Pith. sign in

REVIEW 4 major objections 4 minor 83 references

SCFlow: Implicitly Learning Style and Content Disentanglement with Flow Models

T0 review · 4 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read SCFlow claims that training a flow-matching model only to merge style and content yields an invertible mapping whose inverse separates them, transferring zero-shot to real image datasets.

desk verdict A genuinely new idea — train only the merge, get separation for free — but the whole case rests on zero-shot transfer that I can't verify from the corrupted text. read the letter →

arxiv 2508.03402 v1 pith:TLTMQ4D4 submitted 2025-08-05 cs.CV cs.AIcs.LG

Pingchuan Ma , Xiaopei Yang , Yusong Li , Ming Gui , Felix Krause , Johannes Schusterbauer , Björn Ommer This is my paper
classification cs.CVcs.AIcs.LG
keywords styleandcontentdisentanglementflowmatchinginvertiblemappingcontrollableimagegenerationzero-shottransfersyntheticpaireddataset
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

SCFlow claims that style and content disentanglement does not need to be trained as a separation task. Instead, the paper trains a flow-matching model to invertibly merge a style code and a content code into an image, and the inverse mapping then separates them as a byproduct. Because merging is a well-defined task, the model learns to disentangle without explicit supervision. The method uses flow matching to bridge arbitrary distributions rather than relying on restrictive Gaussian priors, and it is trained on a synthetic dataset of 510,000 paired samples. If the claim holds, a single invertible model can perform both controllable generation and zero-shot style and content transfer on real images.

What carries the argument

The load-bearing object is a flow-matching bidirectional map between an entangled image and a disentangled pair $(s, c)$ of style and content codes. Flow matching here means learning a probability path between arbitrary source and target distributions, so the model can bridge an image distribution to a product of style and content distributions without forcing Gaussian latents. The invertibility of the learned map is what makes separation exact: once merging is trained, splitting is simply the inverse direction of the same map.

What would settle it

Take a model trained only on synthetic pairs and run it on real photographs whose true style and content are known, for example by rendering the same 3D scene under two different painterly filters, then check whether the extracted style code matches the filter and the extracted content code matches the unfiltered geometry; failure would show that merge training did not yield transferable disentanglement.

Watch

Extended reading notes

Core claim

The central claim is that invertible disentanglement emerges naturally from learning to merge style and content, rather than from an explicit separation objective. SCFlow learns a bidirectional mapping between an entangled image representation and a disentangled pair consisting of a style code and a content code. Training only to fuse the two factors into an image makes the inverse mapping a separation operation. The paper further claims that flow matching lets the model map between arbitrary distributions, avoiding the Gaussian prior restrictions of diffusion models and normalizing flows, and that the resulting mapping transfers zero-shot to ImageNet-1k and WikiArt with competitive performance.

Load-bearing premise

The 510,000 synthetic images, made by pairing 51 styles with 10,000 content samples, capture enough of the real style-content relationship that a model trained only on them separates real photos and artworks correctly.

Editorial extensions

If this is right

  • Style and content editing become a matter of swapping one code and applying the inverse map, giving a single model for both generation and manipulation.
  • Zero-shot transfer to ImageNet-1k and WikiArt would show that synthetic style-content pairing can induce factor representations usable on real photographs and artworks.
  • Avoiding Gaussian priors allows the disentangled factors to follow arbitrary distributions, which may preserve richer structure than standard normalizing-flow latents.
  • Because the merge is invertible, content details are not discarded during generation, so the content code can be recovered exactly from an image.
  • The approach reframes disentanglement as a side effect of a well-defined fusion task, potentially removing the need for adversarial or hand-crafted separation losses.

Reading between the lines

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

  • The paper leaves implicit that its notion of style is defined operationally by the 51 synthetic style groups; real-world style categories may not align with that operational definition, and the method's transfer would then reflect whichever factors those groups share.
  • The same invertible merge-then-split recipe could in principle be applied to other entangled factor pairs, such as lighting and geometry or pose and identity, by constructing analogous synthetic pairings.
  • The zero-shot result suggests a testable extension: transfer quality may depend more on the diversity and systematic pairing of synthetic styles than on their photorealism, a hypothesis the paper does not directly test.
  • If synthetic pairing is the only supervision, failure on real images would expose that the synthetic distribution misses some real style-content correlation; this makes dataset construction the main lever for broader applicability.
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

4 major / 4 minor

Summary. The paper proposes SCFlow, a flow-matching framework for learning bidirectional mappings between entangled images and disentangled style/content representations. The central claim is that training solely to merge style and content, without explicit disentanglement supervision, yields an invertible representation from which style and content can be separated naturally. To support this, the authors curate a synthetic dataset of 510,000 images (51 styles × 10,000 content samples) and report zero-shot generalization to ImageNet-1k and WikiArt. However, the provided full text is severely corrupted and unreadable beyond the abstract, so the derivations, training details, experimental protocol, and quantitative results could not be inspected.

Significance. The proposed idea is conceptually attractive: if a model learns to fuse style and content in an invertible way, explicit disentanglement losses may be unnecessary. The use of flow matching on arbitrary distributions is a reasonable alternative to Gaussian-prior diffusion models, and the synthetic dataset is large and systematic. The zero-shot evaluation on real benchmarks is a valuable check of external validity. That said, the abstract alone provides no quantitative evidence of disentanglement, and the unreadable full text prevents verification of the central claim. The significance therefore cannot be confirmed from this submission.

major comments (4)
  1. [Abstract] The central claim of emergent disentanglement is not supported by the readable evidence. In the abstract, the only reported outcome is 'competitive performance' on zero-shot ImageNet-1k and WikiArt, with no quantitative disentanglement metric such as attribute-swap accuracy, latent-dimension probing, or human evaluation. Because the full text is corrupted, I could not verify whether such metrics appear later; as presented, the claim is unfalsifiable.
  2. [Abstract / dataset description] The synthetic dataset's product structure (51 styles × 10,000 content samples) may allow the model to trivially recover the (style, content) index pair rather than learn semantically meaningful factors. The abstract does not report controls or ablations such as held-out style-content pairs, non-product data, or varying the number of styles to rule out this shortcut. This concern is load-bearing because it threatens the interpretation of the synthetic results as emergent disentanglement.
  3. [Abstract] The 'zero-shot setting' is not defined. The abstract does not specify what is held out during training, what the model encounters at test time, or how the synthetic training distribution relates to ImageNet-1k and WikiArt. Without this protocol and quantitative comparisons to baselines, the zero-shot claim cannot be assessed.
  4. [Full text] The provided full text is corrupted (mojibake), making it impossible to read equations, algorithm pseudocode, experimental tables, or limitations. A clean version of the manuscript is required before the scientific content can be reviewed; I have therefore based my assessment entirely on the abstract.
minor comments (4)
  1. [Abstract] The term 'arbitrary distributions' should be made precise; flow matching does not work on literally arbitrary distributions, so the intended class (e.g., compactly supported continuous distributions) should be stated.
  2. [Abstract] The abstract should spell out the acronym SCFlow or define it at first use.
  3. [Abstract] The phrase 'competitive performance' should be replaced with concrete numbers, error bars, or a pointer to specific tables.
  4. [Abstract / dataset description] The dataset curation should specify what 'styles' are (e.g., artistic filters, texture families) and how the 10,000 content samples were selected.

Circularity Check

1 steps flagged · score 4.0 of 10

Synthetic-data 'emergence' is partly built into the dataset construction; zero-shot real-image results provide the only independent check.

  1. self definitional [Abstract, key insights 1 and 3 (synthetic dataset description)]
    "Our approach is built upon three key insights: 1) Training solely to merge style and content, a well-defined task, enables invertible disentanglement without explicit supervision; 2) flow matching bridges on arbitrary distributions, avoiding the restrictive Gaussian priors of diffusion models and normalizing flows; and 3) a synthetic dataset of 510,000 samples (51 styles × 10,000 content samples) was curated to simulate disentanglement through systematic style-content pairing."

    Each synthetic image is generated from a known (style, content) pair, so the ground-truth factorization is fixed by the dataset-construction rule. The flow is trained as a bidirectional mapping between entangled and disentangled representations, meaning the inverse direction is trained to recover exactly those generative pairs. On the synthetic distribution, 'disentanglement' therefore reduces by construction to inverting the data-generation function: the latent style and content variables coincide with the variables used to build the training pairs. This makes the synthetic demonstration of emergent separation self-definitional rather than an unsupervised discovery.

full rationale

The paper's derivation chain contains one load-bearing self-definitional step: the synthetic dataset encodes the style/content split by design as a product of 51 styles and 10,000 contents. An invertible flow trained to merge and unmerge such pairs can solve the synthetic disentanglement task by recovering the construction rule, so the synthetic evidence for 'disentanglement naturally emerges' is not independent of the dataset's built-in labels. However, the abstract also reports zero-shot generalization to ImageNet-1k and WikiArt, which is outside the synthetic manifold and constitutes genuine external evidence. I found no self-citation chains, imported uniqueness theorems, or ansatz-smuggling in the available material, and the corruption of the full text prevents identification of further internal equations. The appropriate score is moderate partial circularity, not a fully circular derivation.

Assumptions & free parameters 0 free parameters · 3 assumptions · 0 invented entities

No free parameters are visible from the abstract. The listed axioms are load-bearing assumptions about flow matching expressiveness, the validity of the synthetic dataset, and the sufficiency of invertibility. These are drawn directly from the abstract's stated insights and are not independently verified.

assumptions (3)
  • domain assumption Flow matching can effectively learn bidirectional invertible mappings between arbitrary distributions.
    The method's second key insight assumes that flow matching bridges any two distributions, which is a strong modeling assumption not proven in the abstract.
  • domain assumption The synthetic dataset of 51 styles and 10,000 contents is a sufficient simulation of real style-content entanglement.
    The third insight relies on this dataset to make disentanglement learnable, but the abstract provides no validation that synthetic pairings transfer to real images beyond the zero-shot claim.
  • ad hoc to paper Invertibility of the learned mapping preserves all information needed to separate style and content.
    The core mechanism assumes invertibility alone guarantees that the merged representation contains no lossy mix that prevents clean separation.

how reviews work

0 comments
Cite this review

Pith. "Pith review of SCFlow: Implicitly Learning Style and Content Disentanglement with Flow Models." pith.science (2026). https://pith.science/paper/TLTMQ4D4

@misc{pith2026250803402,
  author       = {Pith},
  title        = {Pith review of: SCFlow: Implicitly Learning Style and Content Disentanglement with Flow Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/TLTMQ4D4}},
  note         = {Machine review of arXiv:2508.03402}
}
abstract

Explicitly disentangling style and content in vision models remains challenging due to their semantic overlap and the subjectivity of human perception. Existing methods propose separation through generative or discriminative objectives, but they still face the inherent ambiguity of disentangling intertwined concepts. Instead, we ask: Can we bypass explicit disentanglement by learning to merge style and content invertibly, allowing separation to emerge naturally? We propose SCFlow, a flow-matching framework that learns bidirectional mappings between entangled and disentangled representations. Our approach is built upon three key insights: 1) Training solely to merge style and content, a well-defined task, enables invertible disentanglement without explicit supervision; 2) flow matching bridges on arbitrary distributions, avoiding the restrictive Gaussian priors of diffusion models and normalizing flows; and 3) a synthetic dataset of 510,000 samples (51 styles $\times$ 10,000 content samples) was curated to simulate disentanglement through systematic style-content pairing. Beyond controllable generation tasks, we demonstrate that SCFlow generalizes to ImageNet-1k and WikiArt in zero-shot settings and achieves competitive performance, highlighting that disentanglement naturally emerges from the invertible merging process.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

83 extracted references · 56 canonical work pages

  1. [1]

    Unsplash https://unsplash.com/data

  2. [2]

    Building normalizing flows with stochastic interpolants

    Michael S Albergo and Eric Vanden-Eijnden. Building normalizing flows with stochastic interpolants. In ICLR, 2023

    2023

  3. [3]

    Stochastic interpolants: A unifying framework for flows and diffusions

    Michael S Albergo, Nicholas M Boffi, and Eric Vanden-Eijnden. Stochastic interpolants: A unifying framework for flows and diffusions. arXiv, 2023

    2023

  4. [4]

    Coyo-700m: Image-text pair dataset

    Minwoo Byeon, Beomhee Park, Haecheon Kim, Sungjun Lee, Woonhyuk Baek, and Saehoon Kim. Coyo-700m: Image-text pair dataset. https://github.com/kakaobrain/coyo-dataset, 2022

    2022

  5. [5]

    Emerging properties in self-supervised vision transformers

    Mathilde Caron, Hugo Touvron, Ishan Misra, Herv\'e J\'egou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, 2021

    work page 2021

  6. [6]

    Neural ordinary differential equations

    Ricky TQ Chen, Yulia Rubanova, Jesse Bettencourt, and David K Duvenaud. Neural ordinary differential equations. Advances in neural information processing systems, 31, 2018

    2018

  7. [7]

    Learning a similarity metric discriminatively, with application to face verification

    Sumit Chopra, Raia Hadsell, and Yann LeCun. Learning a similarity metric discriminatively, with application to face verification. In 2005 IEEE computer society conference on computer vision and pattern recognition (CVPR'05), pages 539--546. IEEE, 2005

    work page 2005

  8. [8]

    Flow matching in latent space

    Quan Dao, Hao Phung, Binh Nguyen, and Anh Tran. Flow matching in latent space. arXiv, 2023

    work page 2023

Show all 83 references
  1. [9]

    Imagenet: A large-scale hierarchical image database

    Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248--255. Ieee, 2009

    2009

  2. [10]

    Nice: Non-linear independent components estimation

    Laurent Dinh, David Krueger, and Yoshua Bengio. Nice: Non-linear independent components estimation. ICLR, 2015

    2015

  3. [11]

    Density estimation using real nvp

    Laurent Dinh, Jascha Sohl-Dickstein, and Samy Bengio. Density estimation using real nvp. ICLR, 2017

    2017

  4. [12]

    The use of multiple measurements in taxonomic problems

    Ronald A Fisher. The use of multiple measurements in taxonomic problems. Annals of eugenics, 7 0 (2): 0 179--188, 1936

    1936

  5. [13]

    Discriminatory analysis: nonparametric discrimination, consistency properties

    Evelyn Fix. Discriminatory analysis: nonparametric discrimination, consistency properties. USAF school of Aviation Medicine, 1985

    1985

  6. [14]

    Implicit style-content separation using b-lora

    Yarden Frenkel, Yael Vinker, Ariel Shamir, and Daniel Cohen-Or. Implicit style-content separation using b-lora. In ECCV, 2024

    2024

  7. [15]

    Diffusion models and representation learning: A survey

    Michael Fuest, Pingchuan Ma, Ming Gui, Johannes Schusterbauer, Vincent Tao Hu, and Bjorn Ommer. Diffusion models and representation learning: A survey. arXiv preprint arXiv:2407.00783, 2024

    2024 arXiv

  8. [16]

    An image is worth one word: Personalizing text-to-image generation using textual inversion

    Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv, 2022

    2022

  9. [17]

    Sliderspace: Decomposing the visual capabilities of diffusion models, 2025

    Rohit Gandikota, Zongze Wu, Richard Zhang, David Bau, Eli Shechtman, and Nick Kolkin. Sliderspace: Decomposing the visual capabilities of diffusion models, 2025

    2025

  10. [18]

    Gatys, Alexander S

    Leon A. Gatys, Alexander S. Ecker, and Matthias Bethge. Image style transfer using convolutional neural networks. In CVPR, 2016 a

    2016

  11. [19]

    Gatys, Alexander S

    Leon A. Gatys, Alexander S. Ecker, and Matthias Bethge. Image style transfer using convolutional neural networks. In CVPR, 2016 b

    2016

  12. [20]

    Depthfm: Fast monocular depth estimation with flow matching

    Ming Gui, Johannes Schusterbauer, Ulrich Prestel, Pingchuan Ma, Dmytro Kotovenko, Olga Grebenkova, Stefan Andreas Baumann, Vincent Tao Hu, and Björn Ommer. Depthfm: Fast monocular depth estimation with flow matching. AAAI, 2025

    2025

  13. [21]

    Flowtok: Flowing seamlessly across text and image tokens

    Ju He, Qihang Yu, Qihao Liu, and Liang-Chieh Chen. Flowtok: Flowing seamlessly across text and image tokens. arXiv preprint arXiv:2503.10772, 2025

    2025

  14. [22]

    Momentum contrast for unsupervised visual representation learning

    Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. arxiv e-prints, art. arXiv preprint arXiv:1911.05722, 6 0 (7), 2019

    1911 arXiv

  15. [23]

    Prompt-to-prompt image editing with cross attention control

    Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv, 2022

    2022

  16. [24]

    Style aligned image generation via shared attention

    Amir Hertz, Andrey Voynov, Shlomi Fruchter, and Daniel Cohen-Or. Style aligned image generation via shared attention. In CVPR, 2024

    2024

  17. [25]

    Clipscore: A reference-free evaluation metric for image captioning

    Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning. arXiv preprint arXiv:2104.08718, 2021

    2021 arXiv

  18. [26]

    Denoising diffusion probabilistic models

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020

    2020

  19. [27]

    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, 2021

    2021

  20. [28]

    Tao Hu, David W Zhang, Pascal Mettes, Meng Tang, Deli Zhao, and Cees G.M. Snoek. Latent space editing in transformer-based flow matching. In AAAI, 2024

    2024

  21. [29]

    Gpt-4o system card

    Aaron Hurst, Adam Lerer, Adam P Goucher, et al. Gpt-4o system card. arXiv, 2024

    2024

  22. [30]

    Product quantization for nearest neighbor search

    Herve Jegou, Matthijs Douze, and Cordelia Schmid. Product quantization for nearest neighbor search. IEEE transactions on pattern analysis and machine intelligence, 33 0 (1): 0 117--128, 2010

    2010

  23. [31]

    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. In Proceedings of the European Conference on Computer Vision (ECCV). Springer, 2016

    2016

  24. [32]

    Recognizing image style

    Sergey Karayev, Matthew Trentacoste, Helen Han, Aseem Agarwala, Trevor Darrell, Aaron Hertzmann, and Holger Winnemoeller. Recognizing image style. arXiv preprint arXiv:1311.3715, 2013

    2013 arXiv

  25. [33]

    Elucidating the design space of diffusion-based generative models

    Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. In NeurIPS, 2022

    2022

  26. [34]

    Glow: Generative flow with invertible 1x1 convolutions

    Durk P Kingma and Prafulla Dhariwal. Glow: Generative flow with invertible 1x1 convolutions. Advances in neural information processing systems, 31, 2018

    2018

  27. [35]

    Rethinking style transfer: From pixels to parameterized brushstrokes

    Dmytro Kotovenko, Matthias Wright, Arthur Heimbrecht, and Bjorn Ommer. Rethinking style transfer: From pixels to parameterized brushstrokes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021

    2021

  28. [36]

    Dongxu Li, Junnan Li, and Steven C. H. Hoi. Blip-diffusion: Pre-trained subject representation for controllable text-to-image generation and editing, 2023

    2023

  29. [37]

    Learning linear transformations for fast image and video style transfer

    Xueting Li, Sifei Liu, Jan Kautz, and Ming-Hsuan Yang. Learning linear transformations for fast image and video style transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019

    2019

  30. [38]

    Flow matching for generative modeling

    Yaron Lipman, Ricky TQ Chen, Heli Ben-Hamu, Maximilian Nickel, and Matt Le. Flow matching for generative modeling. ICLR, 2023 a

    2023

  31. [39]

    Yaron Lipman, Ricky T. Q. Chen, Heli Ben-Hamu, Maximilian Nickel, and Matthew Le. Flow matching for generative modeling. In ICLR, 2023 b

    2023

  32. [40]

    Improved baselines with visual instruction tuning

    Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 26296--26306, 2024

    2024

  33. [41]

    Flowing from words to pixels: A noise-free framework for cross-modality evolution

    Qihao Liu, Xi Yin, Alan Yuille, Andrew Brown, and Mannat Singh. Flowing from words to pixels: A noise-free framework for cross-modality evolution. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 2755--2765, 2025

    2025

  34. [42]

    Flow straight and fast: Learning to generate and transfer data with rectified flow

    Xingchao Liu, Chengyue Gong, and Qiang Liu. Flow straight and fast: Learning to generate and transfer data with rectified flow. ICLR, 2023

    2023

  35. [43]

    Sit: Exploring flow and diffusion-based generative models with scalable interpolant transformers

    Nanye Ma, Mark Goldstein, Michael S Albergo, Nicholas M Boffi, Eric Vanden-Eijnden, and Saining Xie. Sit: Exploring flow and diffusion-based generative models with scalable interpolant transformers. ECCV, 2024

    2024

  36. [44]

    Some methods for classification and analysis of multivariate observations

    James MacQueen. Some methods for classification and analysis of multivariate observations. In Proceedings of the Fifth Berkeley Symposium on Mathematical Statistics and Probability, Volume 1: Statistics, pages 281--298. University of California press, 1967

    1967

  37. [45]

    An introduction to information retrieval

    Christopher D Manning. An introduction to information retrieval. 2009

    2009

  38. [46]

    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. arXiv preprint arXiv:2108.01073, 2021

    2021 arXiv

  39. [47]

    Action matching: Learning stochastic dynamics from samples

    Kirill Neklyudov, Rob Brekelmans, Daniel Severo, and Alireza Makhzani. Action matching: Learning stochastic dynamics from samples. In ICML, 2023

    2023

  40. [48]

    Representation learning with contrastive predictive coding

    Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018

    2018 arXiv

  41. [49]

    Deadiff: An efficient stylization diffusion model with disentangled representations

    Tianhao Qi, Shancheng Fang, Yanze Wu, Hongtao Xie, Jiawei Liu, Lang Chen, Qian He, and Yongdong Zhang. Deadiff: An efficient stylization diffusion model with disentangled representations. In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), page 8693–...

    2024

  42. [50]

    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, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning (IC...

    2021

  43. [51]

    Zero-shot text-to-image generation

    Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In ICML, 2021

    2021

  44. [52]

    Hierarchical text-conditional image generation with clip latents, 2022

    Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents, 2022

    2022

  45. [53]

    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. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022

    2022

  46. [54]

    Aladin: All layer adaptive instance normalization for fine-grained style similarity

    Dan Ruta, Saeid Motiian, Baldo Faieta, Zhe Lin, Hailin Jin, Alex Filipkowski, Andrew Gilbert, and John Collomosse. Aladin: All layer adaptive instance normalization for fine-grained style similarity. In ICCV, 2021

    2021

  47. [55]

    Large-scale classification of fine-art paintings: Learning the right metric on the right feature

    Babak Saleh and Ahmed Elgammal. Large-scale classification of fine-art paintings: Learning the right metric on the right feature. arXiv preprint arXiv:1505.00855, 2015

    2015 arXiv

  48. [56]

    Improving deep metric learning by divide and conquer

    Artsiom Sanakoyeu, Pingchuan Ma, Vadim Tschernezki, and Bj \"o rn Ommer. Improving deep metric learning by divide and conquer. IEEE Transactions on pattern analysis and machine intelligence, 44 0 (11): 0 8306--8320, 2021

    2021

  49. [57]

    Laion-400m: Open dataset of clip-filtered 400 million image-text pairs

    Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021

    2021 arXiv

  50. [58]

    Laion-5b: An open large-scale dataset for training next generation image-text models

    Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. NeurIPS, 2022

    2022

  51. [59]

    Boosting latent diffusion with flow matching

    Johannes Schusterbauer, Ming Gui, Pingchuan Ma, Nick Stracke, Stefan A Baumann, and Björn Ommer. Boosting latent diffusion with flow matching. ECCV, 2024

    2024

  52. [60]

    Ziplora: Any subject in any style by effectively merging loras

    Viraj Shah, Nataniel Ruiz, Forrester Cole, Erika Lu, Svetlana Lazebnik, Yuanzhen Li, and Varun Jampani. Ziplora: Any subject in any style by effectively merging loras. In ECCV, 2024

    2024

  53. [61]

    Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning

    Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), ...

    2018

  54. [62]

    Very deep convolutional networks for large-scale image recognition, 2015

    Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition, 2015

    2015

  55. [63]

    Deep unsupervised learning using nonequilibrium thermodynamics

    Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In ICML, 2015

    2015

  56. [64]

    Measuring style similarity in diffusion models

    Gowthami Somepalli, Anubhav Gupta, Kamal Gupta, Shramay Palta, Micah Goldblum, Jonas Geiping, Abhinav Shrivastava, and Tom Goldstein. Measuring style similarity in diffusion models. ECCV, 2024 a

    2024

  57. [65]

    Measuring style similarity in diffusion models, 2024 b

    Gowthami Somepalli, Anubhav Gupta, Kamal Gupta, Shramay Palta, Micah Goldblum, Jonas Geiping, Abhinav Shrivastava, and Tom Goldstein. Measuring style similarity in diffusion models, 2024 b

    2024

  58. [66]

    Denoising diffusion implicit models

    Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In ICLR, 2021 a

    2021

  59. [67]

    Score-based generative modeling through stochastic differential equations

    Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In ICLR, 2021 b

    2021

  60. [68]

    Cleandift: Diffusion features without noise

    Nick Stracke, Stefan Andreas Baumann, Kolja Bauer, Frank Fundel, and Bj \"o rn Ommer. Cleandift: Diffusion features without noise. ECCV, 2024 a

    2024

  61. [69]

    Ctrloralter: Conditional loradapter for efficient 0-shot control and altering of t2i models

    Nick Stracke, Stefan Andreas Baumann, Joshua Susskind, Miguel Angel Bautista, and Björn Ommer. Ctrloralter: Conditional loradapter for efficient 0-shot control and altering of t2i models. In European Conference on Computer Vision, pages 87--103. Springer, 2024 b

    2024

  62. [70]

    Emergent correspondence from image diffusion

    Luming Tang, Menglin Jia, Qianqian Wang, Cheng Perng Phoo, and Bharath Hariharan. Emergent correspondence from image diffusion. Advances in Neural Information Processing Systems, 36: 0 1363--1389, 2023

    2023

  63. [71]

    Improving and generalizing flow-based generative models with minibatch optimal transport

    Alexander Tong et al. Improving and generalizing flow-based generative models with minibatch optimal transport. In ICML Worshop, 2023

    2023

  64. [72]

    Visualizing data using t-sne

    Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine learning research, 9 0 (11), 2008

    2008

  65. [73]

    p+: Extended textual conditioning in text-to-image generation

    Andrey Voynov, Qinghao Chu, Daniel Cohen-Or, and Kfir Aberman. p+: Extended textual conditioning in text-to-image generation. arXiv preprint arXiv:2303.09522, 2023

    2023 arXiv

  66. [74]

    Instantstyle: Free lunch towards style-preserving in text-to-image generation

    Haofan Wang, Matteo Spinelli, Qixun Wang, Xu Bai, Zekui Qin, and Anthony Chen. Instantstyle: Free lunch towards style-preserving in text-to-image generation. arXiv preprint arXiv:2404.02733, 2024

    2024 arXiv

  67. [75]

    Evaluating data attribution for text-to-image models

    Sheng-Yu Wang, Alexei A Efros, Jun-Yan Zhu, and Richard Zhang. Evaluating data attribution for text-to-image models. In ICCV, 2023 a

    2023

  68. [76]

    Glstylenet: exquisite style transfer combining global and local pyramid features

    Zhizhong Wang, Lei Zhao, Sihuan Lin, Qihang Mo, Huiming Zhang, Wei Xing, and Dongming Lu. Glstylenet: exquisite style transfer combining global and local pyramid features. IET Computer Vision, 14 0 (8), 2020

    2020

  69. [77]

    Microast: Towards super-fast ultra-resolution arbitrary style transfer

    Zhizhong Wang, Lei Zhao, Zhiwen Zuo, Ailin Li, Haibo Chen, Wei Xing, and Dongming Lu. Microast: Towards super-fast ultra-resolution arbitrary style transfer. In Proceedings of the AAAI Conference on Artificial Intelligence, 2023 b

    2023

  70. [78]

    Bam! the behance artistic media dataset for recognition beyond photography

    Michael J Wilber, Chen Fang, Hailin Jin, Aaron Hertzmann, John Collomosse, and Serge Belongie. Bam! the behance artistic media dataset for recognition beyond photography. In ICCV, 2017

    2017

  71. [79]

    Csgo: Content-style composition in text-to-image generation, 2024

    Peng Xing, Haofan Wang, Yanpeng Sun, Qixun Wang, Xu Bai, Hao Ai, Renyuan Huang, and Zechao Li. Csgo: Content-style composition in text-to-image generation, 2024

    2024

  72. [80]

    A tale of two features: Stable diffusion complements dino for zero-shot semantic correspondence

    Junyi Zhang, Charles Herrmann, Junhwa Hur, Luisa Polania Cabrera, Varun Jampani, Deqing Sun, and Ming-Hsuan Yang. A tale of two features: Stable diffusion complements dino for zero-shot semantic correspondence. Advances in Neural Information Processing Systems, 36: 0 45533--45...

    2023

  73. [81]

    Adding conditional control to text-to-image diffusion models, 2023 b

    Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models, 2023 b

    2023

  74. [82]

    Domain enhanced arbitrary image style transfer via contrastive learning

    Yuxin Zhang, Fan Tang, Weiming Dong, Haibin Huang, Chongyang Ma, Tong-Yee Lee, and Changsheng Xu. Domain enhanced arbitrary image style transfer via contrastive learning. In ACM SIGGRAPH 2022 Conference Proceedings, 2022

    2022

  75. [83]

    Style fader generative adversarial networks for style degree controllable artistic style transfer

    Zhiwen Zuo, Lei Zhao, Shuobin Lian, Haibo Chen, Zhizhong Wang, Ailin Li, Wei Xing, and Dongming Lu. Style fader generative adversarial networks for style degree controllable artistic style transfer. In Proc. Int. Joint Conf. on Artif. Intell.(IJCAI), 2022

    2022

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.