Pith. sign in

REVIEW 3 major objections 5 minor 43 references

LumaGuide: Distribution Shaping for Training-Free HDR Generation in Diffusion Models

T0 review · 3 major / 5 minor · reviewed 2026-08-01 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Aligning a frozen diffusion model's PQ-luminance histogram to an HDR target at sampling time is sufficient to produce HDR-consistent images without retraining.

desk verdict LumaGuide is a genuinely useful training-free recipe for steering diffusion backbones toward HDR-looking luminance distributions, but the headline distribution metrics are measured against targets from the authors' own regressor, so external validation is needed before the HDR claim is fully grounded. read the letter →

arxiv 2607.26237 v1 pith:KPS3V2JA submitted 2026-07-28 cs.CV

Bowen Chen , Shreshth Saini , Balu Adsumilli , Alan C. Bovik This is my paper
classification cs.CV
keywords HDRimagegenerationtraining-freeguidancediffusionmodelsluminancehistogramalignmentPQ/perceptuallyuniformencodingWasserstein-1distanceenergy-basedsamplingvideotemporalconsistency
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

LumaGuide claims that a pretrained diffusion model can be made to generate high-dynamic-range images without any retraining, by steering its sampling trajectory toward a target luminance distribution. The paper's central assertion is that aligning the output's luminance histogram in perceptually uniform PQ space is sufficient to induce HDR-consistent behavior: structured highlights, preserved shadow detail, and a wider dynamic range, while semantic content stays intact. This matters because existing HDR generation approaches either retrain or fine-tune the generative backbone, add reconstruction pipelines, or fuse exposure brackets; LumaGuide proposes a training-free, model-agnostic alternative that works on image and video models alike. The authors back the claim with theory (energy descent, spatial decoupling, PQ gradient stability, and an impossibility result for simple brightness scaling) and with experiments across multiple backbones, target-specification modes, and a video extension. A sympathetic reader would take the paper as establishing that low-level distribution shaping is a viable control mechanism for generative model outputs.

What carries the argument

The load-bearing object is the differentiable soft histogram of per-pixel PQ luminance, paired with the Wasserstein-1 distance on that histogram. PQ (SMPTE ST.2084) is a perceptually uniform luminance encoding, so histogram bins behave like visually meaningful brightness levels and gradients remain scale-balanced across the HDR range. Wasserstein-1 is used because it respects the ordering of bins and, through its closed-form CDF representation, keeps nonzero gradients flowing toward sparse highlight bins where KL and ℓ2 fail. This energy is backpropagated through the decoder into the latent, with a constant guidance schedule. Supporting results: Theorem 1 establishes average energy descent u

What would settle it

Calculate uW1 between guided and unguided outputs on a video-disjoint, held-out set of HDR prompts whose targets come from the regressor; if the guided distribution does not move meaningfully toward the target, the sufficiency claim fails. A sharper probe: deliberately feed an infeasible target (e.g., a nearly uniform mid-gray histogram for a prompt about fireworks); the paper's own logic predicts persistent distribution mismatch and visible over-constraint, not HDR-like behavior.

Watch

Extended reading notes

Core claim

The paper's central claim is that, for a frozen pretrained diffusion model, matching the PQ-encoded luminance histogram of the predicted image to a target HDR distribution is sufficient to make the output 'behave like HDR': highlights gain structure and roll-off, shadows retain detail, and the robust luminance span widens, all without changing the model weights. The authors instantiate this as differentiable energy guidance: at each sampling step they decode the flow estimate, compute a soft histogram of its BT.2020 luminance mapped through the PQ transfer function, measure the Wasserstein-1 distance to the target histogram, and backpropagate that scalar energy into the latent trajectory. Be

Load-bearing premise

The load-bearing premise is that the target HDR luminance histograms produced by the small text-to-histogram regressor are correct, representative, and feasible for the pretrained diffusion prior; the paper itself concedes that some targets are infeasible and that saturated pixels can become unguidable (Appendix A.11), so the claim stands or falls on target validity and achievability.

Editorial extensions

If this is right

  • Text-to-image models can produce HDR content at inference time with no retraining or fine-tuning, making HDR capability a plug-in property of existing backbones.
  • Users can choose their HDR look by selecting target histograms from presets, reference images, or text descriptions, since all three plug into the same objective.
  • If the sufficiency claim holds, video diffusion models gain zero-shot HDR synthesis when the same per-frame guidance is paired with a temporal coherence penalty.
  • A simple global brightness adjustment is mathematically insufficient to create HDR distributions, which refocuses HDR generation on distribution shape rather than exposure scaling.
  • Because the guidance is permutation-invariant, spatial layout and semantics stay under the control of the diffusion prior, so distribution control can be decoupled from structure control.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Editorial extension: the same soft-histogram energy could be applied to other perceptually meaningful scalar features (e.g., chroma saturation or local contrast) with targets drawn from references or text, extending the 'distribution shaping' claim beyond luminance.
  • Editorial extension: the paper's own failure cases suggest a testable refinement: an adaptive feasibility check that detects when a target is unachievable on the model's manifold and relaxes the constraint, rather than forcing a compromise solution.
  • Editorial extension: if PQ histogram alignment is genuinely sufficient for HDR appearance, then cheap distributional metrics (uW1, percentile errors) could replace subjective HDR studies for tuning; the paper's reported Q-Eval mismatch hints that current SDR-trained perceptual metrics are the wrong yardstick for HDR.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper proposes LumaGuide, a training-free framework that steers pretrained diffusion models at sampling time by adding an energy-based gradient term that matches the PQ-luminance histogram of the predicted image to a target histogram via Wasserstein-1 distance. The method is instantiated for HDR generation, with targets provided by presets, reference images, or a text-driven regressor, and is extended to video through a temporal luminance coherence term. The authors prove energy descent, spatial decoupling, PQ gradient stability, and a no-go result for affine brightness scaling, and support the method with experiments on Flux.1, SD3, SDXL, and CogVideoX, including ablations, cross-backbone comparisons, and a subjective study. The central claim is that aligning luminance histograms in PQ space is sufficient to induce HDR-consistent behavior, such as coherent highlights and preserved shadow detail, without modifying model weights.

Significance. If the core HDR claim were independently validated, LumaGuide would be a practically valuable plug-and-play control method: it requires no retraining, supports multiple target-specification modes, and extends to video with a lightweight temporal term. The empirical pipeline is extensive, with three image backbones, one video backbone, ablations, and an honest failure analysis in A.11. The theoretical structure, especially the spatial-decoupling argument (Proposition 1) and the no-go result for affine brightness scaling (Theorem 2), provides useful framing for why histogram shaping can preserve semantics. However, the significance is tempered by the fact that the target distributions that define both the guidance and the headline metrics are generated by a regressor trained on the authors' own dataset (A.7), so the quantitative evidence for 'HDR-consistent behavior' is, without external validation, agreement with a self-defined target rather than with an independent ground-truth HDR distribution.

major comments (3)
  1. [Sec. 4.1, Sec. 4.2, A.7, Tables 1-2] The default target luminance distributions are produced by the text-to-histogram regressor described in A.7, trained on the authors' Beyond8Bits dataset. All distribution-alignment metrics (uW1, p50dist, p99dist) measure distance to these regressor-predicted targets. Consequently, the quantitative claim that histogram alignment 'is sufficient to induce HDR-consistent behavior' (Abstract) is currently grounded only in matching a target that the authors themselves define. Since A.11 concedes that some targets are infeasible for the prior, and A.10 notes that Q-Eval quality decreases with guidance, the central HDR claim needs external validation. Please evaluate against ground-truth HDR luminance distributions from an independent dataset, or at least calibrate the regressor targets against real HDR captures and report the sensitivity of the results to regressor error.
  2. [Theorem 1, A.2.6] Theorem 1's descent guarantee relies on L-smoothness of z -> E(x_hat_0(z)) and on a bounded velocity-error assumption. The proof in A.2.6 refers to 'assumptions (A1)-(A3)', but these are never stated in the paper. Moreover, L is not estimated for any of the tested backbones, so the step-size condition eta_t <= 2/L cannot be verified, and the practical guidance scale s0=2000 is not connected to the theory. Please state the assumptions explicitly, provide empirical evidence of L-smoothness along sampled trajectories, or reclassify Theorem 1 as a formal statement under idealized conditions that are not validated in the experiments.
  3. [A.10, A.8, Table 2] The paper reports in A.10 that Q-Eval quality decreases with guidance even in the moderate regime, yet Table 2 lists Q-quality as a headline metric; LumaGuide's Q-quality (0.568) is below X2HDR's (0.579). The paper attributes this to Q-Eval's SDR bias, but this claim is not independently substantiated. The sole independent evidence for HDR fidelity is the subjective JOD study, whose stimuli are normalized so that every image's 99.5th-percentile maps to 4000 cd/m2 (A.8). This normalization removes absolute-luminance differences and may reward contrast rather than HDR accuracy. Please provide a validation of Q-Eval's behavior on real HDR images, or an alternative perceptual-quality metric that tracks the claimed HDR improvement, and discuss explicitly what the subjective normalization does and does not test.
minor comments (5)
  1. [A.2.6] The proof of Theorem 1 uses 'assumptions (A1)-(A2)' and 'under (A3)', but these assumptions are never defined anywhere in the manuscript. Define them before the proof or remove the references.
  2. [A.2.5] Proposition 3 refers to 'the TLC term of Definition 8', but no Definition 8 exists in the paper. Fix the cross-reference.
  3. [A.2.3, A.5.1] The no-go result for brightness scaling is said to be 'observed empirically' in Table 7, but the table labeled Table 7 in the appendix is about bin count, not brightness scaling. Renumber the tables or correct the citation.
  4. [Sec. 3.4] The heading 'Uniform target specification' appears to be a typo for 'Unified target specification' or 'Target specification modes'; clarify the wording.
  5. [A.2.8] The proof of Proposition 2 states that the inverse PQ Jacobian scales as L^(1-1/m2), but the displayed derivative formula and the leading behavior in the text suggest L^m1. Make the derivation from the PQ formula to the stated scaling explicit.

Circularity Check

1 steps flagged · score 4.0 of 10

Quantitative HDR-consistency evidence is measured against the same predicted target histogram that LumaGuide optimizes, so the distribution-alignment results are largely self-consistent rather than independently validated.

  1. fitted input called prediction [Section 2.2 (Eq. 4), Section 3, Algorithm 1, Section 4.1; Appendix A.7]
    "The energy used in our framework is E(x)=D(hσ(x),h∗), where h∗ is the target histogram and D is the Wasserstein-1 distance ... By default, target luminance distributions are obtained using the text-driven regressor described in the Appendix. ... We use uW1 for unweighted Wasserstein-1 distance between generated and target PQ-luminance histograms over K=32 bins, p50dist and p99dist for absolute percentile errors in PQ space."

    Algorithm 1 minimizes E = W1(hσ, h*) at every sampling step, while the headline metric uW1 is the W1 distance between the generated and the same target PQ histogram. Therefore Tables 1, 2, 5, 12, and 13 mostly demonstrate that the optimizer reduces the quantity it optimizes, not that the result is HDR. This would still be informative if h* were an independent ground-truth HDR histogram for each prompt, but h* is the output of a 270K-parameter MLP trained on the authors' own Beyond8Bits dataset (A.7), with no external validation against independent HDR ground truth. The paper itself (A.10) concedes Q-Eval quality decreases as guidance increases, so the remaining independent evidence is primarily subjective and normalized to a common 99.5th-percentile of 4000 cd/m2, which removes absolute lu

full rationale

Not all of LumaGuide is circular: the W1-in-PQ gradient is a principled first-principles construction, Theorem 1 and Proposition 1 are standard descent/permutation statements, and the method is genuinely training-free. However, the paper's central empirical claim—that histogram alignment induces HDR-consistent behavior—is evaluated largely with the same object being optimized. The default target h* is not an external ground truth for the prompt: it is predicted by a regressor trained on the authors' HDR UGC dataset, and the quantitative distribution metrics measure distance to that predicted target. Because the guidance scale monotonically drives outputs toward exactly this target, the observed uW1 improvements (Table 1) are the expected result of running the optimizer, not an independent confirmation of HDR fidelity. The independent evidence is the subjective study, yet it normalizes all outputs to a common 4000 cd/m2 99.5th percentile, and the paper's own A.10 notes that Q-Eval quality decreases with guidance. Thus the load-bearing 'HDR' label is self-referential to a target the authors defined; still, the method has independent mathematical and qualitative content, so this is a partial circularity rather than a collapse of the derivation.

Assumptions & free parameters 6 free parameters · 6 assumptions · 0 invented entities

No new physical entities, forces, or latent dimensions are introduced. 'Distribution shaping' and the TLC loss are procedural constructs rather than entities with external falsifiable handles. The free-parameter count is dominated by the per-backbone tuning of the guidance scale and the fitted text-to-histogram regressor that defines the target distribution.

free parameters (6)
  • guidance scale s0 = 2000 (Flux), 1000 (SD3), 50 (SDXL), 3000 (CogVideoX)
    Selected per backbone by trading uW1/p99 against artifacts (Tables 1, 12, 13; A.5.2); the central method depends on this hand-tuned strength.
  • soft histogram bandwidth sigma = 0.5/K with K=32
    Chosen to keep boundary leakage controlled; no ablation on sigma is reported (Eq. 3, A.2.1).
  • number of histogram bins K = 32
    Ablated in A.5.1; K=32 is a trade-off point and uW1 degrades at K=64+.
  • TLC weights lambda_TLC = lambda_f = lambda_c = 10 (video)
    Set for video experiments (A.9); no sensitivity study is reported.
  • text-to-histogram regressor weights = 270K-param MLP trained on Beyond8Bits
    The target distributions used in guidance and evaluation are outputs of this fitted model (A.7); despite the 'training-free' framing, a trained component is involved.
  • perceptual-log weighting = log-damped bin weighting (form in appendix)
    Shown best in Table 4; a design choice selected on the same benchmark set and not specified in the main text.
assumptions (6)
  • domain assumption Pretrained diffusion backbones (Flux.1-dev, SD3, SDXL, CogVideoX-5B) are capable of rendering semantically coherent images from prompts.
    LumaGuide preserves semantics by relying on the prior; if the prior fails, the method cannot rescue it (invoked throughout Section 4).
  • domain assumption Pretrained LDR VAEs faithfully reconstruct PQ/PU21-encoded HDR signals without retraining.
    Algorithm 1 backpropagates through the decoder; the pipeline assumes Dec is invertible enough on PQ-encoded content. This is credited to the unpublished preprint X2HDR [38] and not independently verified here (Section A.3).
  • ad hoc to paper The composed energy E(x0_hat(z)) is L-smooth in z.
    Theorem 1 requires it; no bound is derived for the actual soft-histogram-W1-decoder composition (Appendix A.2.6).
  • ad hoc to paper Bounded velocity error with finite second moment, E||v_theta - v*||^2 <= sigma_v^2, and the unguided-drift term is bounded (A3).
    Used in the proof of Theorem 1; no empirical estimate of sigma_v^2 or the drift bound is provided (Appendix A.2.6).
  • domain assumption PQ gradient scale-balance properties hold over four orders of magnitude in luminance.
    Analytic claim about the PQ OETF in Proposition 2; plausible for standard constants but not validated on real decoder Jacobians.
  • ad hoc to paper For Theorem 3, trajectory sensitivity scales as (1-t)^2 and endpoint variance is regularized.
    The paper itself labels these 'idealized assumptions' (A.2.4) and then argues in Remark 1 that they do not hold in practice.

how reviews work

0 comments
Cite this review

Pith. "Pith review of LumaGuide: Distribution Shaping for Training-Free HDR Generation in Diffusion Models." pith.science (2026). https://pith.science/paper/KPS3V2JA

@misc{pith2026260726237,
  author       = {Pith},
  title        = {Pith review of: LumaGuide: Distribution Shaping for Training-Free HDR Generation in Diffusion Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KPS3V2JA}},
  note         = {Machine review of arXiv:2607.26237}
}
read the original abstract

Pretrained diffusion models generate realistic images but are constrained by the statistical biases of their training data, limiting their ability to produce high dynamic range (HDR) content. In this work, we introduce LumaGuide, a training-free framework for distribution shaping in diffusion models. Instead of modifying model parameters, LumaGuide steers the sampling process to match target feature distributions via differentiable energy-based guidance. We instantiate this framework for HDR generation by controlling luminance distributions in perceptually uniform PQ space. Our results show that aligning luminance histograms is sufficient to induce HDR-consistent behavior, including coherent highlights and preserved shadow detail, while maintaining semantic fidelity. Beyond HDR, LumaGuide enables flexible specification of target distributions through data-driven presets, reference images, or text-driven predictors, and extends naturally to video generation with temporal consistency constraints. More broadly, our work demonstrates that controllable generation can be achieved by directly shaping output distributions at sampling time, without retraining diffusion models.

Figures

Figures reproduced from arXiv: 2607.26237 by the authors.

Figure 1
Figure 1. LumaGuide steers a pretrained diffusion model toward HDR-consistent luminance distribu￾tions at sampling time, without modifying any model weights. Generated outputs preserve semantic content while producing structured highlights and faithful shadow detail under exposure adjustment, matching target HDR statistics in PQ space. Abstract Pretrained diffusion models generate realistic images but are constrained by the s… view at source ↗
Figure 2
Figure 2. Per-pixel PQ-luminance surface for an identical seed and prompt. The Flux.1 [5] baseline (top) overshoots into clipped highlights; LumaGuide (bottom) redistributes mass into mid￾tones and preserves the highlight gradient, match￾ing the target HDR distribution. Pretrained diffusion models [13, 35, 29, 30, 26] have demonstrated remarkable abilities to gener￾ate realistic images from text prompts, yet their outputs rem… view at source ↗
Figure 3
Figure 3. HDR visualization under different exposure levels. In PQ space, Flux.1 [5] often yields over-exposed, non-HDR luminance distributions. LumaGuide steers the luminance distribution toward a target HDR distribution while preserving scene structure. an SNR-weighted bell-shaped schedule and a time-windowed variant on every reported metric. Appendix discusses this in more detail. 3.4 Extension to Video Generation The same… view at source ↗
Figures from the paper (14 more)
Figure 4
Figure 4. Figure 4: Qualitative comparison with existing HDR generation methods. For fair comparison, we [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: Subjective study. Left: Bradley–Terry JOD scores with 95% CIs. Right: Per-trial rank dis￾tribution. LumaGuide is ranked first in 50% of trials. Highlight-aware distribution shaping [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]
Figure 6
Figure 6. Figure 6: Additional qualitative results across diverse prompts and scenes. [PITH_FULL_IMAGE:figures/full_fig_p019_6.png]
Figure 7
Figure 7. Figure 7: Resolution comparison between 512 × 512 and 1024 × 1024 generation. Higher-resolution results exhibit improved texture detail and spatial coherence while preserving consistent luminance distribution shaping. K-dimensional PQ histograms, ensuring consistency with the bi…
Figure 8
Figure 8. Figure 8: Luminance histogram alignment in PQ space. [PITH_FULL_IMAGE:figures/full_fig_p021_8.png]
Figure 9
Figure 9. Figure 9: Effect of guidance strength s0 on luminance distribution and visual appearance. Each row corresponds to a different guidance scale, and columns show representative scenes under increasing exposure levels [PITH_FULL_IMAGE:figures/full_fig_p022_9.png]
Figure 10
Figure 10. Figure 10: Qualitative results on SD3 and luminance distribution matching at [PITH_FULL_IMAGE:figures/full_fig_p025_10.png]
Figure 11
Figure 11. Figure 11: Qualitative results on SDXL and luminance distribution matching at [PITH_FULL_IMAGE:figures/full_fig_p026_11.png]
Figure 12
Figure 12. Figure 12: Qualitative video examples comparing vanilla CogVideoX, [PITH_FULL_IMAGE:figures/full_fig_p028_12.png]
Figure 13
Figure 13. Figure 13: Distribution alignment of generated video under three representative prompts. Top row: [PITH_FULL_IMAGE:figures/full_fig_p028_13.png]
Figure 14
Figure 14. Figure 14: Per-pixel temporal standard deviation in highlight regions (PQ [PITH_FULL_IMAGE:figures/full_fig_p029_14.png]
Figure 15
Figure 15. Figure 15: Illustration of the mismatch between Q-Eval quality scores and perceptual HDR quality. [PITH_FULL_IMAGE:figures/full_fig_p030_15.png]
Figure 16
Figure 16. Figure 16: Failure case involving saturation-induced unguidable regions. [PITH_FULL_IMAGE:figures/full_fig_p031_16.png]
Figure 17
Figure 17. Figure 17: Failure case involving target distribution infeasibility. [PITH_FULL_IMAGE:figures/full_fig_p031_17.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

43 extracted references · 9 linked inside Pith

  1. [1]

    Wasserstein generative adversarial networks

    Martin Arjovsky, Soumith Chintala, and Léon Bottou. Wasserstein generative adversarial networks. InInternational Conference on Machine Learning (ICML), pages 214–223, 2017

    2017

  2. [2]

    Pu21: A novel perceptually uniform encoding for adapting existing quality metrics for hdr

    Maryam Azimi et al. Pu21: A novel perceptually uniform encoding for adapting existing quality metrics for hdr. In2021 Picture Coding Symposium (PCS), pages 1–5. IEEE, 2021

    2021

  3. [3]

    Universal guidance for diffusion models

    Arpit Bansal, Hong-Min Chu, Avi Schwarzschild, Soumyadip Sengupta, Micah Goldblum, Jonas Geiping, and Tom Goldstein. Universal guidance for diffusion models. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 843–852, 2023

    2023

  4. [4]

    Bracket diffusion: Hdr image generation by consistent ldr denoising

    Mojtaba Bemana, Thomas Leimkühler, Karol Myszkowski, Hans-Peter Seidel, and Tobias Ritschel. Bracket diffusion: Hdr image generation by consistent ldr denoising. InComputer Graphics Forum, volume 44, page e70086. Wiley Online Library, 2025

    2025

  5. [5]

    FLUX.1: Text-to-image generation models

    Black Forest Labs. FLUX.1: Text-to-image generation models. https://github.com/ black-forest-labs/flux, 2024

    2024

  6. [6]

    Align your latents: High-resolution video synthesis with latent diffusion models

    Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 22563–22575, 2023

    2023

  7. [7]

    Diffu- sion posterior sampling for general noisy inverse problems.arXiv preprint arXiv:2209.14687, 2022

    Hyungjin Chung, Jeongsol Kim, Michael T Mccann, Marc L Klasky, and Jong Chul Ye. Diffu- sion posterior sampling for general noisy inverse problems.arXiv preprint arXiv:2209.14687, 2022

    arXiv 2022

  8. [8]

    Diffusion models beat gans on image synthesis

    Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780–8794, 2021

    2021

Show all 43 references
  1. [9]

    Tweedie’s formula and selection bias.Journal of the American Statistical Association, 106(496):1602–1614, 2011

    Bradley Efron. Tweedie’s formula and selection bias.Journal of the American Statistical Association, 106(496):1602–1614, 2011

    2011

  2. [10]

    Hdr image reconstruction from a single exposure using deep cnns.ACM transactions on graphics (TOG), 36(6):1–15, 2017

    Gabriel Eilertsen, Joel Kronander, Gyorgy Denes, Rafał K Mantiuk, and Jonas Unger. Hdr image reconstruction from a single exposure using deep cnns.ACM transactions on graphics (TOG), 36(6):1–15, 2017

    2017

  3. [11]

    Scaling rectified flow trans- formers for high-resolution image synthesis

    Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow trans- formers for high-resolution image synthesis. InForty-first international conference on machine...

    2024

  4. [12]

    Generative adversarial networks.Communications of the ACM, 63(11):139–144, 2020

    Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks.Communications of the ACM, 63(11):139–144, 2020

    2020

  5. [13]

    Denoising diffusion probabilistic models.Advances in neural information processing systems, 33:6840–6851, 2020

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models.Advances in neural information processing systems, 33:6840–6851, 2020

    2020

  6. [14]

    Classifier-free diffusion guidance.arXiv preprint arXiv:2207.12598, 2022

    Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance.arXiv preprint arXiv:2207.12598, 2022

    2022 arXiv

  7. [15]

    Video diffusion models.Advances in neural information processing systems, 35:8633–8646, 2022

    Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models.Advances in neural information processing systems, 35:8633–8646, 2022. 10

    2022

  8. [16]

    Lora: Low-rank adaptation of large language models.Iclr, 1(2):3, 2022

    Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Liang Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models.Iclr, 1(2):3, 2022

    2022

  9. [17]

    International Telecommunication Union. Recommendation ITU-R BT.2100-3: Image parameter values for high dynamic range television for use in production and international programme exchange.https://www.itu.int/rec/R-REC-BT.2100, February 2025

    2025

  10. [18]

    Analyzing and improving the image quality of stylegan

    Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8110–8119, 2020

    2020

  11. [19]

    Noise2Score: Tweedie’s approach to self-supervised image denoising without clean images

    Kwanyoung Kim and Jong Chul Ye. Noise2Score: Tweedie’s approach to self-supervised image denoising without clean images. InAdvances in Neural Information Processing Systems (NeurIPS), 2021

    2021

  12. [20]

    Learning blind video temporal consistency

    Wei-Sheng Lai, Jia-Bin Huang, Oliver Wang, Eli Shechtman, Ersin Yumer, and Ming-Hsuan Yang. Learning blind video temporal consistency. InProceedings of the European conference on computer vision (ECCV), pages 170–185, 2018

    2018

  13. [21]

    Hdr-vdp-2: A calibrated visual metric for visibility and quality predictions in all luminance conditions.ACM Transactions on graphics (TOG), 30(4):1–14, 2011

    Rafat Mantiuk, Kil Joong Kim, Allan G Rempel, and Wolfgang Heidrich. Hdr-vdp-2: A calibrated visual metric for visibility and quality predictions in all luminance conditions.ACM Transactions on graphics (TOG), 30(4):1–14, 2011

    2011

  14. [22]

    Deep hdr hallucination for inverse tone mapping.Sensors, 21(12):4032, 2021

    Demetris Marnerides, Thomas Bashford-Rogers, and Kurt Debattista. Deep hdr hallucination for inverse tone mapping.Sensors, 21(12):4032, 2021

    2021

  15. [23]

    Ex- pandnet: A deep convolutional neural network for high dynamic range expansion from low dynamic range content

    Demetris Marnerides, Thomas Bashford-Rogers, Jonathan Hatchett, and Kurt Debattista. Ex- pandnet: A deep convolutional neural network for high dynamic range expansion from low dynamic range content. InComputer graphics forum, volume 37, pages 37–49. Wiley Online Library, 2018

    2018

  16. [24]

    Energy-based cross attention for bayesian context update in text-to-image diffusion models.Advances in Neural Information Processing Systems, 36:76382–76408, 2023

    Geon Yeong Park, Jeongsol Kim, Beomsu Kim, Sang Wan Lee, and Jong Chul Ye. Energy-based cross attention for bayesian context update in text-to-image diffusion models.Advances in Neural Information Processing Systems, 36:76382–76408, 2023

    2023

  17. [25]

    Sdxl: Improving latent diffusion models for high-resolution image synthesis.arXiv preprint arXiv:2307.01952, 2023

    Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis.arXiv preprint arXiv:2307.01952, 2023

    2023 arXiv

  18. [26]

    Hierarchical text-conditional image generation with clip latents.arXiv preprint arXiv:2204.06125, 1(2):3, 2022

    Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents.arXiv preprint arXiv:2204.06125, 1(2):3, 2022

    2022 arXiv

  19. [27]

    Zero-shot text-to-image generation

    Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea V oss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. InInternational conference on machine learning, pages 8821–8831. Pmlr, 2021

    2021

  20. [28]

    Morgan Kaufmann, 2010

    Erik Reinhard, Wolfgang Heidrich, Paul Debevec, Sumanta Pattanaik, Greg Ward, and Karol Myszkowski.High dynamic range imaging: acquisition, display, and image-based lighting. Morgan Kaufmann, 2010

    2010

  21. [29]

    High- resolution image synthesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High- resolution image synthesis with latent diffusion models. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684–10695, 2022

    2022

  22. [30]

    Photorealistic text-to-image diffusion models with deep language understanding.Advances in neural information processing systems, 35:36479–36494, 2022

    Chitwan Saharia, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding.Advances in neural information ...

    2022

  23. [31]

    Bovik, Neil Birkbeck, Yilin Wang, and Balu Adsumilli

    Shreshth Saini, Alan C. Bovik, Neil Birkbeck, Yilin Wang, and Balu Adsumilli. Chug: Crowd- sourced user-generated hdr video quality dataset. In2025 IEEE International Conference on Image Processing (ICIP), pages 2504–2509, 2025. 11

    2025

  24. [32]

    Shreshth Saini, Bowen Chen, Neil Birkbeck, Yilin Wang, Balu Adsumilli, and Alan C. Bovik. Seeing beyond8bits: Subjective and objective quality assessment of hdr-ugc videos.arXiv preprint arXiv:2603.00938, 2026

    2026

  25. [33]

    Shreshth Saini, Bowen Chen, Yilin Wang, Neil Birkbeck, Balu Adsumilli, and Alan C. Bovik. Brightrate: Quality assessment for user-generated hdr videos. InProceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 1522–1532, March 2026

    2026

  26. [34]

    Shreshth Saini, Shashank Gupta, and Alan C. Bovik. Rectified-cfg++ for flow based models. In Advances in Neural Information Processing Systems (NeurIPS), 2025

    2025

  27. [35]

    Score-based generative modeling through stochastic differential equations.arXiv preprint arXiv:2011.13456, 2020

    Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations.arXiv preprint arXiv:2011.13456, 2020

    2011 arXiv

  28. [36]

    Qwen2.5 technical report.arXiv preprint arXiv:2412.15115, 2024

    Qwen Team. Qwen2.5 technical report.arXiv preprint arXiv:2412.15115, 2024

    2024 arXiv

  29. [37]

    Lediff: Latent exposure diffusion for hdr generation

    Chao Wang, Zhihao Xia, Thomas Leimkuhler, Karol Myszkowski, and Xuaner Zhang. Lediff: Latent exposure diffusion for hdr generation. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 453–464, 2025

    2025

  30. [38]

    X2hdr: Hdr image generation in a perceptually uniform space.arXiv preprint arXiv:2602.04814, 2026

    Ronghuan Wu, Wanchao Su, Kede Ma, Jing Liao, and Rafał K Mantiuk. X2hdr: Hdr image generation in a perceptually uniform space.arXiv preprint arXiv:2602.04814, 2026

    2026

  31. [39]

    Cogvideox: Text-to-video diffusion models with an expert transformer.arXiv preprint arXiv:2408.06072, 2024

    Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, et al. Cogvideox: Text-to-video diffusion models with an expert transformer.arXiv preprint arXiv:2408.06072, 2024

    2024 arXiv

  32. [40]

    Scaling autoregressive models for content-rich text-to-image generation.arXiv preprint arXiv:2206.10789, 2(3):5, 2022

    Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, et al. Scaling autoregressive models for content-rich text-to-image generation.arXiv preprint arXiv:2206.10789, 2(3):5, 2022

    2022 arXiv

  33. [41]

    Freedom: Training- free energy-guided conditional diffusion model

    Jiwen Yu, Yinhuai Wang, Chen Zhao, Bernard Ghanem, and Jian Zhang. Freedom: Training- free energy-guided conditional diffusion model. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 23174–23184, 2023

    2023

  34. [42]

    Diffhdr: Re-exposing ldr videos with video diffusion models.arXiv preprint arXiv:2604.06161, 2026

    Zhengming Yu, Li Ma, Mingming He, Leo Isikdogan, Yuancheng Xu, Dmitriy Smirnov, Pablo Salamanca, Dao Mi, Pablo Delgado, Ning Yu, et al. Diffhdr: Re-exposing ldr videos with video diffusion models.arXiv preprint arXiv:2604.06161, 2026

    2026 arXiv

  35. [43]

    Q-eval-100k: Evaluating visual quality and alignment level for text-to-vision content

    Zicheng Zhang, Tengchuan Kou, Shushi Wang, Chunyi Li, Wei Sun, Wei Wang, Xiaoyu Li, Zongyu Wang, Xuezhi Cao, Xiongkuo Min, et al. Q-eval-100k: Evaluating visual quality and alignment level for text-to-vision content. InProceedings of the IEEE/CVF Conference on Computer Vision ...

    2025

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.