Pith. sign in

REVIEW 5 major objections 5 minor 2 cited by

The paper claims that a per-timestep spatial mask built from attention correspondence and pixel entropy turns Diffusion-DPO into a training procedure that preserves multiple subjects' identities, cuts attribute leakage, and reaches state-of

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

FocusDPO adds dynamic spatial weighting to preference-based fine-tuning, improving subject fidelity and reducing attribute leakage in multi-subject personalized image generation.

T0 review reviewed 2026-08-05 challenge →

load-bearing objection A plausible, empirically supported training-time method for multi-subject personalization, with a real but non-fatal theoretical gap around the masked DPO objective and a hyperparameter-tuning concern. the 5 major comments →

arxiv 2509.01181 v1 pith:XVEH6Y3H submitted 2025-09-01 cs.CV cs.AI

FocusDPO: Dynamic Preference Optimization for Multi-Subject Personalized Image Generation via Adaptive Focus

Qiaoqiao Jin , Siming Fu , Dong She , Weinan Jia , Hualiang Wang , Mu Liu , Jidong Jiang This is my paper
classification cs.CV cs.AI
keywords multi-subject personalized generationdirect preference optimizationdiffusion modelsattribute leakagesubject fidelityattention correspondenceentropy weightingDisrupted-Instance Pair dataset
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The reading

FocusDPO targets a specific failure mode in multi-subject personalized image generation: when a diffusion model is aligned with user preferences, spreading the optimization pressure evenly over the whole image lets easy regions and backgrounds absorb the training signal while subject details drift and bleed into one another. The paper claims that routing DPO's per-timestep loss through a spatial mask — built from cross-attention correspondence between generated and reference subjects plus a Shannon-entropy measure of local detail — makes existing pre-trained personalization models preserve each subject's identity better and leak fewer attributes across subjects. On the DreamBench benchmark the method reports state-of-the-art scores when layered onto two backbones, IP-Adapter-Plus with SDXL and UNO with FLUX, with consistent gains over plain diffusion DPO and over the fixed-patch PatchDPO. The reason to care is practical: the upgrade lives entirely in the training loss, so it adds no cost or machinery at generation time.

Core claim

The central claim is that preference-optimization pressure should land neither on the whole image nor on a fixed patch grid, but on a per-timestep focus field that shifts as denoising progresses. FocusDPO builds the field from two masks: a Structure-Preserving Attention Field, which uses cross-layer multi-modal attention to find where reference-subject tokens map into the noisy latent and flags subject regions the model misaligns, and a Detail-Preserving Complexity Estimator, a normalized Shannon-entropy map that up-weights texture-rich patches. The two fuse adaptively: if the attention field already covers enough of the prior subject mask it drives the loss alone; otherwise complexity weigh

What carries the argument

The load-bearing object is the dynamic focus mask M multiplied into the Diffusion-DPO loss. M fuses two components. Ms, the Structure-Preserving Attention Field, pools reference-image token embeddings, matches them by cosine similarity against every target token across the multi-modal attention layers, keeps the top-K target positions, and subtracts them from the segmentation prior Mprior to expose subject regions the model currently fails to align. Md, the Detail-Preserving Complexity Estimator, scores each local patch by Shannon entropy of its grayscale intensities and normalizes globally to [0,1]. A focus-coverage ratio Afocus = ||Ms||1/||Mprior||1 decides fusion: above threshold τ=0.1, M

Load-bearing premise

That inserting the spatial mask into the per-noise-residual norms of the Diffusion-DPO loss still defines a valid preference-optimization objective: the masked loss is written down, not derived from any preference model, so if it does not correspond to a coherent reward, the reported gains could come from spatial denoising regularization rather than preference alignment.

What would settle it

Train FocusDPO on the same DIP pairs with the same mask coverage but with the focus mask randomly permuted per timestep; if the random-mask run reproduces the DINO/CLIP gains over plain DPO, the effect is generic spatial regularization rather than semantic correspondence. Alternatively, test the implied reward: after training, check whether the masked loss assigns higher likelihood to winning than losing pairs under the Bradley–Terry model — genuine preference alignment predicts a clean gap, spatial regularization does not.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Personalization backbones can be upgraded purely at training time: the masks live in the loss, so any model that already supports DPO-style LoRA fine-tuning inherits FocusDPO without architectural or inference changes.
  • Attribute leakage shrinks because the attention field directs corrective pressure at exactly the subject regions the model misaligns, while regions it already matches — including background — are down-weighted.
  • The reported gains reproduce across two backbone families (U-Net and DiT), suggesting the masking strategy transfers rather than fitting one architecture.
  • Because DIP pairs match in prompt and composition and differ only inside the subject mask, the preference signal is less confounded by background differences than global preference data, which the authors argue is what lets DPO focus on identity rather than scene.
  • Following the authors' timestep argument, training signal automatically reallocates from global structure at high noise to fine texture at low noise, the schedule they claim fixed-patch methods cannot express.

Where Pith is reading between the lines

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

  • The weighting mechanism is loss-agnostic: nothing in the attention field or entropy estimator depends on DPO's Bradley–Terry framing, so the same masks could weight reward-gradient objectives (DDPO- or DRaFT-style) or plain denoising losses — a transfer the paper does not test.
  • The top-K attention matching maps each reference region to its strongest target token; inverting that correspondence would yield per-subject masks, which could extend the method to an arbitrary number of subjects rather than the pairwise fusion shown here.
  • A direct way to verify the timestep-dependence claim: measure how often Afocus falls below τ at high versus low noise. The paper's reported τ=0.1 would predict the attention field alone dominates early denoising and complexity weighting engages late — a split the published ablations do not show.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

5 major / 5 minor

Summary. The paper proposes FocusDPO, a fine-tuning method for multi-subject personalized image generation. The authors construct a synthetic preference-pair dataset (DIP) whose negative samples are generated by localized inpainting/disruption of subject regions, and then train a spatially weighted Diffusion-DPO objective on top of existing backbones (IP-Adapter-Plus on SDXL and UNO on FLUX.1 dev). The key novelty is a dynamic, per-timestep spatial mask M that combines a Structure-Preserving Attention Field (derived from cross-attention correspondence between generated latents and reference tokens) and a Detail-Preserving Complexity Estimator (based on grayscale entropy), with a schedule depending on the focus coverage ratio. The paper reports consistent improvements over SFT, DPO, and PatchDPO baselines on DreamBench, and claims state-of-the-art single- and multi-subject results.

Significance. If the central claim is accepted, FocusDPO would be a useful recipe for making preference-based fine-tuning more spatially selective in subject-driven generation, and the DIP dataset construction might be reused by other methods. The paper has useful components: it evaluates on two backbones, compares against the relevant PatchDPO baseline, includes qualitative ablations of the two mask components, and provides a hyperparameter sensitivity study. However, the method's theoretical status as a DPO variant is questionable, the DIP dataset is self-generated by the same model family as the main test backbone, and the quantitative evidence lacks error bars or significance testing. These issues currently prevent me from endorsing the state-of-the-art claim as established.

major comments (5)
  1. [Eq. (10), §4.2] The masked loss in Eq. (10) is not derived from any preference model. Diffusion-DPO's Eq. (1) is a lower bound on a Bradley-Terry preference likelihood for whole images; inserting the spatial mask M inside each noise-residual norm produces a new objective, but no per-patch Bradley-Terry factorization or reward model is provided. Since M is computed from xw_t and xr (Eqs. 3–9), the negative-sample residual is weighted by a mask derived from the positive sample and references. The reported gains could therefore come from spatial regularization or curriculum weighting rather than from preference alignment. The paper should either derive the masked objective from a preference model, or explicitly characterize it as a heuristic spatial weighting and add a control ablation isolating the preference-alignment contribution from the spatial-mask contribution.
  2. [Eqs. (3)–(6), §4.2] The definition of the Structure-Preserving Attention Field is underspecified at a load-bearing point. Mprior is defined as a binary pixel/latent mask of size H×W, while M′ is constructed on attention tokens of size pxt, and Eq. (6) writes Ms = Mprior \ M′ as a set difference between objects of different spaces. The definition of K in Eq. (4) is also circular/incomplete: J is said to contain the K largest scores, with K described only as 'the number of tokens of reference image xr in xt', and arg max returning a set of indices is not well-defined. The paper needs a precise statement of the token-to-latent coordinate mapping, how masks are resized/registered, and how K is computed. Without this, the central mask cannot be reproduced.
  3. [§4.1, §5.3.2] The DIP preference pairs are synthetically generated by FLUX, and the main DiT backbone (UNO) uses FLUX.1 dev. This makes the training signal partly self-generated by the same model family under test. The statement in §5.3.2 that the results 'demonstrate that improvements stem from our optimization strategy rather than dataset biases' is not supported: the SFT/DPO/FocusDPO comparison on the same DIP controls for the optimization scheme, but it does not control for the bias introduced by training on FLUX-generated preference data. The paper should discuss the implications of self-generated preference data and, if possible, include an evaluation on human-annotated or independently generated preference pairs.
  4. [Tables 1–2, Appendix A.3] The quantitative evaluation reports single unreplicated numbers with no error bars, confidence intervals, or significance tests. Several key differences are very small (e.g., CLIP-I 0.840 vs 0.838 for IP-Adapter-Plus, and 0.739 vs 0.733 for multi-subject UNO), so the stated gains may not be robust. Moreover, the hyperparameters τ=0.1 and γ=0.3 are selected using 'validation performance measured using our standard evaluation metrics'; if this validation set overlaps with the DreamBench test set used in Tables 1 and 2, the state-of-the-art claim is vulnerable to selection on the test benchmark. The authors should report the validation/evaluation split, the number of seeds, and variance or significance statistics.
  5. [§5.3.1, Figs. 6–7, Appendix A] The ablation of the two main components (Ms and Md) is only qualitative. There are no quantitative tables reporting DINO/CLIP-I/CLIP-T for the w/o Ms, w/o Md, and w/o Mprior configurations. Since the paper's central mechanism is the interaction of these masks, quantitative ablations on DreamBench are needed to substantiate the claim that both components contribute, especially given that the full method's gains over DPO are small on CLIP-I.
minor comments (5)
  1. [Table 2] Typo: 'sythesis' should be 'synthesis'.
  2. [§2.1] Typo: 'everages' should be 'leverages'. Also, reference [39] (OmniControl human motion generation) appears to be the wrong citation for the image-generation OmniControl method; the related work text seems to conflate two similarly named works.
  3. [§4.2, Eq. (3)] The notation 'CLS_xr' is confusing because no CLS token is defined; if it is a pooled embedding, state the pooling operation explicitly.
  4. [Fig. 9] The label 'FoucusDPO' in Figure 9 is a typo; should be 'FocusDPO'. Also, the appendix refers to 'Section 4.3' for ablations, but the ablation results are in Section 5.3; the cross-reference should be corrected.
  5. [Eq. (9) and Algorithm 1] The fusion mask M is written with set-like operations (Ms) and elementwise products (Md ⊙ Mprior); the paper should clarify whether M is binary or real-valued and how it is normalized before being applied in Eq. (10), since this affects the scale of the loss.

Circularity Check

0 steps flagged

No significant circularity: FocusDPO's central claim is an empirical improvement evaluated on external DreamBench benchmarks; the masked DPO loss is a heuristic extension, not a circular reduction.

full rationale

The paper's core claim is that FocusDPO improves subject fidelity and reduces attribute leakage over existing personalized generation models. This is supported by quantitative results on DreamBench using DINO, CLIP-I, and CLIP-T, which are external benchmarks not derived from the training loss. The training data (DIP) is constructed by generating positive/negative pairs with FLUX, and the UNO backbone is also FLUX-based; while this introduces a potential distributional bias, it does not make the benchmark evaluation a restatement of the training objective. The masked DPO loss in Eq. 10 is a heuristic modification of Diffusion-DPO, and the paper does not claim it is derived from a Bradley-Terry preference model; lack of theoretical grounding is a correctness concern, not circularity. Hyperparameters tau and gamma are tuned on validation, not fitted to the test predictions in a way that forces the reported numbers. Self-citations (e.g., [14], [15]) are minor and not load-bearing: the entropy measure is standard, and the evaluation protocol reference is ancillary. No equation or prediction reduces by construction to the inputs, so no circular step is present.

Axiom & Free-Parameter Ledger

3 free parameters · 5 axioms · 0 invented entities

The method's central claim rests on hand-set thresholds (tau, gamma, curation scores), a self-generated preference dataset, and several domain assumptions about attention correspondence and entropy as complexity. No new physical or architectural entities are introduced beyond the naming of the two mask components.

free parameters (3)
  • tau (focus coverage threshold) = 0.1
    Threshold deciding whether to use only the structure mask Ms or fuse it with the complexity mask Md (Eq. 9). Selected by hyperparameter sweep in Appendix A.3 on the evaluation benchmark.
  • gamma (fusion coefficient) = 0.3
    Balances the structural attention field and the detail complexity estimator in the fused mask (Eq. 9). Selected by hyperparameter sweep in Appendix A.3.
  • DIP curation thresholds = positive >=9, negative <=6 (10-point scale)
    Consensus inclusion criteria used to filter the synthetic preference pairs in the DIP dataset (Appendix B.3.1). Hand-set and directly determine the training signal.
axioms (5)
  • standard math Diffusion-DPO loss in Eq. 1 is a valid alignment objective for fine-tuning diffusion models.
    The paper builds on Wallace et al. [34] without re-deriving the loss; background from cited prior literature.
  • domain assumption Multimodal attention similarity (Eq. 3) between noised latent and reference image encodes semantically meaningful subject correspondence.
    Assumed in Section 4.2 and used to construct the Structure-Preserving Attention Field.
  • domain assumption Grayscale Shannon entropy of local patches is a valid proxy for preference-critical visual complexity.
    Introduced in Eq. 7-8 in Section 4.2; motivates the Detail-Preserving Complexity Estimator.
  • domain assumption Synthetic preference pairs produced by FLUX generation plus inpainted disruptions are valid win/lose training signals for DPO.
    DIP dataset construction in Section 4.1 assumes these pairs isolate subject-level inconsistency without other confounders.
  • domain assumption DreamBench evaluation and CLIP/DINO cosine similarities are reliable measures of subject fidelity and text alignment.
    Used in Section 5 as the sole quantitative validation of the central claim.

reviewed 2026-08-05 · how reviews work

0 comments
Cite this review

Pith. "Pith review of FocusDPO: Dynamic Preference Optimization for Multi-Subject Personalized Image Generation via Adaptive Focus." pith.science (2026). https://pith.science/paper/XVEH6Y3H

@misc{pith2026250901181,
  author       = {Pith},
  title        = {Pith review of: FocusDPO: Dynamic Preference Optimization for Multi-Subject Personalized Image Generation via Adaptive Focus},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/XVEH6Y3H}},
  note         = {Machine review of arXiv:2509.01181}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Multi-subject personalized image generation aims to synthesize customized images containing multiple specified subjects without requiring test-time optimization. However, achieving fine-grained independent control over multiple subjects remains challenging due to difficulties in preserving subject fidelity and preventing cross-subject attribute leakage. We present FocusDPO, a framework that adaptively identifies focus regions based on dynamic semantic correspondence and supervision image complexity. During training, our method progressively adjusts these focal areas across noise timesteps, implementing a weighted strategy that rewards information-rich patches while penalizing regions with low prediction confidence. The framework dynamically adjusts focus allocation during the DPO process according to the semantic complexity of reference images and establishes robust correspondence mappings between generated and reference subjects. Extensive experiments demonstrate that our method substantially enhances the performance of existing pre-trained personalized generation models, achieving state-of-the-art results on both single-subject and multi-subject personalized image synthesis benchmarks. Our method effectively mitigates attribute leakage while preserving superior subject fidelity across diverse generation scenarios, advancing the frontier of controllable multi-subject image synthesis.

discussion (0)

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

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Fusion in Your Way: Aligning Image Fusion with Heterogeneous Demands via Direct Preference Optimization

    cs.CV 2026-05 unverdicted novelty 7.0

    DPOFusion uses direct preference optimization on property-aligned and preference-controllable latent diffusion models to produce adaptive infrared-visible image fusions aligned with heterogeneous human and machine vis...

  2. Scaling Multi-Reference Image Generation with Dynamic Reward Optimization

    cs.CV 2026-06 unverdicted novelty 6.0

    Introduces OmniRef-Bench benchmark and DyRef two-stage framework using Difficulty-aware Advantage Reweighting and Discriminative Reward Scaling to improve open-source models on complex multi-reference image generation.

Reference graph

Works this paper leans on

73 extracted references · 47 canonical work pages · cited by 2 Pith papers · 1 internal anchor

  1. [1]

    Training a helpful and harmless assistant with reinforcement learning from human feedback.arXiv preprint arXiv:2204.05862, 2022

    Yuntao Bai, Andy Jones, Kamal Ndousse, Amanda Askell, Anna Chen, Nova DasSarma, Dawn Drain, Stanislav Fort, Deep Ganguli, Tom Henighan, et al. Training a helpful and harmless assistant with reinforcement learning from human feedback.arXiv preprint arXiv:2204.05862, 2022

    Pith/arXiv arXiv 2022

  2. [2]

    Training diffusion models with reinforcement learning

    Kevin Black, Michael Janner, Yilun Du, Ilya Kostrikov, and Sergey Levine. Training diffusion models with reinforcement learning. arXiv preprint arXiv:2305.13301, 2023

    Pith/arXiv arXiv 2023

  3. [3]

    Open problems and fundamental limitations of reinforcement learning from human feedback.arXiv preprint arXiv:2307.15217, 2023

    Stephen Casper, Xander Davies, Claudia Shi, Thomas Krendl Gilbert, Jérémy Scheurer, Javier Rando, Rachel Freedman, Tomasz Korbak, David Lindner, Pedro Freire, et al. Open problems and fundamental limitations of reinforcement learning from human feedback.arXiv preprint arXiv:2307.15217, 2023

    Pith/arXiv arXiv 2023

  4. [4]

    Xverse: Consistent multi-subject control of identity and semantic attributes via dit modulation.arXiv preprint arXiv:2506.21416, 2025

    Bowen Chen, Mengyi Zhao, Haomiao Sun, Li Chen, Xu Wang, Kang Du, and Xinglong Wu. Xverse: Consistent multi-subject control of identity and semantic attributes via dit modulation.arXiv preprint arXiv:2506.21416, 2025

    Pith/arXiv arXiv 2025

  5. [5]

    Customcontrast: A multilevel contrastive perspective for subject-driven text-to-image customization, 2025

    Nan Chen, Mengqi Huang, Zhuowei Chen, Yang Zheng, Lei Zhang, and Zhendong Mao. Customcontrast: A multilevel contrastive perspective for subject-driven text-to-image customization, 2025

    work page 2025

  6. [6]

    Directly fine-tuning diffusion models on differentiable rewards

    Kevin Clark, Paul Vicol, Kevin Swersky, and David J Fleet. Directly fine-tuning diffusion models on differentiable rewards. arXiv preprint arXiv:2309.17400, 2023

    Pith/arXiv arXiv 2023

  7. [7]

    Dpok: Reinforcement learning for fine-tuning text-to-image diffusion models

    Ying Fan, Olivia Watkins, Yuqing Du, Hao Liu, Moonkyung Ryu, Craig Boutilier, Pieter Abbeel, Mohammad Ghavamzadeh, Kangwook Lee, and Kimin Lee. Dpok: Reinforcement learning for fine-tuning text-to-image diffusion models. Advances in Neural Information Processing Systems, 36:79858–79885, 2023

    2023

  8. [8]

    An image is worth one word: Personalizing text-to-image generation using textual inversion.arXiv preprint arXiv:2208.01618, 2022

    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 preprint arXiv:2208.01618, 2022

    Pith/arXiv arXiv 2022

  9. [9]

    Denoising diffusion probabilistic models

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840–6851, 2020

    2020

  10. [10]

    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. InThe TenthInternational Conference on Learning Representations, ICLR 2022, Virtual Event, April 25-29, 2022

    work page 2022

  11. [11]

    In-context lora for diffusion transformers.CoRR, abs/2410.23775, 2024

    Lianghua Huang, Wei Wang, Zhi-Fan Wu, Yupeng Shi, Huanzhang Dou, Chen Liang, Yutong Feng, Yu Liu, and Jingren Zhou. In-context lora for diffusion transformers.CoRR, abs/2410.23775, 2024

    Pith/arXiv arXiv 2024

  12. [12]

    RealCustom: Narrowing Real Text Word for Real-Time Open-Domain Text-to-Image Customization

    Mengqi Huang, Zhendong Mao, Mingcong Liu, Qian He, and Yongdong Zhang. Realcustom: Narrowing real text word for real-time open-domain text-to-image customization.CoRR, abs/2403.00483, 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  13. [13]

    Patchdpo: Patch-level DPO for finetuning-free personalized image generation

    Qihan Huang, Long Chan, Jinlong Liu, Wanggui He, Hao Jiang, Mingli Song, and Jie Song. Patchdpo: Patch-level DPO for finetuning-free personalized image generation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2025, Nashville, TN, USA, June 11-15, 2025, pages 18369–18378. Computer Vision Foundation / IEEE, 2025

    work page 2025

  14. [14]

    Resolving multi-condition confusion for finetuning-free personalized image generation

    Qihan Huang, Siming Fu, Jinlong Liu, Hao Jiang, Yipeng Yu, and Jie Song. Resolving multi-condition confusion for finetuning-free personalized image generation. InAAAI-25, Sponsored bythe Association for the Advancement of Artificial Intelligence, February 25 - March 4, 2025, Philadelphia, PA, USA, pages 3707–3714, 2025

    work page 2025

  15. [15]

    D^2it: Dynamic diffusion transformer for accurate image generation

    Weinan Jia, Mengqi Huang, Nan Chen, Lei Zhang, and Zhendong Mao. D^2it: Dynamic diffusion transformer for accurate image generation. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12860–12870, June 2025

    work page 2025

  16. [16]

    Multi-concept customization of text-to-image diffusion

    Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customization of text-to-image diffusion. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1931–1941, 2023

    work page 1931

  17. [17]

    Flux.https://github.com/black-forest-labs/flux, 2024

    Black Forest Labs. Flux.https://github.com/black-forest-labs/flux, 2024

    2024

  18. [18]

    Aligning text-to-image models using human feedback.arXiv preprint arXiv:2302.12192, 2023

    Kimin Lee, Hao Liu, Moonkyung Ryu, Olivia Watkins, Yuqing Du, Craig Boutilier, Pieter Abbeel, Mohammad Ghavamzadeh, and Shixiang Shane Gu. Aligning text-to-image models using human feedback.arXiv preprint arXiv:2302.12192, 2023. 11

    Pith/arXiv arXiv 2023

  19. [19]

    Dongxu Li, Junnan Li, and Steven C. H. Hoi. Blip-diffusion: Pre-trained subject representation for controllable text- to-image generation and editing. InAdvances in Neural Information Processing Systems 36: Annual Conference on Neural Information Processing Systems 2023, NeurIPS 2023, New Orleans, LA, USA, December 10 - 16, 2023, 2023

    work page 2023

  20. [20]

    Grounding dino: Marrying dino with grounded pre-training for open-set object detection.arXiv preprint arXiv:2303.05499, 2023

    Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection.arXiv preprint arXiv:2303.05499, 2023

    Pith/arXiv arXiv 2023

  21. [21]

    Subject-diffusion: Open domain personalized text-to- image generation without test-time fine-tuning

    Jian Ma, Junhao Liang, Chen Chen, and Haonan Lu. Subject-diffusion: Open domain personalized text-to- image generation without test-time fine-tuning. InACM SIGGRAPH 2024 Conference Papers, SIGGRAPH 2024, Denver, CO, USA, 27 July 2024- 1 August 2024, page 25. ACM, 2024

    work page 2024

  22. [22]

    Realcustom++: Representing images as real-word for real-time customization.arXiv preprint arXiv:2408.09744, 2024

    Zhendong Mao, Mengqi Huang, Fei Ding, Mingcong Liu, Qian He, and Yongdong Zhang. Realcustom++: Representing images as real-word for real-time customization.arXiv preprint arXiv:2408.09744, 2024

    arXiv 2024

  23. [23]

    Dreamo: A unified framework for image customization.CoRR, abs/2504.16915, 2025

    Chong Mou, Yanze Wu, Wenxu Wu, Zinan Guo, Pengze Zhang, Yufeng Cheng, Yiming Luo, Fei Ding, Shiwen Zhang, Xinghui Li, Mengtian Li, Songtao Zhao, Jian Zhang, Qian He, and Xinglong Wu. Dreamo: A unified framework for image customization.CoRR, abs/2504.16915, 2025

    arXiv 2025

  24. [24]

    Maxime Oquab, Timothée Darcet, Théo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Mido Assran, Nicolas Ballas, Wojciech Galuba, Russell Howes, Po-Yao Huang, Shang-Wen Li, Ishan Misra, Michael Rabbat, Vasu Sharma, Gabriel Synnaeve, Hu Xu, Hervé Jégou, Julien Mairal, Patrick Lab...

    work page 2024

  25. [25]

    Scalable diffusion models with transformers

    William Peebles and Saining Xie. Scalable diffusion models with transformers. InProceedings of the IEEE/CVF international conference on computer vision, pages 4195–4205, 2023

    2023

  26. [26]

    SDXL: improving latent diffusion models for high-resolution image synthesis

    Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. SDXL: improving latent diffusion models for high-resolution image synthesis. InThe Twelfth International Conference on Learning Representations, ICLR 2024, Vienna, Austria, May 7-11, 2024, 2024

    2024

  27. [27]

    Aligning text-to-image diffusion models with reward backpropagation

    Mihir Prabhudesai, Anirudh Goyal, Deepak Pathak, and Katerina Fragkiadaki. Aligning text-to-image diffusion models with reward backpropagation. 2023

    work page 2023

  28. [28]

    Learning transferable visual models from natural language supervision

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. InProceedings of the 38th International Conference on Machine Learning, ICML 2021, 18-24 July 2021, ...

    work page 2021

  29. [29]

    Direct preference optimization: Your language model is secretly a reward model.Advances in neural information processing systems, 36:53728–53741, 2023

    Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. Direct preference optimization: Your language model is secretly a reward model.Advances in neural information processing systems, 36:53728–53741, 2023

    2023

  30. [30]

    Sam 2: Segment anything in images and videos, 2024

    Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, Eric Mintun, Junting Pan, Kalyan Vasudev Alwala, Nicolas Carion, Chao-Yuan Wu, Ross Girshick, Piotr Dollár, and Christoph Feichtenhofer. Sam 2: Segment anything in images and videos, 2024. URLhttps://arxiv.org/a...

    Pith/arXiv arXiv 2024

  31. [31]

    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. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684–10695, 2022

    2022

  32. [32]

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

    Nataniel Ruiz, Yuanzhen 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

    2023

  33. [33]

    Ominicontrol: Minimal and universal control for diffusion transformer.CoRR, abs/2411.15098, 2024

    Zhenxiong Tan, Songhua Liu, Xingyi Yang, Qiaochu Xue, and Xinchao Wang. Ominicontrol: Minimal and universal control for diffusion transformer.CoRR, abs/2411.15098, 2024

    Pith/arXiv arXiv 2024

  34. [34]

    Diffusion model alignment using direct preference optimization

    Bram Wallace, Meihua Dang, Rafael Rafailov, Linqi Zhou, Aaron Lou, Senthil Purushwalkam, Stefano Ermon, Caiming Xiong, Shafiq Joty, and Nikhil Naik. Diffusion model alignment using direct preference optimization. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA,USA, June 16-22, 2024, pages 8228–8238, 2024. 12

    work page 2024

  35. [35]

    Ms-diffusion: Multi-subject zero-shot image personalization with layout guidance

    Xierui Wang, Siming Fu, Qihan Huang, Wanggui He, and Hao Jiang. Ms-diffusion: Multi-subject zero-shot image personalization with layout guidance. InThe Thirteenth International Conference on Learning Representations, ICLR 2025, Singapore, April 24-28, 2025, 2025

    work page 2025

  36. [36]

    Omnigen2: Exploration to advanced multimodal generation

    Chenyuan Wu, Pengfei Zheng, Ruiran Yan, Shitao Xiao, Xin Luo, Yueze Wang, Wanli Li, Xiyan Jiang, Yexin Liu, Junjie Zhou, Ze Liu, Ziyi Xia, Chaofan Li, Haoge Deng, Jiahao Wang, Kun Luo, Bo Zhang, Defu Lian, Xinlong Wang, Zhongyuan Wang, Tiejun Huang, and Zheng Liu. Omnigen2: Exploration to advanced multimodal generation. CoRR, abs/2506.18871, 2025

    Pith/arXiv arXiv 2025

  37. [37]

    Less-to-more generalization: Unlocking more controllability by in-context generation.arXiv preprint arXiv:2504.02160, 2025

    Shaojin Wu, Mengqi Huang, Wenxu Wu, Yufeng Cheng, Fei Ding, and Qian He. Less-to-more generalization: Unlocking more controllability by in-context generation.arXiv preprint arXiv:2504.02160, 2025

    Pith/arXiv arXiv 2025

  38. [38]

    Omnigen: Unified image generation

    Shitao Xiao, Yueze Wang, Junjie Zhou, Huaying Yuan, Xingrun Xing, Ruiran Yan, Chaofan Li, Shuting Wang, Tiejun Huang, and Zheng Liu. Omnigen: Unified image generation. InIEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2025, Nashville, TN, USA, June 11-15, 2025, pages 13294–13304. Computer Vision Foundation / IEEE, 2025

    work page 2025

  39. [39]

    Omnicontrol: Control any joint at any time for human motion generation.arXiv preprint arXiv:2310.08580, 2023

    Yiming Xie, Varun Jampani, Lei Zhong, Deqing Sun, and Huaizu Jiang. Omnicontrol: Control any joint at any time for human motion generation.arXiv preprint arXiv:2310.08580, 2023

    Pith/arXiv arXiv 2023

  40. [40]

    Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models.CoRR, abs/2308.06721, 2023

    Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models.CoRR, abs/2308.06721, 2023

    Pith/arXiv arXiv 2023

  41. [41]

    Token-level direct preference optimization

    Yongcheng Zeng, Guoqing Liu, Weiyu Ma, Ning Yang, Haifeng Zhang, and Jun Wang. Token-level direct preference optimization. arXiv preprint arXiv:2404.11999, 2024

    Pith/arXiv arXiv 2024

  42. [42]

    Ssr-encoder: Encoding selective subject representation for subject-driven generation

    Yuxuan Zhang, Yiren Song, Jiaming Liu, Rui Wang, Jinpeng Yu, Hao Tang, Huaxia Li, Xu Tang, Yao Hu, Han Pan, and Zhongliang Jing. Ssr-encoder: Encoding selective subject representation for subject-driven generation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA,USA, June 16-22, 2024, pages 8069–8078. IEEE, 2024

    work page 2024

  43. [43]

    Beyond one-preference-for-all: Multi-objective direct preference optimization

    Zhanhui Zhou, Jie Liu, Chao Yang, Jing Shao, Yu Liu, Xiangyu Yue, Wanli Ouyang, and Yu Qiao. Beyond one-preference-for-all: Multi-objective direct preference optimization. 2023. 13 Appendix A Supplementary Ablation Study A.1 Optimization Strategy Performance Analysis To evaluate the effectiveness of our FocusDPO, we conduct ablation studies across two arc...

    work page 2023

  44. [44]

    **Single Subject**: Does each image contain exactly one human subject (no crowds/obstructions)?

    work page

  45. [45]

    **Clothing - Critical Aspects**: Are the styles and patterns of the garments different?

    work page

  46. [46]

    **Hair Accessories - Critical Aspects**: Do the number of hair accessories differ? If the difference in quantity is due to occlusion caused by different character movements, return "Yes"

    work page

  47. [47]

    **Hairstyle, Color**: Are the hairstyle and hair color identical in both images?

    work page

  48. [48]

    **Shoes and Socks**: Are the style and color of shoes, as well as the color and length of socks, identical? If obscured or not visible, return "Y es"

    work page

  49. [49]

    **Iris/Pupils**: Are the pupil colors in both images similar? If obscured or not visible, return "Y es"

    work page

  50. [50]

    **Facial Features**: Are the features of eyes, nose, and mouth basically the same (ignore the influence of facial expression)? If obscured or not visible, return "Yes"

    work page

  51. [51]

    **Clothing**: Are there differences in the garment details including length, cuff style, collar style, and decorative elements such as buttons?

    work page

  52. [52]

    **Hair Accessories**: Are the hair characteristics consistent between both images, including hair length, hair accessory size, position, and style?

    work page

  53. [53]

    Human Identity Consistency Evaluation Criteria Animal Identity Consistency Evaluation Criteria

    **Additional Accessories - Critical Aspects**: Do the presence or absence of accessories differ? Items held in hands (such as swords, knives, and other props) are not within the scope of evaluation. Human Identity Consistency Evaluation Criteria Animal Identity Consistency Evaluation Criteria

    work page

  54. [54]

    ** Single Subject ** : Does each image contain exactly one animal subject (no multiple animals/obstructions)?

    work page

  55. [55]

    ** Species and Breed ** : Are the species and breed characteristics identical in both images?

    work page

  56. [56]

    ** Fur/Coat Pattern ** : Are the fur patterns, markings, and colorations consistent between both images?

    work page

  57. [57]

    ** Body Proportions ** : Are the body size, limb proportions, and overall physical structure similar? If partially obscured, return "Yes"

    work page

  58. [58]

    ** Facial Features ** : Are the eye shape, nose/snout characteristics, and ear structure basically the same? If obscured or not visible, return "Y es"

    work page

  59. [59]

    ** Eye Color ** : Are the eye colors similar in both images? If obscured or not visible, return "Y es"

    work page

  60. [60]

    ** Distinctive Markings ** : Are unique identifying marks, spots, stripes, or patches consistent between images?

    work page

  61. [61]

    ** Tail Characteristics ** : Are the tail length, shape, and fur texture identical? If obscured or not visible, return "Y es"

    work page

  62. [62]

    ** Accessories - Critical Aspects ** : Do the presence or absence of collars, tags, or other worn accessories differ?

    work page

  63. [63]

    ** Overall Appearance** : Are there significant differences in the animal's general appearance that would affect individual identification? Minor pose variations and natural movement should return "Y es"

    work page

  64. [64]

    ** Single Object ** : Does each image contain exactly one target object (no multiple objects/obstructions)?

    work page

  65. [65]

    ** Object Type and Category ** : Are the object type and category characteristics identical in both images?

    work page

  66. [66]

    ** Color and Material ** : Are the colors, textures, and material properties consistent between both images?

    work page

  67. [67]

    ** Size and Proportions ** : Are the object dimensions, scale, and proportional relationships similar? If partially obscured, return "Yes"

    work page

  68. [68]

    ** Shape and Structure ** : Are the overall form, geometric properties, and structural elements basically the same? If obscured or not visible, return "Yes"

    work page

  69. [69]

    ** Surface Details ** : Are surface patterns, textures, and finish characteristics similar in both images? If obscured or not visible, return "Y es"

    work page

  70. [70]

    ** Distinctive Features ** : Are unique identifying elements, markings, labels, or design details consistent between images?

    work page

  71. [71]

    ** Component Parts ** : Are the visible components, attachments, and sub-elements identical? If obscured or not visible, return "Y es"

    work page

  72. [72]

    ** Functional Elements ** - Critical Aspects: Do the presence or absence of buttons, handles, switches, or other functional components differ?

    work page

  73. [73]

    Object Identity Consistency Evaluation Criteria Figure 13 Identity Consistency Evaluation Criteria of human, animal and object

    ** Overall Appearance ** : Are there significant differences in the object's general appearance that would affect individual identification? Minor angle variations and lighting changes should return "Yes". Object Identity Consistency Evaluation Criteria Figure 13 Identity Consistency Evaluation Criteria of human, animal and object. 19

    work page

This paper was first reviewed by deepseek-v4-flash on August 5, 2026.