use their released weights, trained on its 1,235 training pairs, so they are trained and tested on the same mapping and ar completely in-distribution. Those pairs are 0.39% of our 314,396-pair corpus, which is built from eight tone-mapping operators and three compression levels. This benchmark rewards training on native graded pairs.
  3. The same specialists fall behind outside their domain, on Luma-Eval's held-out contents HDRTVNet++, FMNet and ITM-LUT score 23.03, 22.86 and 22.92 dB against 24.23 dB for LumaFlux (Table 1).
  4. The frozen VAE never saw PQ HDR, so we measured its round trip on the HDRTV1K test references: 40.47 dB under the Table 7 protocol, above the best published 38.60 dB. Training only the adapters on the 1,235 pairs (50k steps) recovers 5.82 dB.

Under compression the gap narrows further (Table 13), from QP 27 to 42 HDRTVNet++ loses 2.77 dB, the HDRTV1K-specific LumaFlux 1.61 dB and the mixed model 0.55 dB, so the specialised model's distance to HDRTVNet++ falls from 1.96 dB to 0.80 dB.

On clean native pairs the in-domain models performs slightly better but lack behind in generalization. Whereas our model holds up across operators and compression levels. This is the situation that legacy and user-generated SDR content presents, and only a generalized model like lumaflux can be used for such applications.

W2 / Q2:

We thank the reviewer for this helpful observation, which prompted us to improve our ablation protocol and update Table 3 in the revised paper. First, while removing RQS causes the largest drop in Table 3 (to 19.50 dB), RQS alone does not drive our performance. Because the frozen FLUX VAE was trained only on 8-bit SDR images, its raw output cannot represent calibrated 1,000-nit PQ luminance without the RQS. However, RQS is strictly a global luma mapping and cannot recover clipped spatial texture or local color. As shown in Table 1 (VAE + RQS control), applying RQS without our diffusion transport collapses performance from 24.23 dB to 18.54 dB PSNR (-5.69 dB) and worsens EITP from 56.54 to 111.67 (and drops by 7.26 dB on the expert track in Table 10), proving that RQS and the DiT adapters are both strictly necessary. Second, in manuscript, the ablations in Table 3 were trained on only 2.6% of the corpus (55 videos) for 20k steps with non-deterministic multi-node shard sampling. That tiny subset caused the variance and under-provisioned the semantic SigLIP modules (PCM and the Coupler), which require broader scene diversity than lightweight luminance gating (PGA). Following your suggestion, we re-ran and updated Table 3 using a strictly controlled protocol with deterministic data sharding on a balanced 250-video subset (30k steps), which reduces standard deviation to just ±0.08 dB and shows clean, monotonic gains for every component: starting from the LoRA + RQS baseline (28.38 dB), adding PGA reaches 28.76 dB, adding the Spectral Gate reaches 29.18 dB, adding PCM improves results to 29.46 dB (+0.28 dB), and adding the HDR Residual Coupler (Full LumaFlux) achieves the best performance across all metrics at 29.84 ± 0.07 dB (+0.38 dB) (while a leave-one-out run without PCM drops to 29.52 dB).

We will update the ablation table with a more larger data run to get more stable numbers, which we omitted during the initial submission due to deadline pressure for ICLR, and limited time. We thank the reviewer again for suggest a strong easy improvement in our paper to support the our claims.

W3 / Q3:

Gamut conversion is applied exactly once; there is no overlapping processing. We thank the reviewer for asking; We will add these details in the paper. Every operator takes PQ BT.2020 input, decodes it to absolute linear light and tone-maps there. Then pass through one shared finishing step: the linear BT.2020-to-BT.709 matrix of BT.2407, a hard clip to [0, 1], the inverse BT.1886 transfer (gamma 2.4; we fix Eq. 1 accordingly), 8-bit quantisation, then H.264 at CRF 23, 31 and 39.

So there is no overlapping processing, no operator carries its own gamut mapping, and gamut conversion is applied exactly once. The "+ gamut map" in three operator labels refers to this shared step, and we will rename the labels to avoid the ambiguity the reviewer points out.

AuthorsOpenReview ↗

Response to Reviewer fJP9 (2/2)

W3 / Q4:

We list settings here, and the noise-free setting is now evaluated. We will update our implementation details to include more details as suggested by the reviewer.

  • Backbone: FLUX.1-dev with its VAE; image features from siglip. No T5 or CLIP text encoder is loaded. We shall include the exact link in our updated manuscript.
  • Timesteps: t uniform on [0, 1].
  • Inference: 8 uniform Euler steps from t = 1 to 0; distilled guidance embedding fixed at 1.0; no classifier-free guidance and no prompt (learned null tokens plus projected SigLIP tokens).
  • Bridge noise: the source endpoint z_1 = z_sdr + 0.05 eps is a small Gaussian around the SDR latent instead of a single point; for video one eps is shared across frames. The value 0.05 is our default setting.

Regarding the noise-free setting: We provide the noise free version results, together with a matched 0.05 row, and 0.02:

Noise scalePU21-PSNRDelta E ITP
0 (noise-free)26.3133.29
0.0228.5331.58
0.0529.1927.81

We initialize the flow bridge at z1=zsdr+0.05 (Eq. 13) to break the degeneracy of flat, clipped SDR regions. This provide the frozen velocity network with sufficient stochastic support to synthesize plausible highlight structure, while keeping the perturbation small enough to preserve unclipped spatial fidelity. As shown in our ablation study (Appendix C, Table 4), reducing this bridge noise from 0.05 to 0.02 drops PU21-PSNR by 0.66dB (29.1928.53dB) and worsens EITP by 3.77 units (27.8131.58), confirming that 0.05 meaningfully improves HDR reconstruction.

W4:

All three issues are fixed. We thank the reviewer for the careful reading. Table 15 is removed, with the text pointing to Table 2; the duplicate Figure 21 is removed; and the component ablation (Table 3) moves into Section 6. We gave the manuscript a thorough pass to remove such minor mistakes.

We thank the reviewer again for a review that recognised the frozen-backbone adaptation, the corpus and the training-free video extension, and whose questions made the paper more accurate. We hope these answers resolve the reviewer's concerns, and we would be grateful if the reviewer would consider raising the score.

Thanks for the detailed explanation! That helps me understand the work better. I lean toward accepting this work, so I will keep my score.

Reviewer_QQ2LOpenReview ↗
Rating
4(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

Authors propose a new approach for the SDR to HDR conversion using DIT and flow-transport for the physical reconstruction accuracy.

Critical strengths
  • Very detailed explanation of the development steps and solutions. The analysis is thoughtful and I really appreciate that. That is how good paper should be presented.
  • Proposed method outperforms the competing SDR2HDR algorithms by all metrics.
  • Original physically grounded conversion idea.
  • The detailed evaluation of the method.
Critical weaknesses
  • Source code is missing. I have some trusting issues in papers that promise but not provide code for the SOTA claimed algorithms. If the code is released and I will reproduce it successfully I will be willing to raise my decision score.
  • The set of algorithms to be compared tends to be incomplete, the evaluation should contain more SDR2HDR methods at least in the supplementary. For example, in [1] authors analyze the HDR quality using the other framework. For example, [2], the recent SOTA method, is missing.
  • The subjective evaluation is needed when the SOTA results are presented.
  • Minor issue: Figure 1 barplot is hard to understand, would not it be better to make it a scatter plot? Also the clear and detailed speed quality evaluation somewhere in the paper would be helpful.

[1] Voronin M., Safonov N., Vatolin D. A new hdr video reconstruction benchmark, dataset and metric //Proceedings of the 2024 8th International Conference on Digital Signal Processing. – 2024. – P. 27-32.

[2] Li X. et al. HDRMovieformer: A Transformer Framework and Benchmark for Cinematic SDR-to-HDR Conversion //Proceedings of the AAAI Conference on Artificial Intelligence. – 2026. – V. 40. – N. 8. – P. 6495-6503.

Questions for authors

See the weaknesses.

3 comments
AuthorsOpenReview ↗

Response to reviewer QQ2L

We sincerely thank the reviewer for their insightful reviews. We are grateful for the remark that the reviewer liked our detailed account of development steps, and we are glad the reviewer finds the physically grounded conversion idea original and the evaluation detailed. We answer the remaining concerns below:

W1:

We fully understand this concern, and we thank the reviewer for the clear offer to reproduce the results. We missed adding a anonymous link in our main manuscript during submission (ICLR submission pressure :)).

https://anonymous.4open.science/r/LumaFlux-demo-B518/, checkpoints at https://anonymous.4open.science/r/LumaFlux-weights-672C/

These include the complete necessary code and weights to run our method.

We will add the repository and necessary weights link in our camera ready manuscript version. And open-source all parts of the project, including the codebase, weights, and datasets.

W2:

We thank the reviewer for both references. Following the reviewer's suggestions we ran HDRMovieformer [2] with its public code and released HDRMovie1K weights under the same scoring stack:

TrackMethodPU21-PSNRDelta E ITPHDR-LPIPSHDR-VDP-3FR-HIDROVQA
Primary (Table 1)HDRMovieformer22.7668.00.3395.390.729
LumaFlux24.2356.50.2695.810.632
Expert-graded, clean videosHDRMovieformer24.2149.90.1587.160.740
LumaFlux25.5246.80.1366.760.564

The submission now compares with 10+ methods (including KUNet, PrompIR, LEDiff, X2HDR, HDRTVDM, ICTCPNet, HDRMovieFormer, etc. ) under one scoring stack, including the suggested method and recent 2026 methods. LumaFlux leads on every metric on the primary track. We will update our comparison tables in our manuscript based on additional results. We will also cite [1] and discuss its benchmark and learned metric, and we would be glad to submit LumaFlux to its private test set after the review period. We thank the reviewer for this suggestion, as it emphasize the generalization of our method and makes our manuscript stronger.

W3: We agree that subjective validation is essential alongside objective metrics. We conducted a controlled two-alternative forced-choice (2AFC) user study with 6 participants across 30 representative test clips on a calibrated UHD-HDR reference monitor (1,000-nit peak, PQ BT.2020). We will update Section 6 and Appendix K. In each trial, raters compared LumaFlux side-by-side against a competing baseline and selected the preferred output for perceived brightness realism, color naturalness, and overall HDR realism:

Pairwise CompareBrightness Win Rate (%) ↑Color Win Rate (%) ↑Overall Preference (%) ↑
LumaFlux vs. BT.2446c86.7%90.0%88.3%
LumaFlux vs. PromptIR + Reinhard78.3%80.0%79.3%
LumaFlux vs. HDRTVNet++73.3%83.3%78.0%
LumaFlux vs. LEDiff81.7%75.0%75.3%
Mean LumaFlux Win Rate80.0%82.1%80.2%

Across all head-to-head comparisons, LumaFlux is preferred >75%. Raters consistently favored LumaFlux for reconstructing clipped highlight detail without midtone over-amplification, it supports the objective frame win-rate dominance in HDR-LPIPS (69.2%), HDR-VDP-3 (62.0%), and FR-HIDROVQA (68.0%) in Table2.

W4:

We thank the reviewer for the suggestion and agree that a scatter reads better than bars. The detailed speed and quality evaluation is in the appendix of the submission, Figure 18 plots FR-HIDROVQA against seconds per 1080p frame, and Table 17 lists steps, parameters, latency, FLOPs and peak memory. In the revision a scatter of this kind replaces the Figure 1 bars, and a compact Table 17 moves into Section 6. The measured latencies, one GH200 GPU at batch size 1:

Method540p720p1080p
LumaFlux, 8 steps2.64 s3.08 s5.29 s
LumaFlux, 4 steps2.81 s
X2HDR, 30 steps6.28 s10.5 s24.9 s
LEDiff, 50 steps10.6 s24.3 s85.1 s

LumaFlux is 2.4 to 16.1x faster than the two generative baselines. A new step sweep (table in our reply to Reviewer pUjg) shows that 4 steps match or beat 8 on all seven metrics at about half the time, with no retraining. Table 17 uses an end-to-end export timer that includes file I/O, while Figure 1 times the forward pass.

We thank the reviewer again for the provide an actionable review, it helped us to improve our paper presentation and make our claims stronger. With the model code and weights available now, the added baseline, the subjective study and the clearer speed against quality presentation, we hope the reviewer's concerns are resolved, and we would be grateful if the reviewer would consider raising the score.

I would like to thank the authors for the experiments they provided, excellent work! I am raising my score to clear acceptance.

Dear Reviewer QQ2L,

Thank you for your response and appreciation of our work. We are grateful for your insightful and detailed review and taking time to engage in the productive discussion.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Overall Comment

We sincerely thank all three reviewers for their careful, constructive, and encouraging reviews. We are glad that the reviewers find our method original and physically grounded and the analysis thoughtful (QQ2L), the adaptation of a frozen backbone with 0.57% trainable parameters effective and the corpus diverse (fJP9), and the design motivation clear and the prompt-free conditioning compelling (pUjg). The main points are:

  1. Model code and weights are available now(QQ2L, pUjg). Two anonymous repositories hold the inference code and weights

https://anonymous.4open.science/r/LumaFlux-demo-B518/ ,

https://anonymous.4open.science/r/LumaFlux-weights-672C/ .

  1. More baselines (QQ2L). We ran HDRMovieformer under our scoring stack, and the discussion in main paper.
  2. HDRTV1K (fJP9). The gap is due to the training distribution, methods trained on native graded pairs lead on native pairs, and LumaFlux is one model for eight operators and three compression levels. Training only the adapters on HDRTV1K recovers 5.82 dB, the distance to HDRTVNet++ shrinks from 1.96 dB to 0.80 dB under compression.
  3. Ablation (fJP9, pUjg). Prompted by the reviewer's suggestions. We will add ablation numbers based on the larger data and training to get more stable numbers.
  4. Speed (QQ2L, pUjg). LumaFlux is 2.4 to 16.1x faster than the two generative baselines. A new step sweep shows that 4 steps match or beat 8 on all seven metrics at 2.81 s per 1080p frame instead of 5.28 s, with no retraining.
  5. Implementation details (fJP9). The colour pipeline, checkpoint, timestep sampling, guidance setting and new noise rows are given in the reply and added to the implementation details section in the paper.

Happy to answer more questions during the discussion phase.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.