LiBrA-Net: Lie-Algebraic Bilateral Affine Fields for Real-Time 4K Video Dehazing
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.
Tom: We're looking at the paper: "LiBrA-Net: Lie-Algebraic Bilateral Affine Fields for Real-Time 4K Video Dehazing".
Jane: Here's the gist of the paper. LiBrA-Net proposes a novel framework that addresses the limitations of existing video dehazing methods by factorizing spatiotemporal affine fields into low-resolution bilateral sub-grids and fusing them using Lie algebra, achieving state-of-the-art performance on 4K video while maintaining real-time processing speeds.
Tom: Core Problem Addressed The paper tackles the gap in ultra-high-definition (UHD) video dehazing due to a lack of a benchmark and the inability of existing methods to run on consumer GPUs for continuous UHD sequences.
Jane: The key observation is that atmospheric dehazing reduces to a per-pixel affine transform governed by the low-frequency depth field, which can be compactly encoded in bilateral grids whose prediction cost is decoupled from the output resolution.
Tom: Conventional dense-prediction architectures couple spatiotemporal modeling cost to the output resolution, which is the shared root cause of scaling issues.
Paper summary: Tom: So, wrapping up on "LiBrA-Net: Lie-Algebraic Bilateral Affine Fields for Real-Time 4K Video Dehazing <ref:2605.11508#pg1,LiBrA-Net: Lie-Algebraic Bilateral Affine Fields for Real-Time 4K Video>." The authors are focused on solving the benchmark gap and the real-time processing bottleneck for UHD video.
Jane: They achieve this by proposing a novel framework that factorizes the spatiotemporal affine field into low-resolution bilateral sub-grids and fusing them using Lie algebra to restore high-frequency detail through a lightweight input-guided branch.
Lu: The main implication is the resolution decoupling design principle, which shows that modeling spatiotemporal structure at a fixed low resolution and applying the learned transform at the target resolution through trilinear slicing is beneficial for other UHD video tasks with spatially smooth degradation fields.
Meng: For me, it means we have a new baseline for how to approach high-resolution video restoration efficiently when we aren't constrained by needing an infinitely large model just to handle pixel density.
Lalam: They set a state of the art among compared methods, exceeding twenty-three dB PSNR on the UHV-4K benchmark while sustaining twenty-five FPS at native 4K throughput with only six point one two M parameters <ref:2605.11508#pg2>.
Tom: That number is what gets people excited—real-time performance paired with high fidelity on a demanding task like this. It shows that efficiency and quality can coexist better than before in this specific type of problem.
Conclusion: Tom: So we’re looking at LiBrA-Net today, which is "Lie-Algebraic Bilateral Affine Fields for Real-Time 4K Video Dehazing <ref:2605.11508#pg1,Lie-Algebraic Bilateral Affine Fields for Real-Time 4K Video Dehazing>." Basically, they took this problem where existing methods struggle with high definition and real time processing and they built a new structure to handle it.
Jane: The authors are focused on a specific way of looking at the problem—breaking down the overall atmospheric effect into smaller pieces that can be predicted separately. My first question is, how exactly did they manage to break down the 4D spatiotemporal field into these lower resolution grids <ref:2605.11508#pg1>?
Lu: They factorized it. They took the full spatiotemporal affine field and split it into a spatial-color grid and a temporal grid, both of which they predict at a fixed, low resolution—two hundred fifty-six by two hundred fifty-six pixels. That’s the core idea for decoupling the resolution from the model cost.
Meng: Decoupling sounds good on paper, but how does that actually translate to speed when we’re talking about actual deployment? If you're predicting things at a much lower resolution first, what happens when you have to reconstruct that detail back up to four thousand pixels for the final output?
Lalam: They handle the reconstruction using trilinear slicing. That means they take those low-resolution predictions and slice them up to get back to the full resolution input. The fusion part uses a Lie algebra—specifically, aff(three) equals gl(three) plus R3—to combine those two sub-grids.
Tom: Combining them in a Lie algebra... that’s heavy math for video dehazing. Jane, what does that fusion actually achieve compared to just stacking two separate predictions?
Jane: It allows them to map the combined coefficients into an invertible GL(three) transform using something called a Cayley parameterization. That’s how they turn those low-resolution features back into the high-frequency detail we need for a clear image.
Lu: And they use this input-guided branch to restore that high frequency detail, which is key because it's not just guessing texture; it’s using the input data to guide the restoration. They also use group-theoretic regularization there.
Meng: So, what’s the real result here? The benchmark numbers are important for me. Are we talking about something that actually runs smoothly on consumer hardware, or is this still a research toy?
Lalam: They report exceeding twenty-three dB PSNR on the UHV-4K benchmark while maintaining twenty-five frames per second at native four thousand pixels throughput with only six point one two million parameters <ref:2605.11508#pg1>. That’s the number they are pushing.
Tom: Six point one two million parameters for real time 4K processing <ref:2605.11508#pg1>... that’s a serious constraint for any model. Jane, what does this mean for someone who just needs to watch a video on their phone or computer?
Jane: It means the method is highly efficient because the heavy lifting of modeling the atmospheric structure is done at a much smaller resolution first. And it has strong external validity, meaning it works well on real outdoor haze, not just synthetic stuff.
Lu: That’s because that low-frequency transform they learn is governed by smooth atmospheric structure, not scene texture specific details. It transfers well between different types of hazy videos.
Tom: So the authors are saying this resolution-decoupled design principle could be useful for other UHD video tasks where the degradation field is spatially smooth, like maybe deblurring or super-resolution too. Jane, what’s the final word on where they stop?
Jane: They admit their scattering model assumes a single beta and A infinity per video. So, mixed weather conditions—like rain mixed with haze—might fall outside what this specific model is designed to handle best right now.
Lu: Exactly. It's a very clean solution for the core problem they set out to solve, but it has defined boundaries based on the assumptions they made about the atmospheric physics.
Tom: Okay, so it’s a high-performing model that prioritizes efficiency and stability under tight constraints, but we have to watch those limitations regarding mixed weather. Jane, what’s next for this research area?
Jane: The authors are releasing the UHV-4K dataset itself—the first one they've paired with depth, transmission maps, and optical-flow annotations on every frame <ref:2605.11508#pg1>. That should open things up a lot for anyone who wants to test these kinds of methods properly.
Yongcong Wang, Chengchao Shen, Guangwei Gao, Wei Wang, Pengwen Dai, Dianjie Lu, Guijuan Zhang, Zhuoran Zheng
Central South University · Nanjing University of Science and Technology · Sun Yat-sen University · Shandong Normal University · Qilu University of Technology
cs.CV
Submitted: 2026-05-12
Updated: 2026-10-29
Comments: Accepted to NeurIPS 2026. Code: https://github.com/ycwang31/LiBrA-Net
Code: https://github.com/ultralytics/ultralytics
License: http://creativecommons.org/licenses/by/4.0/
Importance score: 90/100
The gist: LiBrA-Net proposes a novel framework that addresses the limitations of existing video dehazing methods by factorizing spatiotemporal affine fields into low-resolution bilateral sub-grids and fusing…
Terminology
Summary
LiBrA-Net proposes a novel framework that addresses the limitations of existing video dehazing methods by factorizing spatiotemporal affine fields into low-resolution bilateral sub-grids and fusing them using Lie algebra, achieving state-of-the-art performance on 4K video while maintaining real-time processing speeds. p. 8 7 Conclusion, Limitations and Future Works
We have presented LiBrA-Net, a bilateral-grid framework that decouples spatiotemporal modeling
from output resolution for real-time 4K video dehazing. The method predicts affine color transforms
on compact sub-grids at fixed resolution and applies them at native 4K through trilinear slicing,
achieving 25 FPS with 6.Where the paper says it
Core Problem Addressed
The paper tackles the gap in ultra-high-definition (UHD) video dehazing due to a lack of a benchmark and the inability of existing methods to run on consumer GPUs for continuous UHD sequences. The key observation is that atmospheric dehazing reduces to a per-pixel affine transform governed by the low-frequency depth field, which can be compactly encoded in bilateral grids whose prediction cost is decoupled from the output resolution. p. 0 In this paper, we address both issues with a new
benchmark and an efficient method. Our key observation is that atmospheric dehazing reduces to a per-pixel affine transform governed by the low-frequency depth
field, which can be compactly encoded in bilateral grids whose prediction cost is
decoupled from the output resolution. context that the ASM requires or discard the high-frequency detail that 4K sensors exist to capture. The shared root cause is that conventional dense-prediction architectures couple spatiotemporal
modeling cost to the output resolution.Where the paper says it
Where the paper says it
LiBrA-Net Architecture and Mechanism
LiBrA-Net factorizes the full spatiotemporal affine field into a spatial–color and a temporal sub-grid predicted at a fixed low resolution, fusing their coefficients in the gl(3) Lie algebra under group-theoretic regularization. It maps the result to invertible GL(3) transforms via a Cayley parameterization and restores high-frequency detail through a lightweight input-guided branch. p. 0 Our key observation is that atmospheric dehazing reduces to a per-pixel affine transform governed by the low-frequency depth
field, which can be compactly encoded in bilateral grids whose prediction cost is
decoupled from the output resolution. Building on this, we propose LiBrA-Net,
which factorizes the spatiotemporal affine field into a spatial–color and a temporal
bilateral sub-grid predicted at a fixed low resolution, fuses their coefficients in the
gl(3) Lie algebra under group-theoretic regularization, maps the result to invertible
GL(3) transforms via a Cayley parameterization, and restores high-frequency detail
through a lightweight input-guided branch. 2 Predicting the Two Bilateral Affine Fields
We predict Gχ and Gτ with two lightweight branches that share a stride-16 encoder structure but
use independent weights and consume different inputs, all at a fixed resolution S × S = 256 × 256
regardless of the input size. Chromatic Affine Field.Where the paper says it
Where the paper says it
Resolution Decoupling and Fusion
The method achieves resolution decoupling by modeling spatiotemporal structure at low resolution and applying the learned transform at the target resolution through trilinear slicing. The spatial–color grid Gχ is organized across the full chromatic axis, exploiting all eight color bins rather than collapsing to a depth-only template. p. 8 Three findings emerge. First,
the spatial–color grid exploits the full
chromatic axis across all eight color bins
rather than collapsing to a depth-only template; t-SNE visualization confirms that
grid cells cluster by color bin, indicating
a structured affine descriptor whose Chromatic Field alone accounts for ∼2. The matrix part M is mapped to a per-pixel invertible affine by
the Cayley map,
A = Cay(M) = (I − M/2)−1
(I + M/2), (7)
the [1/1] Padé approximant of exp(M), computed by one closed-form 3 × 3 solve per pixel. Two
properties make this parameterization useful here.Where the paper says it
Where the paper says it
Performance and Validation
LiBrA-Net sets a new state of the art among compared video dehazing methods, exceeding 23 dB PSNR on the UHV-4K benchmark while sustaining 25 FPS at native 4K throughput with only 6.12 M parameters. p. 0 Across UHV-4K, REVIDE, and HazeWorld, LiBrA-Net
sets a new state of the art among compared video dehazing methods while running
native 4K at 25 FPS on a single GPU with only 6. 12 M parameters. Building on this, we propose LiBrA-Net,
which factorizes the spatiotemporal affine field into a spatial–color and a temporal
bilateral sub-grid predicted at a fixed low resolution, fuses their coefficients in the
gl(3) Lie algebra under group-theoretic regularization, maps the result to invertible
GL(3) transforms via a Cayley parameterization, and restores high-frequency detail
through a lightweight input-guided branch. We further release UHV-4K, the first
paired 4K video dehazing benchmark with depth, transmission, and optical-flow annotations on every frame. We feed each method’s 4K output into frozen detection and segmentation
models. LiBrA-Net ranks first on 8 of 10 no-reference metrics, with detection confidence approaching
the clean-frame oracle and segmentation boundaries closest to the ground truth (Appendix G, Table 7).Where the paper says it
Where the paper says it
Where the paper says it
Temporal Consistency and Generalization
Temporal consistency is stabilized by the Temporal Field's self-attention combined with the temporal smoothness component of the Lie temporal term, which regularizes grid trajectories. The framework shows strong external validity, as LiBrA-Net is the only video method that improves all three no-reference image quality assessment metrics simultaneously over raw input on eight real-world 4K hazy videos. p. 8 Real-world generalization. On eight hand-collected 4K hazy videos without ground truth, LiBrANet is the only video method that improves all three NR-IQA scores over raw input; competing
methods degrade when atmospheric statistics deviate from training. On eight hand-collected 4K hazy videos without ground truth, LiBrANet is the only video method that improves all three NR-IQA scores over raw input; competing
methods degrade when atmospheric statistics deviate from training. The bilateral grid’s low-frequency
transform—governed by smooth atmospheric structure rather than scene-specific texture—facilitates
transfer from synthetic to real outdoor haze; qualitative results are in Appendix G.Where the paper says it
Where the paper says it
Conclusion and Future Scope
The resolution-decoupled design principle is shown to be beneficial for other UHD video tasks governed by spatially smooth degradation fields. While the current temporal window covers roughly one second, modeling minute-scale coherence at 4K remains an open area. p. 9 The resolution-decoupled design principle is not specific to dehazing and may benefit
other UHD video tasks governed by spatially smooth degradation fields. Our current temporal window
covers roughly one second; modeling minute-scale coherence at 4K remains open. Our current temporal window
covers roughly one second; modeling minute-scale coherence at 4K remains open. The scattering
model assumes a single β and A∞ per video, so mixed-weather conditions fall outside UHV-4K’s
scope, and paired real 4K hazy/clean data does not yet exist. Building on this, we propose LiBrA-Net,
which factorizes the spatiotemporal affine field into a spatial–color and a temporal
bilateral sub-grid predicted at a fixed low resolution, fuses their coefficients in the
gl(3) Lie algebra under group-theoretic regularization, maps the result to invertible
GL(3) transforms via a Cayley parameterization, and restores high-frequency detail
through a lightweight input-guided branch. We further release UHV-4K, the first
paired 4K video dehazing benchmark with depth, transmission, and optical-flow annotations on every frame.Where the paper says it
Where the paper says it
Where the paper says it
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Adaptive resolution scaling for video dehazing systems: The system can process 4K video at real-time speeds by modeling spatiotemporal structures at a fixed low resolution and applying learned transforms at the target resolution, as shown by
model spatiotemporal structure at low resolution and apply the learned transform at the target resolution.
-
Real-time 4K video dehazing with high speed: The system can achieve 25 FPS on native 4K video using only 6.12 M parameters, surpassing existing methods like HazeWorld and REVIDE in PSNR while maintaining real-time throughput on a single GPU.
-
Enhanced downstream perception: The improved system can drive off-the-shelf YOLOv8-X and SegFormer-B5 close to clean image detection and segmentation quality, leading to better detection confidence and tighter segmentation boundaries for thin structures.
-
Robust temporal coherence: The system maintains high per-frame fidelity while ensuring inter-frame consistency by using a Temporal Field with self-attention, resulting in a lower tOF score compared to other methods while achieving the highest PSNR among compared methods.
-
Generalization to real-world conditions: The system exhibits superior no-reference image quality and better performance on real-world 4K hazy videos without ground truth, due to its bilateral grid's low-frequency transform being governed by smooth atmospheric structure rather than scene-specific texture.
Abstract
Currently, there is a gap in the field of ultra-high-definition (UHD) video dehazing due to the lack of a benchmark for evaluation. Furthermore, existing video dehazing methods cannot run on consumer-grade GPUs when processing continuous UHD sequences of 3--5 frames at a time. In this paper, we address both issues with a new benchmark and an efficient method. Our key observation is that atmospheric dehazing reduces to a per-pixel affine transform governed by the low-frequency depth field, which can be compactly encoded in bilateral grids whose prediction cost is decoupled from the output resolution. Building on this, we propose LiBrA-Net, which factorizes the spatiotemporal affine field into a spatial--color and a temporal bilateral sub-grid predicted at a fixed low resolution, fuses their coefficients in the gl(3) Lie algebra under group-theoretic regularization, maps the result to invertible GL(3) transforms via a Cayley parameterization, and restores high-frequency detail through a lightweight input-guided branch. We further release UHV-4K, the first paired 4K video dehazing benchmark with depth, transmission, and optical-flow annotations on every frame. Across UHV-4K, REVIDE, and HazeWorld, LiBrA-Net sets a new state of the art among compared video dehazing methods while running native 4K at 25FPS on a single GPU with only 6.12M parameters. Code is available at https://github.com/ycwang31/LiBrA-Net.
Sources
- Depth-Centric Dehazing and Depth-Estimation from Real-World Hazy Driving Video
- Embedding Fourier for Ultra-High-Definition Low-Light Image Enhancement
- The 2017 DAVIS Challenge on Video Object Segmentation
- Correlation Matching Transformation Transformers for UHD Image Restoration
- Video Dehazing via a Multi-Range Temporal Alignment Network with Physical Prior
Related papers
- TEA: Text Encoder Alignment for Robust Concept Erasure in Text-to-Image Models
- Do Egocentric Video-Language Models Capture Both Hand- and Object-Centric Cues?
- Aura: Consistent Multi-Subject Video Generation via VLM-Grounded Semantic Alignment
- PVRF: All-in-one Adverse Weather Removal via Prior-modulated and Velocity-constrained Rectified Flow
- DLWM: Diverse Latent World Models for Efficient Multimodal Reasoning
- Out of Sight, Still in Mind: Token Compression for Omni-LLMs