arxiv: 2604.11788 · v1 · submitted 2026-04-13 · 💻 cs.CV

Recognition: unknown

HDR Video Generation via Latent Alignment with Logarithmic Encoding

Naomi Ken Korem , Mohamed Oumoumad , Harel Cain , Matan Ben Yosef , Urska Jelercic , Ofir Bibi , Yaron Inger , Or Patashnik

show 1 more author

Daniel Cohen-Or

Authors on Pith no claims yet

Pith reviewed 2026-05-10 15:26 UTC · model grok-4.3

classification 💻 cs.CV

keywords HDR video generationlogarithmic encodinglatent space alignmentpretrained generative modelscamera-mimicking degradationshigh dynamic rangevideo synthesis

0 comments

The pith

Logarithmic encoding aligns HDR imagery with the latent space of pretrained generative models, enabling lightweight fine-tuning for high-quality HDR video generation.

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

The paper shows that HDR video generation does not require building new representations from scratch. Instead, a logarithmic encoding already used in film pipelines can map high-dynamic-range scenes into a distribution that fits the latent space of existing pretrained video models. This alignment supports direct adaptation through minimal fine-tuning without retraining any encoder. A separate training step that applies camera-like degradations encourages the model to use its learned priors to fill in unobserved bright or dark details. If the approach holds, generative models can handle fundamentally different radiance data without major redesigns, simply by choosing a representation that matches their existing priors.

Core claim

The central claim is that a logarithmic encoding maps HDR imagery into a distribution naturally aligned with the latent space of pretrained generative models. This alignment permits direct adaptation via lightweight fine-tuning without retraining an encoder. When combined with a training strategy that applies camera-mimicking degradations, the model learns to infer missing high-dynamic-range content from its priors, producing high-quality HDR video across diverse scenes and challenging lighting conditions.

What carries the argument

Logarithmic encoding that remaps HDR radiance values into a distribution matching pretrained model latents, paired with camera-mimicking degradations that force inference of unobserved content from visual priors.

If this is right

Pretrained video models can generate HDR footage after only lightweight fine-tuning once the input is logarithmically encoded.
Camera-mimicking degradations during training enable the model to recover high-dynamic-range details not present in the input.
The method produces strong results on diverse scenes and challenging lighting without redesigning the generative architecture.
HDR can be handled effectively by aligning representations to existing priors rather than learning entirely new ones.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The same log-alignment step could simplify HDR adaptation for still-image generators or editing tools that share similar latent spaces.
Real captured HDR video sequences could serve as a direct test set to check whether the simulated degradations match actual camera behavior.
Other perceptual encodings used in imaging pipelines might offer similar alignment benefits for radiance data in generative tasks.

Load-bearing premise

That logarithmic encoding naturally aligns HDR distributions with the latent space of pretrained models and that the models' priors can reliably infer unobserved high dynamic range content when trained with camera-mimicking degradations.

What would settle it

Fine-tune the same pretrained video model on identical HDR data but replace the logarithmic encoding with a linear or standard tone-mapped representation; if the outputs show clipping, color shifts, or loss of detail in extreme lighting, the alignment claim is weakened.

Figures

Figures reproduced from arXiv: 2604.11788 by Daniel Cohen-Or, Harel Cain, Matan Ben Yosef, Mohamed Oumoumad, Naomi Ken Korem, Ofir Bibi, Or Patashnik, Urska Jelercic, Yaron Inger.

**Figure 1.** Figure 1: LumiVid: HDR video generation from SDR input. Given an input SDR video (top), our method produces an HDR video (bottom), enhancing dynamic range and recovering fine details in challenging regions. Abstract High dynamic range (HDR) imagery offers a rich and faithful representation of scene radiance, but remains challenging for generative models due to its mismatch with the bounded, perceptually compressed d… view at source ↗

**Figure 2.** Figure 2: LumiVid Training Overview. Scene-linear HDR frames are compressed via LogC3 and encoded by the frozen VAE to produce target latents ztgt. The same HDR frames are tonemapped to SDR and degraded ( MP4 compression, blur, contrast) to produce reference latents zref. Both are concatenated and fed to the DiT with LoRA adapters; only the LoRA weights are trained via flow matching loss view at source ↗

**Figure 3.** Figure 3: LumiVid Inference Overview. An SDR video is VAE-encoded to zref, concatenated with noise, and denoised by the DiT+LoRA. The output latents are VAE-decoded and decompressed via LogC3−1 to produce scene-linear float16 EXR. The VAE and DiT remain frozen throughout; only the LoRA adapters (<1% parameters) are trained. 3 Method High dynamic range (HDR) video represents scene radiance in an unbounded, linear spa… view at source ↗

**Figure 4.** Figure 4: HDR transform analysis. (a) Pixel-space distributions after compression, overlaid on the SDR prior (gray) and raw HDR input (purple, extending beyond 1.0). LogC3 and PQ produce distributions closest to SDR; ACES collapses near zero; HLG overflows. (b) Latent-space distributions after VAE encoding. HLG’s wide tails cause reconstruction failure. (c) Measured VAE roundtrip error (compress → encode → decode → … view at source ↗

**Figure 5.** Figure 5: Figure 5: LumiVid qualitative results: For each scene, we compare the input SDR (left) with view at source ↗

**Figure 6.** Figure 6: HDR generation comparison. Comparison of LumiVid against state-of-the-art baselines view at source ↗

read the original abstract

High dynamic range (HDR) imagery offers a rich and faithful representation of scene radiance, but remains challenging for generative models due to its mismatch with the bounded, perceptually compressed data on which these models are trained. A natural solution is to learn new representations for HDR, which introduces additional complexity and data requirements. In this work, we show that HDR generation can be achieved in a much simpler way by leveraging the strong visual priors already captured by pretrained generative models. We observe that a logarithmic encoding widely used in cinematic pipelines maps HDR imagery into a distribution that is naturally aligned with the latent space of these models, enabling direct adaptation via lightweight fine-tuning without retraining an encoder. To recover details that are not directly observable in the input, we further introduce a training strategy based on camera-mimicking degradations that encourages the model to infer missing high dynamic range content from its learned priors. Combining these insights, we demonstrate high-quality HDR video generation using a pretrained video model with minimal adaptation, achieving strong results across diverse scenes and challenging lighting conditions. Our results indicate that HDR, despite representing a fundamentally different image formation regime, can be handled effectively without redesigning generative models, provided that the representation is chosen to align with their learned priors.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

The paper gives a practical recipe for adapting pretrained video generators to HDR output using log encoding plus degradation training, but the alignment benefit is asserted more than measured.

read the letter

The core idea here is to take an existing video model, encode HDR frames with a log transform that cinema already uses, then fine-tune lightly while training on camera-like degradations so the model hallucinates the missing dynamic range from its priors. That combination is the actual new piece: not a new architecture, but a targeted representation trick plus a training hack that avoids full encoder retraining or new HDR datasets from scratch. If the experiments hold up, it is a low-friction way to get HDR video out of models that were never trained on it.

Referee Report

1 major / 1 minor

Summary. The manuscript proposes a method for generating high dynamic range (HDR) videos by applying logarithmic encoding to map HDR content into the latent space of pretrained video generative models. This alignment allows for adaptation through lightweight fine-tuning without retraining the encoder. Additionally, a training strategy using camera-mimicking degradations is introduced to enable the model to infer unobserved high dynamic range details from its priors. The authors claim this leads to high-quality HDR video generation across diverse scenes and challenging lighting conditions with minimal adaptation.

Significance. If the central claims hold, this work could have significant impact by demonstrating that existing generative models can be adapted for HDR tasks without substantial redesign, leveraging established cinematic practices like log encoding. This simplifies the pipeline and reduces data requirements, which is valuable for the computer vision and graphics communities working on video synthesis and HDR imaging. The approach credits the strength of pretrained priors and provides a practical alternative to learning new representations from scratch.

major comments (1)

[Abstract] The central claim in the abstract that logarithmic encoding 'maps HDR imagery into a distribution that is naturally aligned with the latent space of these models' enabling 'direct adaptation via lightweight fine-tuning without retraining an encoder' is load-bearing but unsupported by any direct verification. No latent-space statistics (e.g., distribution moments, histograms, or divergence measures such as KL divergence between log-encoded HDR frames and the pretrained model's training distribution) are reported to confirm the alignment assumption versus the possibility that camera-mimicking degradations are doing the compensatory work.

minor comments (1)

[Abstract] The abstract states that the method achieves 'strong results' and 'high-quality HDR video generation' but provides no quantitative metrics, ablation details, or baseline comparisons, making it difficult to evaluate the practical gains over existing approaches.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the constructive feedback and for recognizing the potential impact of our approach. We address the major comment point-by-point below and will revise the manuscript accordingly to strengthen the presentation of our central claim.

read point-by-point responses

Referee: [Abstract] The central claim in the abstract that logarithmic encoding 'maps HDR imagery into a distribution that is naturally aligned with the latent space of these models' enabling 'direct adaptation via lightweight fine-tuning without retraining an encoder' is load-bearing but unsupported by any direct verification. No latent-space statistics (e.g., distribution moments, histograms, or divergence measures such as KL divergence between log-encoded HDR frames and the pretrained model's training distribution) are reported to confirm the alignment assumption versus the possibility that camera-mimicking degradations are doing the compensatory work.

Authors: We agree that direct statistical verification of the distribution alignment would strengthen the manuscript and make the claim more robust. The original submission relied on the well-established use of logarithmic encoding in cinematic pipelines (which is designed to produce perceptually uniform representations within bounded bit depths) together with the empirical success of lightweight fine-tuning as indirect evidence. However, we acknowledge that this leaves open the question of whether the degradations alone could suffice. In the revised version, we will add a dedicated analysis subsection (likely in Section 3 or 4) that includes: (1) histograms and first- and second-order moments of log-encoded HDR frames versus frames from the pretrained model's training distribution; (2) a quantitative comparison using a divergence measure such as approximated KL divergence or Wasserstein distance between the two distributions; and (3) an expanded ablation that trains the same model with linear (instead of log) encoding while keeping the degradation pipeline identical. These additions will allow readers to assess the relative contributions of the encoding choice versus the degradation strategy. We believe the new analysis will confirm that the log mapping is the key enabler of the observed alignment and minimal adaptation. revision: yes

Circularity Check

0 steps flagged

No circularity; derivation relies on external pretrained priors and empirical representation choice

full rationale

The paper's central steps—selecting logarithmic encoding to align HDR with pretrained latent spaces, then applying lightweight fine-tuning plus camera-mimicking degradations—rest on external pretrained generative models (whose priors are independent of this work) and an observational claim about cinematic log encoding. No equations or claims reduce by construction to fitted parameters, self-definitions, or self-citation chains. The alignment is presented as an empirical observation enabling adaptation, not as a derived result from the paper's own outputs. This is a standard non-circular use of prior models.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

The central claim rests on the domain assumption that log encoding produces natural alignment and that pretrained priors suffice for HDR inference; no free parameters or invented entities are explicitly introduced in the abstract.

axioms (2)

domain assumption Logarithmic encoding maps HDR imagery into a distribution naturally aligned with pretrained generative model latent spaces
Presented as an observation enabling direct adaptation without encoder retraining.
domain assumption Pretrained visual priors can infer missing HDR content when trained with camera-mimicking degradations
Core to the training strategy for recovering unobservable details.

pith-pipeline@v0.9.0 · 5548 in / 1266 out tokens · 61512 ms · 2026-05-10T15:26:44.962792+00:00 · methodology

Review history (2 revisions) →

discussion (0)

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

Generating HDR Video from SDR Video
cs.CV 2026-05 unverdicted novelty 7.0

A multi-exposure video model predicts bracketed linear SDR sequences from single nonlinear SDR input, which a merging model combines into HDR video preserving shadow and highlight detail.
Single-Shot HDR Recovery via a Video Diffusion Prior
cs.CV 2026-05 unverdicted novelty 7.0

Single-shot HDR is achieved by conditioning a video diffusion model on an LDR input to generate an exposure bracket and fusing the bracket with per-pixel weights from a lightweight UNet.

Reference graph

Works this paper leans on

42 extracted references · 6 canonical work pages · cited by 2 Pith papers · 3 internal anchors

[1]

ActionVFX stock footage library

ActionVFX. ActionVFX stock footage library. https://www.actionvfx.com, 2026. Ac- cessed: 2026-04-13. 9

2026
[2]

Camera sample footage and reference images

ARRI. Camera sample footage and reference images. https: //www.arri.com/en/learn-help/learn-help-camera-system/ camera-sample-footage-reference-image, 2026. Accessed: 2026-04-13

2026
[3]

Bracket diffusion: HDR image generation by consistent LDR denoising.Computer Graphics F orum, 2025

Mojtaba Bemana, Karol Myszkowski, Hans-Peter Seidel, and Tobias Ritschel. Bracket diffusion: HDR image generation by consistent LDR denoising.Computer Graphics F orum, 2025

2025
[4]

Avcontrol: Efficient framework for training audio-visual controls.arXiv preprint arXiv:2603.24793, 2026

Matan Ben-Yosef, Tavi Halperin, Naomi Ken Korem, Mohammad Salama, Harel Cain, Asaf Joseph, Anthony Chen, Urska Jelercic, and Ofir Bibi. Avcontrol: Efficient framework for training audio-visual controls.arXiv preprint arXiv:2603.24793, 2026

work page arXiv 2026
[5]

Stable Video Diffusion: Scaling Latent Video Diffusion Models to Large Datasets

Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Do- minik Lorber, Yam Levi, Zion English, Vikram V oleti, Adam Letts, Varun Jampani, and Robin Rombach. Stable video diffusion: Scaling latent video diffusion models to large datasets.arXiv preprint arXiv:2311.15127, 2023

work page internal anchor Pith review arXiv 2023
[6]

Tears of steel

Blender Foundation. Tears of steel. https://mango.blender.org/, 2012. Accessed: 2026- 04-13

2012
[7]

Wong, and Lei Zhang

Guanying Chen, Chaofeng Chen, Shi Guo, Zhetong Liang, Kwan-Yee K. Wong, and Lei Zhang. HDR video reconstruction: A coarse-to-fine network and a real-world benchmark dataset. In ICCV, 2021

2021
[8]

Ren, Lynhoo Tian, Yu Qiao, and Chao Dong

Xiangyu Chen, Zhengwen Zhang, Jimmy S. Ren, Lynhoo Tian, Yu Qiao, and Chao Dong. A new journey from SDRTV to HDRTV. InICCV, 2021

2021
[9]

Debevec and Jitendra Malik

Paul E. Debevec and Jitendra Malik. Recovering high dynamic range radiance maps from photographs. InACM SIGGRAPH, pages 369–378, 1997

1997
[10]

Mantiuk, and Jonas Unger

Gabriel Eilertsen, Joel Kronander, Gyorgy Denes, Rafał K. Mantiuk, and Jonas Unger. HDR image reconstruction from a single exposure using deep CNNs. InACM SIGGRAPH Asia, 2017

2017
[11]

Edge-preserving decom- positions for multi-scale tone and detail manipulation

Zeev Farbman, Raanan Fattal, Dani Lischinski, and Richard Szeliski. Edge-preserving decom- positions for multi-scale tone and detail manipulation. InACM SIGGRAPH, pages 67:1–67:10, 2008

2008
[12]

Gradient domain high dynamic range compression

Raanan Fattal, Dani Lischinski, and Michael Werman. Gradient domain high dynamic range compression. InACM SIGGRAPH, pages 249–256, 2002

2002
[13]

Yilun Guan, Yuhan Zhu, and Rafał K. Mantiuk. HDR-V-Diff: Latent diffusion-based HDR video reconstruction. InECCV Workshops, 2024

2024
[14]

Yilun Guan, Yuhan Zhu, and Rafał K. Mantiuk. GM-Diffusion: Gain-map guided diffusion model for high dynamic range image generation. InICCV, 2025

2025
[15]

LTX-Video: Realtime Video Latent Diffusion

Yoav HaCohen, Nisan Chiprut, Benny Brazowski, Daniel Shalem, Dudu Moshe, Eitan Richard- son, Eran Levin, Guy Shiran, Nir Zabari, Ori Gordon, Poriya Panet, Sapir Weissbuch, Victor Kulikov, Yaki Bitterman, Zeev Melumian, and Ofir Bibi. LTX-Video: Realtime video latent diffusion.arXiv preprint arXiv:2501.00103, 2024

work page internal anchor Pith review arXiv 2024
[16]

Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J. Fleet. Video diffusion models. InNeurIPS, 2022

2022
[17]

Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen

Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. InICLR, 2022

2022
[18]

Deep HDR video from sequences with alternating exposures

Nima Khademi Kalantari and Ravi Ramamoorthi. Deep HDR video from sequences with alternating exposures. InEurographics, 2019

2019
[19]

LuxDiT: Lighting estimation with video diffusion transformer

Zhihao Liang et al. LuxDiT: Lighting estimation with video diffusion transformer. InAdvances in Neural Information Processing Systems (NeurIPS), 2025

2025
[20]

GMNet: Learning gain map for inverse tone mapping

Jiyuan Liao et al. GMNet: Learning gain map for inverse tone mapping. InInternational Conference on Learning Representations (ICLR), 2025

2025
[21]

Mantiuk and Maliha Ashraf

Rafał K. Mantiuk and Maliha Ashraf. PU21: A novel perceptually uniform encoding for adapting existing quality metrics for HDR. InPicture Coding Symposium (PCS), 2021

2021
[22]

2304.13625 , archivePrefix=

Rafał K. Mantiuk, Rajitha Weerakkody, Saverio Andriani, and Marta Mrak. HDR-VDP-3: A multi-metric for predicting image differences, quality and contrast distortions in high dynamic range and normal content.arXiv preprint arXiv:2304.13625, 2023. 10

work page arXiv 2023
[23]

Mantiuk, Aliaksei Chubarau, et al

Rafał K. Mantiuk, Aliaksei Chubarau, et al. ColorVideoVDP: A visual difference predictor for image, video and display distortions. InACM SIGGRAPH, 2024

2024
[24]

Ex- pandNet: A deep convolutional neural network for high dynamic range expansion from low dynamic range content.Computer Graphics F orum, 2018

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.Computer Graphics F orum, 2018

2018
[25]

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
[26]

Aliaksei Mikhailiuk, Maria Perez-Ortiz, Dingcheng Yue, Wilson Suen, and Rafał K. Mantiuk. Consolidated dataset and metrics for high-dynamic-range image quality.IEEE Transactions on Multimedia, 24:2125–2138, 2021

2021
[27]

Scalable diffusion models with transformers

William Peebles and Saining Xie. Scalable diffusion models with transformers. InICCV, 2023

2023
[28]

Poly haven hdri library

Poly Haven. Poly haven hdri library. https://polyhaven.com/hdris, 2026. Accessed: 2026-04-13

2026
[29]

Photographic tone reproduc- tion for digital images

Erik Reinhard, Michael Stark, Peter Shirley, and James Ferwerda. Photographic tone reproduc- tion for digital images. InACM SIGGRAPH, pages 267–276, 2002

2002
[30]

Real-HDRV: A large-scale high-quality dataset for HDR video reconstruction

Hao Shu, Chenming Liu, Kai Zhong, Wenhao Wu, and Tong Lu. Real-HDRV: A large-scale high-quality dataset for HDR video reconstruction. InCVPR, 2024

2024
[31]

RoFormer: Enhanced transformer with rotary position embedding.Neurocomputing, 568:127063, 2024

Jianlin Su, Yu Lu, Shengfeng Pan, Ahmed Murtadha, Bo Wen, and Yunfeng Liu. RoFormer: Enhanced transformer with rotary position embedding.Neurocomputing, 568:127063, 2024

2024
[32]

Wan2.1: A comprehensive and multitask video foundation model.arXiv preprint, 2025

Ang Wang, Zhiyu Gao, Cong Chen, Teng Xiao, Di He, et al. Wan2.1: A comprehensive and multitask video foundation model.arXiv preprint, 2025

2025
[33]

GlowGAN: Unsupervised learning of HDR images from LDR images in the wild

Chao Wang, Ana Serrano, Xingang Pan, Dongdong Chen, Guillaume Gilet, Christian Theobalt, Karol Myszkowski, and Thomas Leimkühler. GlowGAN: Unsupervised learning of HDR images from LDR images in the wild. InICCV, 2023

2023
[34]

LEDiff: Latent exposure diffusion for HDR generation

Chao Wang, Ana Serrano, Xingang Pan, Karol Myszkowski, and Thomas Leimkühler. LEDiff: Latent exposure diffusion for HDR generation. InCVPR, 2025

2025
[35]

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, 2025

work page arXiv 2025
[36]

HDRFlow: Real-time HDR video reconstruction with large motions

Gangwei Xu, Hao Ye, Xinyu Wang, Junda Liu, and Xin Xu. HDRFlow: Real-time HDR video reconstruction with large motions. InCVPR, 2024

2024
[37]

Dual frequency transformer for efficient SDR-to-HDR translation.Machine Intelligence Research, 2024

Jie Xu, Mingze Jiang, Mao Ye, and Yanwei Pang. Dual frequency transformer for efficient SDR-to-HDR translation.Machine Intelligence Research, 2024

2024
[38]

Toward high-quality HDR deghosting with conditional diffusion models.IEEE TCSVT, 2023

Qingsen Yan, Tao Hu, Yuan Sun, Hao Tang, Yu Zhu, Wei Dong, Luc Van Gool, and Yanning Zhang. Toward high-quality HDR deghosting with conditional diffusion models.IEEE TCSVT, 2023

2023
[39]

HDRev-Net: Event-guided high dynamic range video reconstruction

Yunhao Yang, Jin Ye, Jingye He, Conghui Wang, and Yizhen Han. HDRev-Net: Event-guided high dynamic range video reconstruction. InCVPR, 2023

2023
[40]

CogVideoX: Text-to-Video Diffusion Models with An Expert Transformer

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

work page internal anchor Pith review arXiv 2024
[41]

Video inverse tone mapping with temporal clues

Hao Ye, Gangwei Xu, Xinyu Wang, Junda Liu, and Xin Xu. Video inverse tone mapping with temporal clues. InCVPR, 2024

2024
[42]

Yuhan Zhu, Yilun Guan, and Rafał K. Mantiuk. Zero-shot high dynamic range imaging via diffusion models. InCVPR, 2024. Highlight. 11

2024