Pith. sign in

REVIEW 3 major objections 4 minor 37 references

Multimodal 3D LUT Generation via StatLUT with Statistical Features for Photorealistic Style Transfer

T0 review · 3 major / 4 minor · reviewed 2026-07-10 · grok-4.5

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

Pith's one-line read Photorealistic style transfer works better when color is treated as Lab statistics and turned into a smooth 3D LUT by a residual Transformer, not by semantic image encoders.

desk verdict Solid engineering fix for semantic entanglement in PST via Lab stats + residual Transformer LUTs; real gains and a clean stress test, but synthetic training and weaker text path keep it incremental. read the letter →

arxiv 2607.08227 v1 pith:F7PWZL5V submitted 2026-07-09 cs.CV

Yifan Wang , Zhixiang Hao , Yu Wang , Congchao Zhu This is my paper
classification cs.CV
keywords photorealisticstyletransfer3DLUTLabstatisticssemanticdecouplingTransformerresidualmappingtext-drivencolorgradinghistogramdiffuser
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

Photorealistic style transfer must change only color and tone while leaving every edge and texture untouched. The paper argues that the usual deep pipeline fails this test because pretrained image encoders mix high-level object meaning into the color map, producing spatial distortions, and because point-wise color maps ignore the lattice of neighboring colors, producing banding. StatLUT replaces the encoder with a Lab-Extractor that records only three spatially-agnostic histograms (luminance, chrominance, and luminance conditioned on color). Those pure statistics condition a Transformer that predicts a residual 3D lookup table whose topology stays smooth. A small diffusion Transformer can also invent the same statistics from a text prompt, so style can be driven by language alone. On standard photo benchmarks the resulting LUTs sit closest to the ideal content–style trade-off, survive deliberate destruction of the style image’s layout, and keep video frames flicker-free.

What carries the argument

The Lab-Extractor + MR-Mapper pair: three soft-binned Lab histograms (HL, Hab, ML|ab) are mapped by residual and global branches into a condition memory that a Transformer decoder uses to predict a residual on an identity 3D LUT grid, guaranteeing a topologically smooth color manifold.

What would settle it

Measure residual maps after style-image patch-shuffling on a held-out set of extreme real photo pairs outside the synthetic LUT manifold; if residuals rise to the level of encoder-based baselines, the claimed semantic immunity and generalization both fail.

Watch

Extended reading notes

Core claim

Semantic entanglement and color banding in photorealistic style transfer are eliminated by discarding spatial image features entirely, representing style solely as Lab statistical distributions, and generating a residual 3D LUT via topology-aware cross-attention; the same statistics can be synthesized from text, yielding both image-driven and text-driven color grading that preserves structure better than prior LUT and encoder methods.

Load-bearing premise

Training on random synthetic LUTs applied to ordinary photos produces color statistics that still match the distributions of real photographic style pairs, including extreme day-to-night or snow-to-neon gaps.

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 / 4 minor

Summary. The paper proposes StatLUT, a multimodal framework for photorealistic style transfer that generates 3D LUTs from spatially-agnostic statistical features extracted in CIE Lab space by a Lab-Extractor (1D luminance histogram, 2D chrominance histogram, and color-conditioned mean luminance). These features condition a Transformer Seq2Seq residual predictor (MR-Mapper) that maps an identity LUT to a topologically smooth target LUT; a lightweight DiT (H-Diffuser) further synthesizes the same statistical features from CLIP text embeddings for text-driven grading. Training is self-supervised by applying random LUTs to COCO images and using spatially augmented stylized results as style references. Experiments on PhotoNAS and PST50 report superior content-style balance (lowest distance-to-ideal), a user study preference of 70% Top-1, near-zero residuals under style patch-shuffling, and competitive text-driven results, with ablations on feature sets, mapper design, and LUT size.

Significance. If the claims hold, StatLUT offers a practical advance for high-resolution and video PST by replacing heavy semantic encoders with compact statistical priors and residual LUT prediction, eliminating banding and spatial artifacts while adding text control without instance-specific fine-tuning. The patch-shuffling experiment and self-supervised paradigm are concrete, falsifiable contributions that strengthen the semantic-decoupling argument; the compact MR-Mapper (0.38 M parameters) and efficient inference (<50 ms image-driven) are engineering strengths that support real-time deployment. These elements, together with the user study and ablations, make the work a useful reference point for LUT-based color grading even if some generalization questions remain open.

major comments (3)
  1. [Sec. 3.2 / Limitations] Sec. 3.2 training paradigm and Limitations: the self-supervised construction (random LUTs applied to COCO content, then spatial augmentations of the result used as style references) is load-bearing for the claimed generalization of Lab statistics and residual Seq2Seq mapping. Because the Lab-Extractor discards spatial co-occurrence, the network never observes real photographic luminance-chrominance bindings outside the synthetic LUT manifold. The Limitations paragraph itself flags extreme content-style gaps as problematic; without a distribution-coverage analysis or additional unpaired real-style experiments, the SOTA distance-to-ideal numbers and patch-shuffling immunity risk being largely in-distribution confirmations rather than proof of robust semantic decoupling on photographic styles.
  2. [Sec. 4.1.2 / Table 1] Sec. 4.1.2, Table 1 and Fig. 5: evaluation relies exclusively on the specialized Neural Preset protocol (discriminator style similarity + LDC-feature SSIM content similarity, normalized distance-to-ideal). While the authors note that PSNR/LPIPS can misalign with PST perception, the absence of any standard perceptual or histogram-matching metrics, and the large drop for the text-driven variant (style sim 0.498 vs 0.746 on NAS), make it difficult to situate the gains relative to the broader literature or to quantify the cross-modal gap. A small set of conventional metrics or an explicit calibration against human rankings would strengthen the central quantitative claim.
  3. [Sec. 4.1.3] Sec. 4.1.3 and Fig. 6: the text-driven mode is presented as pioneering multimodal PST, yet no baselines (even simple CLIP-conditioned colorization or text-to-histogram methods) are reported, and style similarity trails the image-driven oracle substantially. Given that H-Diffuser is trained on Qwen-generated captions that deliberately exclude object semantics, the claim of flexible natural-language control needs either external comparisons or a controlled human study isolating prompt fidelity; otherwise the multimodal contribution remains incompletely substantiated.
minor comments (4)
  1. [Appendix A] Appendix A soft-binning and square-root normalization of Hab are described clearly, but the precise bin-edge definitions and the value of epsilon are not stated; a short equation or pseudocode block would aid reproducibility.
  2. [Fig. 5] Fig. 5 caption and surrounding text contain residual control characters and formatting artifacts that should be cleaned for the camera-ready version.
  3. [Sec. 3.2 / Appendix B] The loss weights (lambda_lut=1.0, lambda_img=0.5, lambda_mono=5.0, lambda_tv=0.0001) and diffusion lambdas are given only in Appendix B; moving a one-sentence summary into the main text would help readers assess sensitivity without flipping to the appendix.
  4. [References] References include several 2025–2026 entries; ensure final DOIs or arXiv identifiers are stable before publication.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: self-supervised synthetic-LUT training is evaluated on external public benchmarks with independent metrics and a falsifying patch-shuffle test.

full rationale

The paper's load-bearing claims (superior content-style balance on NAS/PST50, near-zero residuals under style patch-shuffling, and multimodal text-driven capability) are empirical results obtained by training a residual Transformer Seq2Seq LUT predictor on self-supervised synthetic pairs (random LUTs applied to COCO content, then spatially augmented to form style references) and then measuring against held-out public benchmarks (PST50, PhotoNAS) plus a user study and an independent patch-shuffling residual test. No equation or definition forces the reported Content Sim./Style Sim./Dist.-to-Ideal numbers or the residual maps by construction; the Lab-Extractor histograms, MR-Mapper residual prediction (Eqs. 1-4), and H-Diffuser X0-prediction objective are design choices whose quality is assessed externally. There are no self-citations of uniqueness theorems, no fitted parameters renamed as predictions of the same quantities, and no ansatz smuggled via prior author work. The synthetic training manifold is a potential generalization risk (explicitly flagged in Limitations), but that is a correctness/coverage concern, not circularity of the derivation chain.

Assumptions & free parameters 5 free parameters · 4 assumptions · 3 invented entities

The central claim rests on the empirical premise that Lab histograms fully capture transferable style, on a set of hand-chosen architectural and loss hyperparameters, and on three newly named modules whose only evidence is the paper’s own experiments. No free parameters are fitted to the evaluation metrics themselves, but many design constants are chosen without external calibration.

free parameters (5)
  • LUT grid size D
    Chosen as D=16 after an ablation showing diminishing returns; cubic complexity makes the choice consequential for both quality and memory.
  • chrominance stretch gamma and scale S
    gamma=0.5, S=128 set by hand to expand central Lab bins; directly shapes the statistical features fed to every downstream module.
  • loss weights lambda_lut, lambda_img, lambda_mono, lambda_tv
    Empirically set to 1.0 / 0.5 / 5.0 / 0.0001; control the trade-off between LUT fidelity, image fidelity and manifold regularity.
  • density-aware mask scale alpha and diffusion lambdas
    alpha and lambda_L/ab/M control the H-Diffuser objective; chosen without external validation set.
  • Transformer depth, hidden size, heads; DiT blocks and hidden size
    6-layer 512-d 8-head decoder and 8-block 768-d DiT are architectural free choices that determine capacity.
assumptions (4)
  • domain assumption Spatially-agnostic Lab histograms (HL, Hab, ML|ab) completely decouple color/tone style from structural semantics, so that a LUT conditioned only on them cannot introduce spatial artifacts.
    Stated as the core motivation in Sec. 3.1 and Introduction; never proved, only supported by the patch-shuffling experiment.
  • domain assumption Predicting residual offsets on an identity 3-D LUT grid plus monotonicity/TV regularizers yields a topologically smooth color manifold free of banding.
    Sec. 3.2; standard in LUT literature but still an unproved modeling choice for the residual Transformer.
  • ad hoc to paper Self-supervised pairs formed by applying random LUTs then spatially augmenting the result are distributionally sufficient for real photographic style transfer.
    Training paradigm in Sec. 3.2; the Limitations section itself notes failure modes on extreme real gaps.
  • ad hoc to paper CLIP text embeddings plus an X0-prediction DiT can synthesize valid Lab statistical features whose induced LUTs match human notions of the prompt.
    Sec. 3.3; text-driven style similarity is markedly lower than image-driven, indicating the axiom is only partially true.
invented entities (3)
  • Lab-Extractor
    purpose: Produces the three spatially-agnostic statistical descriptors that replace deep image encoders.
    New named module; independent evidence is limited to the paper’s own ablations and residual maps.
  • MR-Mapper (Multi-dimensional Residual Mapper)
    purpose: Unifies heterogeneous 1-D/2-D histograms into residual and global tokens for the Transformer decoder.
    Architectural invention; only evidence is the ablation against MLP baselines inside this paper.
  • H-Diffuser
    purpose: Lightweight DiT that maps text prompts directly to Lab statistical features, enabling text-driven LUT generation.
    New module; no external validation or third-party reimplementation exists yet.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Multimodal 3D LUT Generation via StatLUT with Statistical Features for Photorealistic Style Transfer." pith.science (2026). https://pith.science/paper/F7PWZL5V

@misc{pith2026260708227,
  author       = {Pith},
  title        = {Pith review of: Multimodal 3D LUT Generation via StatLUT with Statistical Features for Photorealistic Style Transfer},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/F7PWZL5V}},
  note         = {Machine review of arXiv:2607.08227}
}
read the original abstract

Photorealistic Style Transfer (PST) aims to transfer the color and tonal style of a reference to a content image while strictly preserving its structural integrity. However, existing deep learning-based methods inherently suffer from semantic entanglement caused by pre-trained image encoders, leading to unnatural spatial distortions. Moreover, current pixel-level mapping paradigms often ignore color gamut topology, resulting in color banding, while also lacking the multimodal capability for intuitive text-driven control. To address these bottlenecks, we propose StatLUT, an innovative multimodal framework for 3D LUT generation. First, we bypass traditional encoders and introduce a Lab-Extractor to derive spatially-agnostic statistical features, fundamentally decoupling color distributions from structural semantics to ensure artifact-free rendering. Second, we formulate LUT generation as a Transformer-based Seq2Seq translation task, utilizing a Multi-dimensional Residual Mapper (MR-Mapper) to predict topologically smooth 3D LUTs. Finally, to break the single-modal barrier, we propose the H-Diffuser, a lightweight Diffusion Transformer that directly synthesizes statistical features from natural language prompts, enabling flexible text-driven color grading. Extensive experiments on standard benchmarks demonstrate that StatLUT significantly outperforms state-of-the-art methods in both visual quality and quantitative metrics, pioneering a highly robust and flexible paradigm for multimodal photorealistic style transfer.

Figures

Figures reproduced from arXiv: 2607.08227 by the authors.

Figure 1
Figure 1. Overview of the proposed StatLUT framework. (1) Feature Extraction: The Lab-Extractor [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Architecture of the Multi-dimensional Residual Mapper (MR-Mapper), which unifies [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Visual evaluation of different methods on the NAS dataset. [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: Visual evaluation of different methods on the PST50 dataset. [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: Quantitative evaluation on the (a) NAS and (b) PST50 datasets. The top-right corner [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: Visual comparison of intermediate features and stylized results between the image-driven [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 7
Figure 7. Figure 7: Pipeline of the proposed Lab-Extractor. The downsampled image is converted to the [PITH_FULL_IMAGE:figures/full_fig_p012_7.png]
Figure 8
Figure 8. Figure 8: Qualitative results of video color grading. For each example, the top row shows the original [PITH_FULL_IMAGE:figures/full_fig_p014_8.png]
Figure 9
Figure 9. Figure 9: Visual and quantitative results of the style patch-shuffling experiment. The top row displays [PITH_FULL_IMAGE:figures/full_fig_p015_9.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

37 extracted references · 37 canonical work pages

  1. [1]

    Ultrafast Photorealistic Style Transfer via Neural Architecture Search

    Jie An, Haoyi Xiong, Jun Huan, and Jiebo Luo. Ultrafast Photorealistic Style Transfer via Neural Architecture Search. InAAAI, 2020

    work page 2020

  2. [2]

    Gatys, Alexander S

    Leon A. Gatys, Alexander S. Ecker, and Matthias Bethge. Image Style Transfer Using Convolutional Neural Networks. InCVPR, 2016

    work page 2016

  3. [3]

    Arbitrary Style Transfer in Real-Time with Adaptive Instance Normaliza- tion

    Xun Huang and Serge Belongie. Arbitrary Style Transfer in Real-Time with Adaptive Instance Normaliza- tion. InICCV, 2017

    work page 2017

  4. [4]

    Perceptual Losses for Real-Time Style Transfer and Super-Resolution

    Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual Losses for Real-Time Style Transfer and Super-Resolution. InECCV, 2016

    work page 2016

  5. [5]

    Arbitrary Style Transfer With Style-Attentional Networks

    Dae Young Park and Kwang Hee Lee. Arbitrary Style Transfer With Style-Attentional Networks. InCVPR, 2019

    work page 2019

  6. [6]

    A closed-form solution to photorealistic image stylization

    Yijun Li, Ming-Yu Liu, Xueting Li, Ming-Hsuan Yang, and Jan Kautz. A closed-form solution to photorealistic image stylization. InECCV, 2018

    work page 2018

  7. [7]

    Photorealistic style transfer via wavelet transforms

    Jaejun Yoo, Youngjung Uh, Sanghyuk Chun, Byeongkyu Kang, and Jung-Woo Ha. Photorealistic style transfer via wavelet transforms. InICCV, 2019

    work page 2019

  8. [8]

    PhotoWCT2 : Compact Autoencoder for Photorealistic Style Transfer Resulting from Blockwise Training and Skip Connections of High-Frequency Residuals

    Tai-Yin Chiu and Danna Gurari. PhotoWCT2 : Compact Autoencoder for Photorealistic Style Transfer Resulting from Blockwise Training and Skip Connections of High-Frequency Residuals. InWACV, 2022

    work page 2022

Show all 37 references
  1. [9]

    AdaCM: Adaptive ColorMLP for Real-Time Universal Photo-Realistic Style Transfer

    Tianwei Lin, Honglin Lin, Fu Li, Dongliang He, Wenhao Wu, Meiling Wang, Xin Li, and Yong Liu. AdaCM: Adaptive ColorMLP for Real-Time Universal Photo-Realistic Style Transfer. InAAAI, 2023

    2023

  2. [10]

    D-LUT: Photorealistic style transfer via diffusion process

    Mujing Li, Guanjie Wang, Xingguang Zhang, Qifeng Liao, and Chenxi Xiao. D-LUT: Photorealistic style transfer via diffusion process. InWACV, 2025

    2025

  3. [11]

    Zhanghan Ke, Yuhao Liu, Lei Zhu, Nanxuan Zhao, and Rynson W.H. Lau. Neural Preset for Color Style Transfer. InCVPR, 2023

    2023

  4. [12]

    NLUT: Neural-based 3D Lookup Tables for Video Photorealistic Style Transfer.arXiv preprint arXiv:2303.09170, 2023

    Yaosen Chen, Han Yang, Yuexin Yang, Yuegen Liu, Wei Wang, Xuming Wen, and Chaoping Xie. NLUT: Neural-based 3D Lookup Tables for Video Photorealistic Style Transfer.arXiv preprint arXiv:2303.09170, 2023

    2023 arXiv

  5. [13]

    SA-LUT: Spatial adaptive 4D look-up table for photorealistic style transfer

    Zerui Gong, Zhonghua Wu, Qingyi Tao, Qinyue Li, and Chen Change Loy. SA-LUT: Spatial adaptive 4D look-up table for photorealistic style transfer. InICCV, 2025

    2025

  6. [14]

    Video color grading via look-up table generation

    Seunghyun Shin, Dongmin Shin, Jisu Shin, Hae-Gon Jeon, and Joon-Young Lee. Video color grading via look-up table generation. InICCV, 2025

    2025

  7. [15]

    Scalable diffusion models with transformers

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

    2023

  8. [16]

    Pitie, A.C

    F. Pitie, A.C. Kokaram, and R. Dahyot. N-dimensional probability density function transfer and its application to color transfer. InICCV, 2005

    2005

  9. [17]

    Very Deep Convolutional Networks for Large-Scale Image Recognition

    Karen Simonyan and Andrew Zisserman. Very Deep Convolutional Networks for Large-Scale Image Recognition. InICLR, 2015

    2015

  10. [18]

    Efficientnet: Rethinking model scaling for convolutional neural networks

    Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In ICML, 2019

    2019

  11. [19]

    Color Transfer between Images

    Erik Reinhard, Michael Ashikhmin, Bruce Gooch, and Peter Shirley. Color Transfer between Images. IEEE CG&A, 2001

    2001

  12. [20]

    Photorealistic image style transfer based on explicit affine transformation

    Junjie Kang, Jinsong Wu, and Shiqi Jiang. Photorealistic image style transfer based on explicit affine transformation. InICME, 2024

    2024

  13. [21]

    Joint Bilateral Learning for Real-Time Universal Photorealistic Style Transfer

    Xide Xia, Meng Zhang, Tianfan Xue, Zheng Sun, Hui Fang, Brian Kulis, and Jiawen Chen. Joint Bilateral Learning for Real-Time Universal Photorealistic Style Transfer. InECCV, 2020

    2020

  14. [22]

    Temporal Consistent Semantic Video Color Transfer from Multiple References

    Aupendu Kar and Guan-Ming Su. Temporal Consistent Semantic Video Color Transfer from Multiple References. InCVPR, 2025

    2025

  15. [23]

    Guiding Instruction- based Image Editing via Multimodal Large Language Models

    Tsu-Jui Fu, Wenze Hu, Xianzhi Du, William Yang Wang, Yinfei Yang, and Zhe Gan. Guiding Instruction- based Image Editing via Multimodal Large Language Models. InICLR, 2024. 10

    2024

  16. [24]

    Imagic: Text-Based Real Image Editing with Diffusion Models

    Bahjat Kawar, Shiran Zada, Oran Lang, Omer Tov, Huiwen Chang, Tali Dekel, Inbar Mosseri, and Michal Irani. Imagic: Text-Based Real Image Editing with Diffusion Models. InCVPR, 2023

    2023

  17. [25]

    Tim Brooks, Aleksander Holynski, and Alexei A. Efros. InstructPix2Pix: Learning to Follow Image Editing Instructions. InCVPR, 2023

    2023

  18. [26]

    Learning Transferable Visual Models From Natural Language Supervision

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning Transferable Visual Models From Natural Language Supervision. InICML, 2021

    2021

  19. [27]

    SDEdit: Guided Image Synthesis and Editing with Stochastic Differential Equations

    Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jiajun Wu, Jun-Yan Zhu, and Stefano Ermon. SDEdit: Guided Image Synthesis and Editing with Stochastic Differential Equations. InICLR, 2021

    2021

  20. [28]

    L-CAD: Language-based Colorization with Any-level Descriptions using Diffusion Priors

    Zheng Chang, Shuchen Weng, Peixuan Zhang, Yu Li, Si Li, and Boxin Shi. L-CAD: Language-based Colorization with Any-level Descriptions using Diffusion Priors. InNIPS, 2023

    2023

  21. [29]

    Instant Photorealistic Style Transfer: A Lightweight and Adaptive Approach.arXiv preprint arXiv:2309.10011, 2023

    Rong Liu. Instant Photorealistic Style Transfer: A Lightweight and Adaptive Approach.arXiv preprint arXiv:2309.10011, 2023

    2023 arXiv

  22. [30]

    Color Transfer for Images: A Survey.TOMM, 2024

    Chenlei Lv, Dan Zhang, Shengling Geng, Zhongke Wu, and Hui Huang. Color Transfer for Images: A Survey.TOMM, 2024

    2024

  23. [31]

    Bovik, Hamid R

    Zhou Wang, Alan C. Bovik, Hamid R. Sheikh, and Eero P. Simoncelli. Image quality assessment: From error visibility to structural similarity.TIP, 2004

    2004

  24. [32]

    Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dollár, and C

    Tsung-Yi Lin, Michael Maire, Serge J. Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dollár, and C. Lawrence Zitnick. Microsoft COCO: Common Objects in Context. InECCV, 2014

    2014

  25. [33]

    Denoising Diffusion Implicit Models

    Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising Diffusion Implicit Models. InICLR, 2021

    2021

  26. [34]

    Ldc: Lightweight dense cnn for edge detection.IEEE Access, 2022

    Xavier Soria, Gonzalo Pomboza-Junez, and Angel Domingo Sappa. Ldc: Lightweight dense cnn for edge detection.IEEE Access, 2022

    2022

  27. [35]

    Decoupled Weight Decay Regularization

    Ilya Loshchilov and Frank Hutter. Decoupled Weight Decay Regularization. InICLR, 2019

    2019

  28. [36]

    Qwen3.5: Towards native multimodal agents, February 2026

    Qwen Team. Qwen3.5: Towards native multimodal agents, February 2026. URL https://qwen.ai/ blog?id=qwen3.5. 11 A Detailed Formulation of the Lab-Extractor CIE Lab Conversion Stretching Normalization Soft-binning Downsampling RGB L ab Figure 7: Pipeline of the proposed Lab-Extra...

    2026

  29. [37]

    The learning rate is initialized at 3.0×10 −4 with a weight decay of 0.05, incorporating a 5-epoch linear warmup followed by a cosine annealing schedule decaying to 1.0×10 −7

    for 200 epochs. The learning rate is initialized at 3.0×10 −4 with a weight decay of 0.05, incorporating a 5-epoch linear warmup followed by a cosine annealing schedule decaying to 1.0×10 −7. The empirical weights for the objective functions are set to λlut = 1.0, λimg = 0.5, ...

Pith tools

Reviewed July 10, 2026 · model on record in the stance chip above.