pith. sign in

arxiv: 2605.20807 · v1 · pith:6ATDHDUXnew · submitted 2026-05-20 · 💻 cs.CV

Decomposing Subject-Driven Image Generation via Intermediate Structural Prediction

Hanzhong Guo , Yizhou Yu This is my paper

Pith reviewed 2026-05-21 05:27 UTC · model grok-4.3

classification 💻 cs.CV
keywords subject-driven generationtext-to-image synthesisCanny edge mapsstructural decompositionhigh-frequency detail preservationimage editingdataset constructionknowledge distillation
0
0 comments X

The pith

Predicting an intermediate Canny map before final rendering preserves high-frequency details like logos and text in subject-driven image generation.

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

The paper claims that direct operation in RGB space causes detail loss when substantial edits are needed for subject-driven text-to-image tasks. It introduces a two-stage process that first generates a Canny edge map to capture structure separately from appearance, then conditions the output image on both the predicted map and the original subject reference. A supporting dataset of 100k text-aware pairs is built automatically to handle text consistency across views. Experiments with GPT-4.1 evaluation and distillation tests indicate measurable improvements over baselines, supporting the idea that structural decomposition helps maintain identity fidelity.

Core claim

A two-stage framework decouples structure from appearance by first predicting a Canny map and then rendering the final image conditioned on both the source appearance and the predicted structure, with an automatic pipeline creating a 100k-pair text-aware dataset to aid text handling.

What carries the argument

The two-stage decomposition that first predicts a Canny map as structural guidance before appearance-conditioned rendering.

If this is right

  • High-frequency identity elements remain sharper across edits than in single-stage RGB approaches.
  • Text consistency improves when the dataset construction pipeline enforces cross-view agreement.
  • The method demonstrates gains in both automated metrics and GPT-4.1-based human preference studies.
  • Knowledge distillation from the two-stage model yields a lighter single-stage variant that retains some fidelity benefits.

Where Pith is reading between the lines

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

  • The same structural-intermediate idea could be tested with other edge or depth representations to see if Canny edges are uniquely effective.
  • This decoupling might extend to video or 3D subject-driven generation where consistent structure across frames is required.
  • Future pipelines could insert additional intermediate predictions, such as segmentation or normal maps, for even finer control.

Load-bearing premise

Predicting a Canny map as an intermediate structural representation and conditioning the final rendering on both this map and source appearance will avoid the detail degradation that occurs when methods operate directly in RGB space under substantial edits.

What would settle it

A controlled test on subjects containing fine text, logos, or patterns under large pose or viewpoint changes, measuring whether detail retention metrics exceed those of direct RGB baselines.

Figures

Figures reproduced from arXiv: 2605.20807 by Hanzhong Guo, Yizhou Yu.

Figure 1
Figure 1. Figure 1: High-Fidelity Subject-Driven Generation through Structural Decomposition. Our method excels at preserving the identity of subjects, especially those with high-frequency details like text and patterns. (Left Panel) We demonstrate our primary approach on a frozen FLUX.1-dev backbone. Given a reference image, our method first predicts a target Canny map (Stage 1 Output) that captures the desired structural ch… view at source ↗
Figure 2
Figure 2. Figure 2: An overview of our proposed two-stage framework, data pipeline, and network architecture. (a) Two-Stage Inference Pipeline. Our method decomposes image generation into two distinct stages. Given a source image (cimg) and a text prompt (ctext), Stage 1 is dedicated to predicting the target structure, yielding a Canny edge map (Cˆtgt). This map precisely defines the geometry, pose, and textual layout of the … view at source ↗
Figure 3
Figure 3. Figure 3: Demonstration of the text preservation capability of our Stage 1 Canny predictor. These examples highlight the ef￾fectiveness of training with our ‘TextingSubject100k‘ dataset. tokens {c1, ..., ck}, the operation is: ∆z l t = Projout(MLP(MM-Attn(Norm([z l t , c1, ..., ck])))) (2) This output, ∆z l t , then modulates the original latent via a residual connection: z l+1 t = z l t + ∆z l t . The true elegance… view at source ↗
Figure 4
Figure 4. Figure 4: Qualitative comparison with baseline methods. Each example shows the reference image (Ref) alongside the outputs from baseline models (OminiControl, FLUX-Kontext) and our method. Compared to the baselines, our method demonstrates superior per￾formance in generating images that are more consistent with the text prompt while more faithfully preserving the subject’s identity and intricate textual details [PI… view at source ↗
Figure 5
Figure 5. Figure 5: Additional qualitative results on text-free objects. Each group displays a reference image (Ref) and two different scenes generated by our model. These examples showcase the model’s ability to robustly place the source object into diverse new environments and styles while maintaining its core identity features [PITH_FULL_IMAGE:figures/full_fig_p012_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Additional qualitative results on objects with text. Each group shows a reference image (Ref) with text and two outputs generated in new scenes. Our method successfully preserves the legibility, style, and content of the text on the object even under significant changes in background and lighting, a key contribution of our work [PITH_FULL_IMAGE:figures/full_fig_p013_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Additional qualitative results on objects with text. Each group shows a reference image (Ref) with text and two outputs generated in new scenes. Our method successfully preserves the legibility, style, and content of the text on the object even under significant changes in background and lighting, a key contribution of our work [PITH_FULL_IMAGE:figures/full_fig_p014_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Examples from the augmented ‘Subject200k‘ dataset. These image triplets are created by taking a source image and applying Bagel to generate novel, rotated views. This data augmentation strategy helps the model learn a robust representation of object identity across different viewpoints [PITH_FULL_IMAGE:figures/full_fig_p015_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Examples from our custom ‘TextingSubject100k‘ dataset. Each triplet shows an object with text, with its views rotated by Bagel. We employ a strict OCR filtering process to ensure the text remains consistent and legible across views, providing high-quality training data for our text-aware model [PITH_FULL_IMAGE:figures/full_fig_p016_9.png] view at source ↗
read the original abstract

Subject-driven text-to-image generation still struggles to preserve high-frequency identity details such as logos, patterns, and text. Existing methods typically operate directly in RGB space, which often leads to detail degradation under substantial edits. We propose a two-stage framework that decouples structure from appearance by first predicting a Canny map and then rendering the final image conditioned on both the source appearance and the predicted structure. To improve text handling, we further introduce a fully automatic pipeline that constructs a 100k-pair text-aware dataset with cross-view textual consistency. Experiments, including GPT-4.1-based evaluation and a knowledge distillation study, show clear gains over selected baselines and suggest that intermediate structural prediction is an effective route for high-fidelity subject-driven generation. Our dataset and code will be made publicly available.

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

3 major / 2 minor

Summary. The paper introduces a two-stage framework for subject-driven text-to-image generation. In the first stage, it predicts an intermediate Canny edge map from the subject image and text prompt. In the second stage, it generates the final image by conditioning on both the predicted Canny map and the source appearance features. To address challenges with text and logos, the authors construct a 100k-pair text-aware dataset ensuring cross-view textual consistency using an automatic pipeline. The experiments, including evaluations with GPT-4.1 and a knowledge distillation study, report improvements over selected baselines, suggesting that intermediate structural prediction helps in maintaining high-fidelity details.

Significance. If validated, this decomposition approach could significantly improve the preservation of high-frequency identity details like text, patterns, and logos in subject-driven generation tasks, offering a more robust alternative to direct RGB-space methods that suffer from detail degradation under edits. The public release of the dataset and code would further enhance its impact by enabling reproducibility and further research.

major comments (3)
  1. §3.1: The central claim that first-stage Canny prediction plus appearance conditioning reliably prevents high-frequency detail loss under substantial edits rests on an assumption that may not hold, because Canny is a fixed low-level edge detector that discards texture, color gradients, and fine semantic layout (e.g., exact stroke order in logos or text). If the predicted Canny deviates from the edit-specified geometry on these elements, the second-stage renderer still operates from incomplete structure.
  2. §5.2, Table 3: The GPT-4.1-based evaluation and knowledge distillation study are described as showing 'clear gains,' but the manuscript provides no quantitative metrics, baseline details, ablation results, or statistical significance tests. Without these, it is not possible to verify whether the data actually support the central claim that intermediate structural prediction is effective.
  3. §4.1: The 100k text-aware dataset targets cross-view consistency but does not guarantee that the learned Canny predictor will produce geometrically accurate maps for novel prompts involving text or logos; the automatic construction pipeline lacks reported human verification or error analysis on semantic fidelity.
minor comments (2)
  1. §2: The related-work discussion could include more recent references on structural conditioning and edge-based guidance in diffusion models to better situate the contribution.
  2. Figure 2: The pipeline diagram would benefit from clearer labeling of the conditioning inputs to the second-stage renderer.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the detailed and constructive review. We address each major comment point by point below, indicating where revisions have been made to the manuscript.

read point-by-point responses

  1. Referee: §3.1: The central claim that first-stage Canny prediction plus appearance conditioning reliably prevents high-frequency detail loss under substantial edits rests on an assumption that may not hold, because Canny is a fixed low-level edge detector that discards texture, color gradients, and fine semantic layout (e.g., exact stroke order in logos or text). If the predicted Canny deviates from the edit-specified geometry on these elements, the second-stage renderer still operates from incomplete structure.

    Authors: We agree that Canny is a low-level edge detector and does not encode texture or fine semantic details on its own. In our framework the second stage is explicitly conditioned on appearance features extracted from the source subject image; these features are responsible for supplying texture, color gradients, and high-frequency identity elements while the predicted Canny map supplies only geometric guidance. We have revised §3.1 to clarify this complementary relationship and added qualitative examples that illustrate preservation of text and logos under edits where the Canny map is only approximate. revision: yes

  2. Referee: §5.2, Table 3: The GPT-4.1-based evaluation and knowledge distillation study are described as showing 'clear gains,' but the manuscript provides no quantitative metrics, baseline details, ablation results, or statistical significance tests. Without these, it is not possible to verify whether the data actually support the central claim that intermediate structural prediction is effective.

    Authors: We acknowledge that the original presentation of the GPT-4.1 evaluation and knowledge-distillation study lacked sufficient quantitative detail. In the revised manuscript we have expanded §5.2 and updated Table 3 to report concrete metrics (preference scores, consistency rates), baseline specifications, ablation results, and statistical significance (paired t-tests with p-values). These additions directly support the claim that intermediate structural prediction yields measurable improvements. revision: yes

  3. Referee: §4.1: The 100k text-aware dataset targets cross-view consistency but does not guarantee that the learned Canny predictor will produce geometrically accurate maps for novel prompts involving text or logos; the automatic construction pipeline lacks reported human verification or error analysis on semantic fidelity.

    Authors: The automatic pipeline was designed for scalability while enforcing cross-view textual consistency through filtering heuristics. We accept that human verification strengthens the claim. The revised manuscript now includes a human evaluation on a 500-pair random subset together with an error analysis of semantic fidelity; these results are reported in §4.1 and the supplementary material. revision: yes

Circularity Check

0 steps flagged

No load-bearing circularity; empirical pipeline with external grounding

full rationale

The paper presents a two-stage empirical framework that first predicts a Canny edge map as intermediate structure and then conditions a renderer on both the predicted map and source appearance, augmented by an automatically constructed 100k text-aware dataset. No equations, parameter fits, or derivations are described that reduce to self-definition or rename fitted inputs as predictions. The central claim rests on comparative experiments (including GPT-4.1 evaluation and distillation) rather than any self-citation chain or uniqueness theorem imported from prior author work. Standard Canny detection and conditioning provide independent grounding, making the derivation self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract provides no explicit free parameters, axioms, or invented entities; the method builds on standard Canny edge detection and conditioning techniques from prior literature.

pith-pipeline@v0.9.0 · 5655 in / 1155 out tokens · 41362 ms · 2026-05-21T05:27:24.049296+00:00 · methodology

discussion (0)

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

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

31 extracted references · 31 canonical work pages · 9 internal anchors

  1. [1]

    Stephen Batifol, Andreas Blattmann, Frederic Boesel, Sak- sham Consul, Cyril Diagne, Tim Dockhorn, Jack English, Zion English, Patrick Esser, Sumith Kulal, et al. Flux. 1 kontext: Flow matching for in-context image generation and editing in latent space.arXiv e-prints, pages arXiv–2506,

    work page

  2. [2]

    FLUX: A New Era for Fast and High- Quality Image Generation.https://github.com/ black-forest-labs/flux, 2024

    Black Forest Labs. FLUX: A New Era for Fast and High- Quality Image Generation.https://github.com/ black-forest-labs/flux, 2024. Accessed: Septem- ber 5, 2025. 4, 6

    work page 2024

  3. [3]

    A computational approach to edge detection

    John Canny. A computational approach to edge detection. IEEE Transactions on pattern analysis and machine intelli- gence, (6):679–698, 1986. 2

    work page 1986

  4. [4]

    Textdif- fuser: Diffusion models as text painters.arXiv preprint arXiv:2305.10855, 2023

    Jingye Chen, Yupan Zhang, Qing Li, Zhaoliang Liu, Gyun- gin Yang, Seung-Hwan Lee, and Jinyoung Kim. Textdif- fuser: Diffusion models as text painters.arXiv preprint arXiv:2305.10855, 2023. 3

    work page arXiv 2023

  5. [5]

    Emerging Properties in Unified Multimodal Pretraining

    Chaorui Deng, Deyao Zhu, Kunchang Li, Chenhui Gou, Feng Li, Zeyu Wang, Shu Zhong, Weihao Yu, Xiaonan Nie, Ziang Song, et al. Emerging properties in unified multimodal pretraining.arXiv preprint arXiv:2505.14683, 2025. 6

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  6. [6]

    DreamDA: Generative data augmentation with diffusion models.arXiv preprint arXiv:2403.12803, 2024

    Yunxiang Fu, Chaoqi Chen, Yu Qiao, and Yizhou Yu. DreamDA: Generative data augmentation with diffusion models.arXiv preprint arXiv:2403.12803, 2024. 3

    work page arXiv 2024

  7. [7]

    LaMamba-Diff: Linear-time high-fidelity diffusion models based on local at- tention and mamba

    Yunxiang Fu, Chaoqi Chen, and Yizhou Yu. LaMamba-Diff: Linear-time high-fidelity diffusion models based on local at- tention and mamba. InProceedings of the British Machine Vision Conference, Sheffield, UK, 2025. 3

    work page 2025

  8. [8]

    An Image is Worth One Word: Personalizing Text-to-Image Generation using Textual Inversion

    Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patash- nik, Amit H Bermano, Gal Chechik, and Daniel Cohen- Or. An image is worth one word: Personalizing text-to- image generation using textual inversion.arXiv preprint arXiv:2208.01618, 2022. 2

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  9. [9]

    Seedream 3.0 Technical Report

    Yu Gao, Lixue Gong, Qiushan Guo, Xiaoxia Hou, Zhichao Lai, Fanshi Li, Liang Li, Xiaochen Lian, Chao Liao, Liyang Liu, et al. Seedream 3.0 technical report.arXiv preprint arXiv:2504.11346, 2025. 5

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  10. [10]

    Real-time identity defenses against malicious personalization of diffusion models.arXiv preprint arXiv:2412.09844, 2024

    Hanzhong Guo, Shen Nie, Chao Du, Tianyu Pang, Hao Sun, and Chongxuan Li. Real-time identity defenses against ma- licious personalization of diffusion models.arXiv preprint arXiv:2412.09844, 2024. 3

    work page arXiv 2024

  11. [11]

    Real-time one-step diffusion-based expressive portrait videos generation.arXiv preprint arXiv:2412.13479, 2024

    Hanzhong Guo, Hongwei Yi, Daquan Zhou, Alexan- der William Bergman, Michael Lingelbach, and Yizhou Yu. Real-time one-step diffusion-based expressive portrait videos generation.arXiv preprint arXiv:2412.13479, 2024. 3

    work page arXiv 2024

  12. [12]

    Gans trained by a two time-scale update rule converge to a local nash equilib- rium.Advances in neural information processing systems, 30, 2017

    Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilib- rium.Advances in neural information processing systems, 30, 2017. 6

    work page 2017

  13. [13]

    Lora: Low-rank adaptation of large language models.ICLR, 1(2):3, 2022

    Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen- Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models.ICLR, 1(2):3, 2022. 4

    work page 2022

  14. [14]

    Multi-concept customization of text-to-image diffusion,

    Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Customdiffusion: Multi- concept customization of text-to-image diffusion.arXiv preprint arXiv:2212.04488, 2022. 3

    work page arXiv 2022

  15. [15]

    Photomaker: Customizing re- alistic human photos via stacked id embedding

    Zhen Li, Mingdeng Cao, Xintao Wang, Zhongang Qi, Ming- Ming Cheng, and Ying Shan. Photomaker: Customizing re- alistic human photos via stacked id embedding. InProceed- ings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8640–8650, 2024. 3

    work page 2024

  16. [16]

    Flow Matching for Generative Modeling

    Yaron Lipman, Ricky TQ Chen, Heli Ben-Hamu, Maximil- ian Nickel, and Matt Le. Flow matching for generative mod- eling.arXiv preprint arXiv:2210.02747, 2022. 4

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  17. [17]

    Prodigy: An expeditiously adaptive parameter-free learner.arXiv preprint arXiv:2306.06101, 2023

    Konstantin Mishchenko and Aaron Defazio. Prodigy: An expeditiously adaptive parameter-free learner.arXiv preprint arXiv:2306.06101, 2023. 6

    work page arXiv 2023

  18. [18]

    T2I-Adapter: Learning Adapters to Dig out More Controllable Ability for Text-to-Image Diffusion Models

    Chong Mou, Xintao Wang, Liangbin Xie, Jian Zhang, Zhon- gang Qi, Ying Shan, and Xiaohu Qie. T2I-Adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models.arXiv preprint arXiv:2302.08453, 2023. 3

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  19. [19]

    Scalable diffusion models with transformers

    William Peebles and Saining Xie. Scalable diffusion models with transformers. InProceedings of the IEEE/CVF Inter- national Conference on Computer Vision, pages 4195–4205,

    work page

  20. [20]

    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. Learn- ing transferable visual models from natural language super- vision.arXiv preprint arXiv:2103.00020, 2021. 6

    work page internal anchor Pith review Pith/arXiv arXiv 2021

  21. [21]

    High-resolution image synthesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj ¨orn Ommer. High-resolution image synthesis with latent diffusion models. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684–10695, 2022. 2

    work page 2022

  22. [22]

    Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation

    Nataniel Ruiz, Yuanzhe Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22500– 22510, 2023. 2, 6

    work page 2023

  23. [23]

    Seedream 4.0: Toward Next-generation Multimodal Image Generation

    Team Seedream, Yunpeng Chen, Yu Gao, Lixue Gong, Meng Guo, Qiushan Guo, Zhiyao Guo, Xiaoxia Hou, Weilin Huang, Yixuan Huang, et al. Seedream 4.0: Toward next- generation multimodal image generation.arXiv preprint arXiv:2509.20427, 2025. 7

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  24. [24]

    Ominicontrol: Minimal and uni- versal control for diffusion transformer.arXiv preprint arXiv:2401.15098, 2024

    Zhenxiong Tan, Songhua Liu, Xingyi Yang, Qiaochu Xue, and Xinchao Wang. Ominicontrol: Minimal and uni- versal control for diffusion transformer.arXiv preprint arXiv:2401.15098, 2024. 1, 2, 3, 6, 7

    work page arXiv 2024

  25. [25]

    InstantID: Zero-shot Identity-Preserving Generation in Seconds

    Qixun Wang, Xu Bai, Haofan Wang, Zekui Qin, Anthony Chen, Huaxia Li, Xu Tang, and Yao Hu. Instantid: Zero-shot identity-preserving generation in seconds.arXiv preprint arXiv:2401.07519, 2024. 3

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  26. [26]

    Automatic photo adjustment using deep neu- ral networks.ACM Trans

    Zhicheng Yan, Hao Zhang, Baoyuan Wang, Sylvain Paris, and Yizhou Yu. Automatic photo adjustment using deep neu- ral networks.ACM Trans. Graph., 35(2):11:1–11:15, 2016. 3

    work page 2016

  27. [27]

    IP-Adapter: Text Compatible Image Prompt Adapter for Text-to-Image Diffusion Models

    Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. IP- Adapter: Text compatible image prompt adapter for text-to- image diffusion models.arXiv preprint arXiv:2308.06721,

    work page internal anchor Pith review Pith/arXiv arXiv

  28. [28]

    Glyphcontrol: A conditional control module for accurate and consistent font generation.arXiv preprint arXiv:2402.13426, 2024

    Yuxin Zeng, Yahan Zhang, Yang Chen, Yidong Liu, and Yuan Zhang. Glyphcontrol: A conditional control module for accurate and consistent font generation.arXiv preprint arXiv:2402.13426, 2024. 3

    work page arXiv 2024

  29. [29]

    Adding conditional control to text-to-image diffusion models

    Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3836–3847, 2023. 3

    work page 2023

  30. [30]

    CarveMix: A sim- ple data augmentation method for brain lesion segmentation

    Xinru Zhang, Chenghao Liu, Ni Ou, Xiangzhu Zeng, Zhizheng Zhuo, Yunyun Duan, Xiaoliang Xiong, Yizhou Yu, Zhiwen Liu, Yaou Liu, and Chuyang Ye. CarveMix: A sim- ple data augmentation method for brain lesion segmentation. NeuroImage, 271:120041, 2023. 3

    work page 2023

  31. [31]

    1984” and “Awaken Minds

    Yuxuan Zhang, Yiren Song, Jiaming Liu, Rui Wang, Jinpeng Yu, Hao Tang, Huaxia Li, Xu Tang, Yao Hu, Han Pan, et al. Ssr-encoder: Encoding selective subject representation for subject-driven generation. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8069–8078, 2024. 6 Decomposing Subject-Driven Image Generation vi...

    work page 2024