pith. sign in

arxiv: 2606.27880 · v1 · pith:7RZ6DRKAnew · submitted 2026-06-26 · 💻 cs.CV

OrthoTryOn: Geometric Orthogonalization for Conflict-Free Unified Fashion Generation

Zhaotong Yang , Ying Tai , Jiahui Zhan , Yu Zheng , Jianjun Qian , Jian Yang This is my paper

Pith reviewed 2026-06-29 04:26 UTC · model grok-4.3

classification 💻 cs.CV
keywords unified fashion generationvirtual try-ongarment reconstructionLoRA adaptationorthogonal subspace projectiongradient conflictdiffusion modelsmulti-task learning
0
0 comments X

The pith

Task-specific orthogonal rotations in a shared LoRA map fashion tasks into decorrelated frames to remove gradient conflicts.

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

The paper shows that a single diffusion model can handle multiple fashion tasks such as virtual try-on and garment reconstruction by adding task-specific orthogonal rotations inside a shared low-rank adapter. These rotations send each task's bottleneck features into its own coordinate frame so that gradients no longer fight each other during joint training. A second, parameter-free step called Fisher-guided negative guidance then steers the generation away from any remaining task overlap at inference time. The resulting unified model not only avoids the usual drop in quality from naive sharing but exceeds the accuracy of separately trained task-specific models on standard benchmarks and works with different diffusion backbones.

Core claim

OrthoTryOn mitigates inter-task interference by applying Orthogonal Subspace Projection (OSP) that rotates bottleneck features with task-specific orthogonal matrices inside the shared LoRA module, placing each task in a decorrelated coordinate frame. Fisher-guided Negative Guidance (FNG) then uses the diagonal of the Fisher information matrix to measure sensitivity overlap and repels the sampling trajectory from the most confusable task via classifier-free guidance. Experiments confirm that this approach eliminates the severe degradation seen in naive unified training and surpasses independently trained models across multiple benchmarks while remaining compatible with diverse diffusion archi

What carries the argument

Orthogonal Subspace Projection (OSP) that applies task-specific orthogonal rotations to bottleneck features inside the shared LoRA module.

If this is right

  • Unified training on multiple fashion tasks becomes feasible without the performance penalty of naive parameter sharing.
  • The single model exceeds the accuracy of independently trained task-specific models on virtual try-on and garment reconstruction benchmarks.
  • The same orthogonalization pattern works across multiple diffusion backbones without retraining the core architecture.
  • Fisher-guided negative guidance removes residual task overlap at inference time with no added trainable parameters.

Where Pith is reading between the lines

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

  • The same rotation-based separation could be tested on other multi-task generative problems such as simultaneous editing and captioning.
  • Shared adapters with geometric decorrelation might cut the total compute needed when expanding a fashion model to additional output styles.
  • Applying the method outside clothing domains would test whether the decorrelation benefit is specific to visual fashion semantics or holds more generally.

Load-bearing premise

That task-specific orthogonal rotations on bottleneck features will map the tasks into sufficiently decorrelated spaces without discarding important task information.

What would settle it

A controlled run in which OSP is added to the shared LoRA yet measured gradient conflicts between tasks stay high or final image quality remains below that of separate models.

Figures

Figures reproduced from arXiv: 2606.27880 by Jiahui Zhan, Jianjun Qian, Jian Yang, Ying Tai, Yu Zheng, Zhaotong Yang.

Figure 1
Figure 1. Figure 1: We present OrthoTryOn, a unified generalist model capable of handling diverse fashion tasks within a single architecture, including virtual try-on, garment reconstruc￾tion, and pose transfer. The shared architecture also naturally supports sequential edit￾ing by chaining task-specific conditions. Abstract. Unified fashion generation integrates tasks like virtual try-on and garment reconstruction into a sin… view at source ↗
Figure 2
Figure 2. Figure 2: (a) Task-Specific LoRAs maintain independent adapters for each task. (b) Our OrthoTryOn within a shared LoRA, enforcing decorrelated coordinate frames to mit￾igate inter-task conflict. (c) Gradient magnitudes of different methods across training steps, showing that OrthoTryOn effectively mitigates the gradient decay caused by inter-task conflict. Despite the statistical decorrelation introduced by OSP, res… view at source ↗
Figure 3
Figure 3. Figure 3: Overview of OrthoTryOn. Orthogonal Subspace Projection utilizes task-specific orthogonal matrices Qi in the shared LoRA module to rotate bottleneck features into decorrelated coordinate frames, achieving statistically decorrelated optimization in ex￾pectation. Based on parameter sensitivity tracked during training, Fisher-guided Neg￾ative Guidance explicitly suppresses the interfering task via CFG to preve… view at source ↗
Figure 4
Figure 4. Figure 4: Qualitative comparison of virtual try-on on VITON-HD dataset [12]. with the explicit condition of the interfering task: \hat {v} = v(x_t, t, c_{j^*}) + s \cdot \Big ( v(x_t, t, c_i) - v(x_t, t, c_{j^*}) \Big ), (11) where ci and cj ∗ represent the condition inputs for the target and the hard neg￾ative tasks, respectively, and s is the guidance scale. In the learned conditional velocity field, this operatio… view at source ↗
Figure 5
Figure 5. Figure 5: Qualitative comparison of garment reconstruction on VITON-HD dataset [12] and pose transfer on DeepFashion dataset [39]. Please zoom in for better view. pairs depicting the same identity under different poses. Text conditions are gen￾erated via Qwen2.5-VL-7B-Instruct [3], and human skeletons are extracted using HRNet [50]. 4.2 Comparison with State-of-the-Art Methods To evaluate OrthoTryOn, we benchmark ag… view at source ↗
Figure 7
Figure 7. Figure 7: Performance com￾parison of different virtual try-on variants across multi￾ple metrics. duce Fisher-guided Negative Guidance (FNG). We first evaluate a suboptimal variant FNG∗ (variant (e)), which forces the task with the lowest Fisher similar￾ity (i.e., maximum semantic discrepancy) as the negative prompt. As shown in Tab. 2, while this strategy enhances structural alignment and eliminates the red￾box arti… view at source ↗
read the original abstract

Unified fashion generation integrates tasks like virtual try-on and garment reconstruction into a single model to reduce task-specific adaptation costs. However, naive parameter sharing across semantically distinct tasks induces negative transfer through severe inter-task gradient conflict. We propose OrthoTryOn, a unified framework mitigating this interference within a shared Low-Rank Adaptation (LoRA) module. Its Orthogonal Subspace Projection (OSP) applies task-specific orthogonal rotations to bottleneck features, mapping them into decorrelated coordinate frames. To address residual semantic coupling at inference time, we further propose Fisher-guided Negative Guidance (FNG), a parameter-free strategy that utilizes diagonal Fisher information to quantify inter-task sensitivity overlap and explicitly repels generation trajectories from the most confusable task via Classifier-Free Guidance. Extensive experiments demonstrate that OrthoTryOn avoids the severe performance degradation typical of naive unified training and even surpasses independently trained task-specific models, achieving state-of-the-art results across multiple benchmarks while generalizing robustly across diverse diffusion backbones. Code is available at https://github.com/NJU-PCALab/OrthoTryOn.

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.

Referee Report

2 major / 1 minor

Summary. The paper introduces OrthoTryOn, a unified framework for fashion generation tasks such as virtual try-on and garment reconstruction. It uses a shared LoRA module with Orthogonal Subspace Projection (OSP) that applies task-specific orthogonal rotations to bottleneck features to map them into decorrelated frames, reducing gradient conflicts. Additionally, Fisher-guided Negative Guidance (FNG) is proposed to handle residual semantic coupling at inference using diagonal Fisher information. The authors claim that this approach avoids the performance degradation of naive unified training, surpasses task-specific models, and achieves state-of-the-art results on multiple benchmarks, generalizing across diffusion backbones.

Significance. If the results hold, this work would be significant for multi-task learning in generative models, particularly in reducing the cost of maintaining separate models for related fashion tasks. The geometric approach to orthogonalization and the parameter-free FNG are novel contributions. The release of code supports reproducibility.

major comments (2)
  1. [§3.2] §3.2: The OSP method applies fixed orthogonal matrices R_t to the bottleneck activation h so that the projected features R_t h lie in mutually orthogonal frames. However, if the original features contain linearly inseparable semantic components shared between tasks (e.g., garment shape cues between try-on and reconstruction), the rotation may project out part of the signal or leave residual coupling; no analytic bound is provided showing that the retained subspace still spans the task-specific manifold.
  2. [Experiments] Experiments (quantitative tables and ablations): The central claim that the unified model matches or exceeds independently trained task-specific baselines requires detailed ablation studies, quantitative tables with error bars, and cross-backbone comparisons to evaluate whether OSP and FNG deliver the reported gains without substantial loss of task-relevant information.
minor comments (1)
  1. Clarify whether the diagonal Fisher information in FNG is computed once per task or updated during training, and specify the exact form of the negative guidance scale.

Simulated Author's Rebuttal

2 responses · 1 unresolved

We thank the referee for the constructive feedback and for recognizing the potential significance of OrthoTryOn for multi-task generative modeling. We address each major comment below with clarifications and commitments to revision where appropriate.

read point-by-point responses

  1. Referee: [§3.2] The OSP method applies fixed orthogonal matrices R_t to the bottleneck activation h so that the projected features R_t h lie in mutually orthogonal frames. However, if the original features contain linearly inseparable semantic components shared between tasks (e.g., garment shape cues between try-on and reconstruction), the rotation may project out part of the signal or leave residual coupling; no analytic bound is provided showing that the retained subspace still spans the task-specific manifold.

    Authors: We appreciate this theoretical observation. OSP is motivated by the empirical observation that task-specific directions in the shared LoRA bottleneck can be decorrelated via fixed orthogonal rotations, and our experiments (including component ablations in Section 4) show that the unified model preserves or improves task performance relative to independent baselines. This suggests that, in practice, the projected subspaces retain sufficient task-relevant information. We agree that an analytic bound on subspace preservation would be desirable but is non-trivial to derive for high-dimensional diffusion features and lies outside the current scope; we will add a limitations paragraph discussing this point and possible future analysis in the revision. revision: partial

  2. Referee: [Experiments] Experiments (quantitative tables and ablations): The central claim that the unified model matches or exceeds independently trained task-specific baselines requires detailed ablation studies, quantitative tables with error bars, and cross-backbone comparisons to evaluate whether OSP and FNG deliver the reported gains without substantial loss of task-relevant information.

    Authors: We agree that stronger experimental support would improve the manuscript. The current version already contains quantitative comparisons and component ablations demonstrating that OSP and FNG are responsible for the observed gains. In the revision we will expand these with (i) additional ablation tables isolating each component, (ii) error bars computed from multiple random seeds, and (iii) results on further diffusion backbones beyond those already reported, to confirm that the improvements hold without loss of task-specific fidelity. revision: yes

standing simulated objections not resolved
  • Deriving a rigorous analytic bound guaranteeing that OSP retains the full task-specific manifold when input features contain linearly inseparable shared components.

Circularity Check

0 steps flagged

No circularity: empirical method with independent experimental validation

full rationale

The paper introduces OSP as a geometric rotation applied to LoRA bottlenecks and FNG as a Fisher-based guidance term, both defined directly from first principles without reference to the final performance metrics. Claims of unified training outperforming task-specific models are supported solely by benchmark experiments rather than any equation that re-derives the improvement from the same fitted quantities. No self-citation chains, fitted-input predictions, or ansatz smuggling appear in the provided sections. The derivation chain remains self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract provides no explicit free parameters, background axioms, or postulated entities; the contributions are presented as the OSP rotation and FNG guidance mechanisms themselves.

pith-pipeline@v0.9.1-grok · 5729 in / 1116 out tokens · 37585 ms · 2026-06-29T04:26:42.348616+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

69 extracted references · 8 canonical work pages · 4 internal anchors

  1. [1]

    An, X., Zhao, L., Gong, C., Wang, N., Wang, D., Yang, J.: Sharpose: Sparse high- resolutionrepresentationforhumanposeestimation.In:AAAI.vol.38,pp.691–699 (2024)

    2024

  2. [2]

    In: ICML

    Arjovsky, M., Shah, A., Bengio, Y.: Unitary evolution recurrent neural networks. In: ICML. pp. 1120–1128. PMLR (2016)

    2016

  3. [3]

    Qwen2.5-VL Technical Report

    Bai, S., Chen, K., Liu, X., Wang, J., Ge, W., Song, S., Dang, K., Wang, P., Wang, S., Tang, J., et al.: Qwen2. 5-vl technical report. arXiv preprint arXiv:2502.13923 (2025)

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  4. [4]

    arXiv preprint arXiv:2603.09111 (2026)

    Bao, J., Qian, J., Yan, M., Yang, J.: Progressive representation learning for multimodal sentiment analysis with incomplete modalities. arXiv preprint arXiv:2603.09111 (2026)

    work page arXiv 2026

  5. [5]

    In: AAAI

    Bi, J., Wu, Q., Qian, J., Luo, L., Yang, J.: Dual manifold regularization steered robust representation learning for point cloud analysis. In: AAAI. vol. 39, pp. 1844–1852 (2025)

    2025

  6. [6]

    In: ICLR (2018)

    Bińkowski, M., Sutherland, D.J., Arbel, M., Gretton, A.: Demystifying mmd gans. In: ICLR (2018)

    2018

  7. [7]

    In: CVPR

    Chen, X., Huang, L., Liu, Y., Shen, Y., Zhao, D., Zhao, H.: Anydoor: Zero-shot object-level image customization. In: CVPR. pp. 6593–6602 (2024)

    2024

  8. [8]

    In: ECAI (2023)

    Chen, Z., Gao, R., Xiang, T.Z., Lin, F.: Diffusion model for camouflaged object detection. In: ECAI (2023)

    2023

  9. [9]

    In: ICCV

    Chen, Z., Li, Y., Wang, H., Chen, Z., Jiang, Z., Li, J., Wang, Q., Yang, J., Tai, Y.: Ragd: Regional-aware diffusion model for text-to-image generation. In: ICCV. pp. 19331–19341 (2025)

    2025

  10. [10]

    L2P: Unlocking Latent Potential for Pixel Generation

    Chen, Z., Zhu, J., Chen, X., Zhang, J., Chen, J., Zeng, Z., Zhang, W., Wang, C., Yang, J., Tai, Y.: L2p: Unlocking latent potential for pixel generation. arXiv preprint arXiv:2605.12013 (2026)

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  11. [11]

    In: CVPR

    Chen, Z., Zhu, J., Chen, X., Zhang, J., Hu, X., Zhao, H., Wang, C., Yang, J., Tai, Y.: Dip: Taming diffusion models in pixel space. In: CVPR. pp. 36136–36146 (2026)

    2026

  12. [12]

    In: CVPR

    Choi, S., Park, S., Lee, M., Choo, J.: Viton-hd: High-resolution virtual try-on via misalignment-aware normalization. In: CVPR. pp. 14131–14140 (2021) 16 Yang et al

    2021

  13. [13]

    In: ECCV

    Choi, Y., Kwak, S., Lee, K., Choi, H., Shin, J.: Improving diffusion models for authentic virtual try-on in the wild. In: ECCV. pp. 206–235. Springer (2024)

    2024

  14. [14]

    In: ICLR (2025)

    Chong, Z., Dong, X., Li, H., Zhang, S., Zhang, W., Zhang, X., Zhao, H., Jiang, D., Liang, X.: Catvton: Concatenation is all you need for virtual try-on with diffusion models. In: ICLR (2025)

    2025

  15. [15]

    IEEE TPAMI44(5), 2567–2581 (2020)

    Ding, K., Ma, K., Wang, S., Simoncelli, E.P.: Image quality assessment: Unifying structure and texture similarity. IEEE TPAMI44(5), 2567–2581 (2020)

    2020

  16. [16]

    In: CVPR

    Ding, T., Yang, H., Shi, L., Li, J., Hu, X., Yang, J., Tai, Y.: Adaptive depth lightweight rgb-t tracking with holistic token routing. In: CVPR. pp. 20942–20952 (2026)

    2026

  17. [17]

    IEEE TPAMI47(4), 2865–2881 (2025)

    Du, Y., Zhan, J., Li, X., Dong, J., Chen, S., Yang, M.H., He, S.: One-for-all: towards universal domain translation with a single stylegan. IEEE TPAMI47(4), 2865–2881 (2025)

    2025

  18. [18]

    In: AISTATS

    Farajtabar, M., Azizan, N., Mott, A., Li, A.: Orthogonal gradient descent for con- tinual learning. In: AISTATS. pp. 3762–3773. PMLR (2020)

    2020

  19. [19]

    In: CVPR

    Güler, R.A., Neverova, N., Kokkinos, I.: Densepose: Dense human pose estimation in the wild. In: CVPR. pp. 7297–7306 (2018)

    2018

  20. [20]

    In: ICCV

    Guo, H., Zeng, B., Song, Y., Zhang, W., Zhang, C., Liu, J.: Any2anytryon: Lever- aging adaptive position embeddings for versatile virtual clothing tasks. In: ICCV. pp. 19085–19096 (2025)

    2025

  21. [21]

    In: ICCV

    Han, X., Zhu, X., Deng, J., Song, Y.Z., Xiang, T.: Controllable person image synthesis with pose-constrained latent diffusion. In: ICCV. pp. 22768–22777 (2023)

    2023

  22. [22]

    In: CVPR

    Han, X., Wu, Z., Wu, Z., Yu, R., Davis, L.S.: Viton: An image-based virtual try-on network. In: CVPR. pp. 7543–7552 (2018)

    2018

  23. [23]

    Classifier-Free Diffusion Guidance

    Ho, J., Salimans, T.: Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598 (2022)

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  24. [24]

    In: ICML

    Houlsby, N., Giurgiu, A., Jastrzebski, S., Morrone, B., De Laroussilhe, Q., Ges- mundo, A., Attariyan, M., Gelly, S.: Parameter-efficient transfer learning for nlp. In: ICML. pp. 2790–2799. PMLR (2019)

    2019

  25. [25]

    In: ICLR (2022)

    Hu, E.J., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W., et al.: Lora: Low-rank adaptation of large language models. In: ICLR (2022)

    2022

  26. [26]

    In: AAAI

    Hu,X.,Tai,Y.,Zhao,X.,Zhao,C.,Zhang,Z.,Li,J.,Zhong,B.,Yang,J.:Exploiting multimodal spatial-temporal patterns for video object tracking. In: AAAI. vol. 39, pp. 3581–3589 (2025)

    2025

  27. [27]

    arXiv preprint arXiv:2411.10499 (2024)

    Jiang, B., Hu, X., Luo, D., He, Q., Xu, C., Peng, J., Zhang, J., Wang, C., Wu, Y., Fu, Y.: Fitdit: Advancing the authentic garment details for high-fidelity virtual try-on. arXiv preprint arXiv:2411.10499 (2024)

    work page arXiv 2024

  28. [28]

    In: ECCV

    Jin, J., Shen, Y., Fu, Z., Yang, J.: Customized generation reimagined: Fidelity and editability harmonized. In: ECCV. pp. 410–426. Springer (2024)

    2024

  29. [29]

    In: CVPR

    Jin, J., Yu, Z., Shen, Y., Fu, Z., Yang, J.: Latexblend: Scaling multi-concept customized generation with latent textual blending. In: CVPR. pp. 23585–23594 (2025)

    2025

  30. [30]

    PNAS114(13), 3521–3526 (2017)

    Kirkpatrick, J., Pascanu, R., Rabinowitz, N., Veness, J., Desjardins, G., Rusu, A.A., Milan, K., Quan, J., Ramalho, T., Grabska-Barwinska, A., et al.: Overcoming catastrophic forgetting in neural networks. PNAS114(13), 3521–3526 (2017)

    2017

  31. [31]

    Labs, B.F.: Flux.https://github.com/black-forest-labs/flux(2024)

    2024

  32. [32]

    Labs, B.F.: Flux.2.https://github.com/black-forest-labs/flux2(2025)

    2025

  33. [33]

    In: ECCV

    Lee, S., Gu, G., Park, S., Choi, S., Choo, J.: High-resolution virtual try-on with misalignment and occlusion-handled conditions. In: ECCV. pp. 204–219. Springer (2022) OrthoTryOn 17

    2022

  34. [34]

    In: EMNLP

    Lester, B., Al-Rfou, R., Constant, N.: The power of scale for parameter-efficient prompt tuning. In: EMNLP. pp. 3045–3059 (2021)

    2021

  35. [35]

    Lezcano-Casado, M., Martınez-Rubio, D.: Cheap orthogonal constraints in neural networks:Asimpleparametrizationoftheorthogonalandunitarygroup.In:ICML. pp. 3794–3803. PMLR (2019)

    2019

  36. [36]

    Li, X.L., Liang, P.: Prefix-tuning: Optimizing continuous prompts for generation. In: ACL. pp. 4582–4597 (2021)

    2021

  37. [37]

    In: AAAI

    Li, X., Zhan, J., He, S., Xu, Y., Dong, J., Zhang, H., Du, Y.: Personamagic: Stage- regulated high-fidelity face customization with tandem equilibrium. In: AAAI. vol. 39, pp. 4995–5003 (2025)

    2025

  38. [38]

    In: CVPR

    Liu, J., Zhang, J., Rota, P., Sebe, N.: Multi-focal conditioned latent diffusion for person image synthesis. In: CVPR. pp. 16019–16028 (2025)

    2025

  39. [39]

    In: CVPR

    Liu, Z., Luo, P., Qiu, S., Wang, X., Tang, X.: Deepfashion: Powering robust clothes recognition and retrieval with rich annotations. In: CVPR. pp. 1096–1104 (2016)

    2016

  40. [40]

    In: ICLR (2019)

    Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. In: ICLR (2019)

    2019

  41. [41]

    In: CVPR

    Lu, Y., Zhang, M., Ma, A.J., Xie, X., Lai, J.: Coarse-to-fine latent diffusion for pose-guided person image synthesis. In: CVPR. pp. 6420–6429 (2024)

    2024

  42. [42]

    In: ICLR (2016)

    Mishkin, D., Matas, J.: All you need is a good init. In: ICLR (2016)

    2016

  43. [43]

    In: ICLR

    Nan, K., Xie, R., Zhou, P., Fan, T., Yang, Z., Chen, Z., Li, X., Yang, J., Tai, Y.: Openvid-1m: A large-scale high-quality dataset for text-to-video generation. In: ICLR. vol. 2025, pp. 1045–1064 (2025)

    2025

  44. [44]

    In: CVPR

    Parmar, G., Zhang, R., Zhu, J.Y.: On aliased resizing and surprising subtleties in gan evaluation. In: CVPR. pp. 11410–11420 (2022)

    2022

  45. [45]

    In: ICML

    Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: ICML. pp. 8748–8763. PMLR (2021)

    2021

  46. [46]

    In: CVPR

    Ren,Y.,Fan,X.,Li,G.,Liu,S.,Li,T.H.:Neuraltextureextractionanddistribution for controllable person image synthesis. In: CVPR. pp. 13535–13544 (2022)

    2022

  47. [47]

    In: CVPR

    Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR. pp. 10684–10695 (2022)

    2022

  48. [48]

    In: ICLR (2014)

    Saxe, A.M., McClelland, J.L., Ganguli, S.: Exact solutions to the nonlinear dy- namics of learning in deep linear neural networks. In: ICLR (2014)

    2014

  49. [49]

    In: ICML

    Schwarz, J., Czarnecki, W., Luketina, J., Grabska-Barwinska, A., Teh, Y.W., Pas- canu, R., Hadsell, R.: Progress & compress: A scalable framework for continual learning. In: ICML. pp. 4528–4537. PMLR (2018)

    2018

  50. [50]

    In: CVPR

    Sun, K., Xiao, B., Liu, D., Wang, J.: Deep high-resolution representation learning for human pose estimation. In: CVPR. pp. 5693–5703 (2019)

    2019

  51. [51]

    LongCat-Image Technical Report

    Team, M.L., Ma, H., Tan, H., Huang, J., Wu, J., He, J.Y., Gao, L., Xiao, S., Wei, X., Ma, X., et al.: Longcat-image technical report. arXiv preprint arXiv:2512.07584 (2025)

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  52. [52]

    In: BMVC (2025)

    Velioglu, R., Bevandic, P., Chan, R., Hammer, B.: Tryoffdiff: Virtual-try-off via high-fidelity garment reconstruction using diffusion models. In: BMVC (2025)

    2025

  53. [53]

    IEEE TIP13(4), 600–612 (2004)

    Wang, Z., Bovik, A.C., Sheikh, H.R., Simoncelli, E.P.: Image quality assessment: from error visibility to structural similarity. IEEE TIP13(4), 600–612 (2004)

    2004

  54. [54]

    NeurIPS29(2016)

    Wisdom, S., Powers, T., Hershey, J., Le Roux, J., Atlas, L.: Full-capacity unitary recurrent neural networks. NeurIPS29(2016)

    2016

  55. [55]

    arXiv preprint arXiv:2412.08573 (2024)

    Xarchakos, I., Koukopoulos, T.: Tryoffanyone: Tiled cloth generation from a dressed person. arXiv preprint arXiv:2412.08573 (2024)

    work page arXiv 2024

  56. [56]

    In: CVPR

    Xie, Z., Huang, Z., Dong, X., Zhao, F., Dong, H., Zhang, X., Zhu, F., Liang, X.: Gp-vton: Towards general purpose virtual try-on via collaborative local-flow global-parsing learning. In: CVPR. pp. 23550–23559 (2023) 18 Yang et al

    2023

  57. [57]

    In: AAAI

    Xu, Y., Gu, T., Chen, W., Chen, A.: Ootdiffusion: Outfitting fusion based latent diffusion for controllable virtual try-on. In: AAAI. vol. 39, pp. 8996–9004 (2025)

    2025

  58. [58]

    arXiv preprint arXiv:2510.24260 (2025)

    Yang, Z., Chen, Y., Li, Y., He, S., Xu, Y., Dong, J., Yang, J., Du, Y.: Deshadow- mamba: Deshadowing as 1d sequential similarity. arXiv preprint arXiv:2510.24260 (2025)

    work page arXiv 2025

  59. [59]

    In: ECCV

    Yang,Z.,Jiang,Z.,Li,X.,Zhou,H.,Dong,J.,Zhang,H.,Du,Y.:D 4-vton:Dynamic semantics disentangling for differential diffusion based virtual try-on. In: ECCV. pp. 36–52. Springer (2024)

    2024

  60. [60]

    In: ICCV

    Yang, Z., Li, Y., He, S., Li, X., Xu, Y., Dong, J., Du, Y.: Omnivton: Training-free universal virtual try-on. In: ICCV. pp. 16702–16711 (2025)

    2025

  61. [61]

    NeurIPS33, 5824–5836 (2020)

    Yu, T., Kumar, S., Gupta, A., Levine, S., Hausman, K., Finn, C.: Gradient surgery for multi-task learning. NeurIPS33, 5824–5836 (2020)

    2020

  62. [62]

    In: CVPR

    Zhan, J., Sun, X., Zhu, X., Ji, Y., Liu, R., Zhang, L., Zhang, J.: Towards source- aware object swapping with initial noise perturbation. In: CVPR. pp. 4400–4409 (2026)

    2026

  63. [63]

    In: CVPR

    Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: CVPR. pp. 586–595 (2018)

    2018

  64. [64]

    In: AAAI (2026)

    Zhang, W., Jin, Y., Li, X., Zhang, Y., Cong, X., Wang, C., Qiao, F., et al.: Unifit: Towards universal virtual try-on with mllm-guided semantic alignment. In: AAAI (2026)

    2026

  65. [65]

    IEEE TCSVT (2026)

    Zheng, Y., Fang, W., Hu, X., Fan, J., Weng, J., Li, J., Zhang, K., Yang, J.: Dvdpec: Driving-video dehazing via position embedding-based codebook. IEEE TCSVT (2026)

    2026

  66. [66]

    In: CVPR

    Zheng, Y., Zhan, J., He, S., Dong, J., Du, Y.: Curricular contrastive regularization for physics-aware single image dehazing. In: CVPR. pp. 5785–5794 (2023)

    2023

  67. [67]

    In: CVPR

    Zheng, Y., Zhang, K., Zhu, W., Liu, Q., Hu, X., Li, J., Yang, J.: Dvar: Dynamic visual autoregressive modeling for image super-resolution. In: CVPR. pp. 23378– 23387 (2026)

    2026

  68. [68]

    In: CVPR

    Zhou, X., Zhang, B., Zhang, T., Zhang, P., Bao, J., Chen, D., Zhang, Z., Wen, F.: Cocosnet v2: Full-resolution correspondence learning for image translation. In: CVPR. pp. 11465–11475 (2021)

    2021

  69. [69]

    In: CVPR

    Zhu, Z., Huang, T., Shi, B., Yu, M., Wang, B., Bai, X.: Progressive pose attention transfer for person image generation. In: CVPR. pp. 2347–2356 (2019)

    2019