Pith. sign in

REVIEW 4 major objections 5 minor 7 cited by

UMO claims that treating multi-identity generation as a global assignment problem, with a matching reward and Hungarian-optimal pairing, significantly boosts identity fidelity and reduces identity confusion on existing customization models.

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 ↗

A reinforcement-learning reward based on bipartite face matching improves multi-identity consistency and reduces identity confusion in image customization models.

T0 review reviewed 2026-08-04 challenge →

load-bearing objection A genuinely neat multi-to-multi matching reward for identity customization that probably works, but the headline metrics are too tightly coupled to the training objective to take at face value. the 4 major comments →

arxiv 2509.06818 v1 pith:6X7MWMIU submitted 2025-09-08 cs.CV cs.LG

UMO: Scaling Multi-Identity Consistency for Image Customization via Matching Reward

Yufeng Cheng , Wenxu Wu , Shaojin Wu , Mengqi Huang , Fei Ding , Qian He This is my paper
classification cs.CV cs.LG
keywords multi-identity customizationidentity consistencyidentity confusionreinforcement learningdiffusion modelsassignment problemreference rewardID-Conf metric
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

This paper argues that existing multi-identity image customization methods fail because they pair each reference identity with one generated face in a one-to-one way, which cannot separate natural variation within an identity from differences between identities as the number of people grows. UMO instead reformulates multi-identity generation as a global assignment problem: it matches every generated face to the most appropriate reference identity, maximizing overall identity similarity and minimizing confusion. The method works as a general post-training scheme, adding a reinforcement-learning reward called Multi-Identity Matching Reward (MIMR) on top of existing customization models such as UNO and OmniGen2, using face-recognition embeddings and the Hungarian algorithm. The paper also introduces a scalable multi-reference dataset and a new metric, ID-Conf, that measures identity confusion as the relative margin between the two most similar generated candidates. If correct, UMO offers a model-agnostic way to preserve several distinct identities in a single generated image and to scale that ability beyond what one-to-one mapping achieves.

Core claim

The central claim is that identity confusion in multi-person image generation is an assignment problem, not just a data or architecture problem. UMO's training reward constructs a bipartite graph between M reference identities and N detected faces in the generated image, weights each edge by the cosine similarity of face embeddings, and finds the optimal matching with the Hungarian algorithm. The Multi-Identity Matching Reward then rewards matched pairs positively and non-matched pairs negatively, so the model is pushed to make every generated face look like the most appropriate reference and unlike all others. The paper reports that this training scheme significantly raises ID-Sim and ID-Co

What carries the argument

The key machinery is the Multi-Identity Matching Reward (MIMR), built on a Single Identity Reward (SIR), which is the cosine similarity between face embeddings of the generated image and a reference. For multi-reference inputs, faces are treated as vertices of a bipartite graph; edges carry SIR weights; the Hungarian algorithm finds the optimal assignment; and MIMR sums the matched-pair similarities positively and the unmatched-pair similarities negatively. This global-matching reward, optimized through ReReFL, a diffusion-model reinforcement-learning loop that backpropagates through a partially denoised image, is what lets the model separate intra-identity variation from inter-identity diff

Load-bearing premise

The load-bearing premise is that the face-recognition encoder used both as the training reward and as part of the evaluation metric really measures identity similarity on stylized, photorealistic, and multi-person compositions; if that encoder is biased by style or layout, then both the training signal and the reported ID-Sim and ID-Conf improvements are miscalibrated.

What would settle it

Swap the evaluation face encoder for a different face-recognition model and run a forced-choice human identity test comparing UMO with its baselines on stylized and multi-person outputs; if human agreement drops or the reported gains vanish, the ID-Sim and ID-Conf improvements are partly an artifact of the chosen embedding model.

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

If this is right

  • UMO can be applied on top of existing customization models like UNO and OmniGen2 with only LoRA fine-tuning, so it does not require architectural changes to improve identity consistency.
  • Multi-identity generation can scale to more identities because the assignment formulation removes the one-to-one mapping bottleneck that causes confusion as the number of people grows.
  • The ID-Conf metric provides a quantitative way to measure identity confusion, not just identity similarity, which was missing from prior evaluations.
  • MIMR's combination of positive reward for matched pairs and negative reward for unmatched pairs directly suppresses the blended-face and swapped-attribute failures common in multi-identity generation.
  • The released dataset, code, and model make it possible to reproduce UMO and apply the same reward-based matching recipe to other customization models.

Where Pith is reading between the lines

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

  • Editorial inference: UMO's gains hinge on the face-recognition encoder measuring true identity across styles and compositions; if that encoder is biased by art style, the reward could overfit to the encoder's preferences rather than to human-perceived identity.
  • Editorial inference: The same assignment-based matching reward could extend beyond faces to any multi-instance subject (objects, animals, characters) by replacing face embeddings with a general instance-similarity metric.
  • Editorial inference: The data recipe, real multi-identity video frames for fidelity plus filtered synthetic frames for diversity, could be reused for other customization tasks that need many instances in one image.
  • Editorial inference: The ID-Conf margin metric could generalize to any multi-instance generation task where confusion is defined as the gap between the best and the second-best candidate match.
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

4 major / 5 minor

Summary. The paper proposes UMO, a reinforcement-learning framework for improving multi-identity consistency in image customization models. UMO defines a single-identity reward based on cosine similarity between face-recognition embeddings, extends it to multiple identities through a Hungarian-assignment 'multi-to-multi matching' reward (MIMR), and fine-tunes diffusion-based customization models with a partial-denoising RL objective (ReReFL). The authors also construct a large multi-identity dataset (synthesized and real) and propose a new metric, ID-Conf, to quantify identity confusion. Experiments on XVerseBench and OmniContext report large improvements in ID-Sim and ID-Conf when UMO is applied to UNO and OmniGen2, with ablations showing that both ReReFL and MIMR contribute. Code and models are released, though the training dataset is not.

Significance. If the reported gains are genuine, UMO would be a valuable and general post-training method for improving identity fidelity and reducing identity confusion across existing customization models, which currently suffer from these problems. The paper is clearly written, the method is well specified, and the ablations support the contribution of each component. However, the central empirical claim is currently not fully established: the MIMR reward and the proposed ID-Conf metric are essentially the same margin objective, and the paper does not demonstrate that the gains transfer to an independent face encoder. The absence of error bars and significance tests, and the lack of release of the training data, further weaken the quantitative evidence. For these reasons, the significance is real but conditional on addressing the evaluation concerns.

major comments (4)
  1. [Eq. (4), Eq. (5), Section 4.1.3] The reward–metric circularity concern is valid. MIMR with λ1=1, λ2=−1 rewards increasing the cosine similarity between each generated face and its assigned reference while penalizing similarities to non-assigned references, i.e., it directly maximizes the margin between top-1 and top-2 similarities. ID-Conf, as defined in Eq. (5), measures the very same margin: 1 − cos(top2)/cos(top1). Since ψ in Eq. (4) and Ψ in Eq. (5) are both face-recognition embeddings and the relation between them is not specified, the reported ID-Conf improvements in Tables 2–5 may be substantially by construction. The paper must specify whether ψ and Ψ are the same network, and more importantly, evaluate on an independent face encoder not used in training, and validate ID-Conf against human confusion judgments. Without this, the claim of mitigating identity confusion is not convincingly supported.
  2. [Tables 1–5, Section 4.4] All quantitative results are reported as point estimates without error bars, confidence intervals, or significance tests. For example, Table 2 shows ID-Sim for UNO jumping from 31.82 to 69.09 after UMO training; such large gains are plausible but RL training is stochastic and model selection over checkpoints can inflate results. The paper should report multiple training seeds or at least bootstrap confidence intervals over benchmark items, and state the number of generations used for evaluation. The user study (Section 4.4, Figure 7) is presented only as radar charts, without the number of participants, questionnaire details, or any statistical test. Independent human evidence is important here, especially because it is the only signal not tied to the training reward.
  3. [Eq. (3), Algorithm 1, Section 4.1.2] The assignment formulation requires the number of detected faces N to be at least the number of reference identities M (Eq. (3) sums over n references). The paper does not discuss what happens when N < M, e.g., when the generation misses an identity or the face detector fails. This is directly relevant to the confusion-alleviation claim: if the face detector provides an unfair advantage by ignoring missed identities, ID-Conf could be inflated. The authors should report face-detection statistics (detection rate, N vs M distribution) and describe the failure handling in the Hungarian matching.
  4. [Section 3.2, Section 4.1.1] The paper states that the training dataset is a key contribution ('we develop a scalable customization dataset'), but only code and model weights are released, not the dataset. Since the reward is trained on this data, the reproducibility of the central results is limited. The dataset construction details (e.g., the 'strict face similarity filtering' thresholds, the real/synthetic ratio, and identity-count distribution) should be documented more fully. If the dataset cannot be released, a detailed specification and a sample should be provided.
minor comments (5)
  1. [Algorithm 1, line 12] Line 12 reads 'θ_{i+1} ← θ_i', which is presumably a typo for the parameter update. Please correct.
  2. [Eq. (3)] The notation uses both n and M for the number of reference identities; unify these for clarity.
  3. [Section 4.1.3] The footnote 'We align ID-Conf score with the value range of the other metrics' is unclear. Please specify the exact transformation applied to ID-Conf in Tables 2, 4, 5, and Appendix F.
  4. [Figure 7] Radar charts show mean ranks, but the number of participants and the distribution of ranks are not given. Even a sentence with N and a p-value would help.
  5. [Throughout] Minor language issues: 'wildely' (Section 2.2), 'fails to address' (Section 1), 'reformulating' (Section 5), and similar typos should be corrected.

Circularity Check

1 steps flagged

MIMR reward and ID-Conf metric are the same relative-margin objective, so the reported confusion-mitigation gain is partly by construction; external ID-Sim and user study provide partial independent support.

specific steps
  1. fitted input called prediction [Section 3.3.3 Eq. (4); Section 4.1.1 hyperparameters; Section 4.1.3 Eq. (5)]
    "we define Multi-Identity Matching Reward (MIMR) as: RMIMR = 1/(MN) Σ_j Σ_k (λ1 1{k=σ(j)} + λ2 1{k≠σ(j)}) e_{F_j,\hat{F}_k} where λ1 > 0, λ2 < 0 ... For the training hyperparameters ... λ1 = 1, λ2 = −1 ... ID-Conf = 1/n Σ_i clip(1 − cos(Ψ(F_i), Ψ(\hat{F}_{j[2]_i})) / cos(Ψ(F_i), Ψ(\hat{F}_{j[1]_i})), 0, 1) where Ψ is the model used in XVerseBench to get face embeddings."

    With λ1=1 and λ2=−1, MIMR increases the cosine of each reference to its assigned generated face and decreases the cosine to every non-assigned face, in particular the second-best candidate. ID-Conf is exactly the per-reference relative margin 1 − (second-best cosine / best cosine), and the assignment σ in Eq. (3) is chosen so the paired face is the best-matching one. Thus maximizing MIMR directly maximizes the quantity that ID-Conf measures; the reported ID-Conf improvements in Tables 2–3 are the training objective itself, not an independent validation of confusion mitigation. Because the paper proposes both the metric and the reward, the confusion-mitigation claim is partially circular. The use of different encoder symbols ψ (reward) and Ψ (evaluation) does not remove the construction, si

full rationale

The paper's central empirical claim has two components: improving identity similarity and mitigating identity confusion. The similarity claim is supported by external ID-Sim scores on XVerseBench and by a user study, so it is not circular. The confusion-mitigation claim, however, is evaluated primarily with ID-Conf, a metric newly introduced in the same paper. The Multi-Identity Matching Reward (Eq. 4) is written as a contrastive objective over reference/generated cosine similarities: it rewards the assigned pair and penalizes all non-assigned pairs. This is the loss-side version of the relative top-1/top-2 margin that ID-Conf (Eq. 5) reports. Optimizing with RL therefore forces the ID-Conf score upward by construction, making the 'mitigates confusion' quantitative result partly self-referential. Qualitative examples and a user-study radar chart (Section 4.4) provide some independent evidence, but the radar chart lacks participant counts and significance testing, so they do not remove the circularity. I found no load-bearing self-citation chain: UNO [32] is authored by the same group, but UMO is also applied to OmniGen2 and compared against external methods, so the self-citations are not the sole support. Overall, one central reported metric reduces to the training objective, warranting a score of 6 rather than a higher score.

Axiom & Free-Parameter Ledger

3 free parameters · 4 axioms · 0 invented entities

The paper introduces no new physical or conceptual entities; it proposes a new training objective and a new metric. Its central claim rests on the reliability of face-recognition embeddings, the validity of early-step denoising gradients, and the correctness of assignment during training.

free parameters (3)
  • lambda_1 = 1
    Weight for matched-pair similarity in MIMR (Eq. 4), set by hand.
  • lambda_2 = -1
    Weight for mismatched-pair similarity in MIMR (Eq. 4), set by hand; negative sign encodes penalty.
  • TS/Te denoising step range = [1,10] for UNO, [1,20] for OmniGen2
    Chosen based on observed SIR stability in Figure 4, a data-dependent choice that affects gradient quality and training success.
axioms (4)
  • domain assumption Face recognition embedding psi (reward, Eq. 2) and evaluation encoder Psi (ID-Conf, Eq. 5) faithfully measure identity similarity for photorealistic and stylized faces.
    Both the training reward and the evaluation metric depend entirely on these embeddings; no robustness analysis for stylized or anime faces is provided.
  • domain assumption Gradients through the early denoising steps (TS to Te) are sufficient to improve final output identity fidelity.
    ReReFL backpropagates only through steps in [TS, Te]; Figure 4 shows reward stability but does not prove the gradients are useful for the final image.
  • domain assumption The optimal assignment from Hungarian matching on intermediate noisy outputs correctly identifies reference-to-generated correspondences.
    If the assignment is wrong, the MIMR reward reinforces incorrect pairings and can degrade identity consistency; used in Eq. 3.
  • domain assumption Training data from video frames and synthetic pipelines has the same distribution as the evaluation benchmarks.
    Section 3.2 describes the pipeline but provides no distribution-shift analysis between the training data and XVerseBench or OmniContext.

reviewed 2026-08-04 · how reviews work

0 comments
Cite this review

Pith. "Pith review of UMO: Scaling Multi-Identity Consistency for Image Customization via Matching Reward." pith.science (2026). https://pith.science/paper/6X7MWMIU

@misc{pith2026250906818,
  author       = {Pith},
  title        = {Pith review of: UMO: Scaling Multi-Identity Consistency for Image Customization via Matching Reward},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6X7MWMIU}},
  note         = {Machine review of arXiv:2509.06818}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Recent advancements in image customization exhibit a wide range of application prospects due to stronger customization capabilities. However, since we humans are more sensitive to faces, a significant challenge remains in preserving consistent identity while avoiding identity confusion with multi-reference images, limiting the identity scalability of customization models. To address this, we present UMO, a Unified Multi-identity Optimization framework, designed to maintain high-fidelity identity preservation and alleviate identity confusion with scalability. With "multi-to-multi matching" paradigm, UMO reformulates multi-identity generation as a global assignment optimization problem and unleashes multi-identity consistency for existing image customization methods generally through reinforcement learning on diffusion models. To facilitate the training of UMO, we develop a scalable customization dataset with multi-reference images, consisting of both synthesised and real parts. Additionally, we propose a new metric to measure identity confusion. Extensive experiments demonstrate that UMO not only improves identity consistency significantly, but also reduces identity confusion on several image customization methods, setting a new state-of-the-art among open-source methods along the dimension of identity preserving. Code and model: https://github.com/bytedance/UMO

discussion (0)

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

Forward citations

Cited by 7 Pith papers

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

  1. UniCustom: Unified Visual Conditioning for Multi-Reference Image Generation

    cs.CV 2026-05 unverdicted novelty 7.0

    UniCustom fuses ViT and VAE features before VLM encoding and uses two-stage training plus slot-wise regularization to improve subject consistency in multi-reference diffusion-based image generation.

  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.

  3. Lance: Unified Multimodal Modeling by Multi-Task Synergy

    cs.CV 2026-05 unverdicted novelty 6.0

    Lance presents a dual-stream mixture-of-experts model with modality-aware positional encoding and staged multi-task training that outperforms prior open-source unified models on image and video generation while keepin...

  4. UniCustom: Unified Visual Conditioning for Multi-Reference Image Generation

    cs.CV 2026-05 unverdicted novelty 6.0

    A unified visual conditioning approach fuses semantic and appearance features before VLM processing, with two-stage training and slot-wise regularization, to improve consistency in multi-reference image generation.

  5. Hierarchical Concept-to-Appearance Guidance for Multi-Subject Image Generation

    cs.CV 2026-02 conditional novelty 6.0

    A diffusion-transformer framework with VLM-grounded masked attention and VAE dropout improves identity and prompt fidelity for multi-subject image generation.

  6. iMontage: Unified, Versatile, Highly Dynamic Many-to-many Image Generation

    cs.CV 2025-11 conditional novelty 6.0

    iMontage repurposes a pretrained video diffusion model to generate coherent yet highly dynamic image sets from arbitrary numbers of input images.

  7. Lance: Unified Multimodal Modeling by Multi-Task Synergy

    cs.CV 2026-05 unverdicted novelty 5.0

    Lance introduces a dual-stream MoE model with modality-aware rotary positional encoding and staged multi-task training that outperforms open-source unified models on image and video generation while retaining understa...

Reference graph

Works this paper leans on

41 extracted references · 7 canonical work pages · cited by 5 Pith papers

  1. [1]

    Improving image generation with better captions.Computer Science

    James Betker, Gabriel Goh, Li Jing, Tim Brooks, Jianfeng Wang, Linjie Li, Long Ouyang, Juntang Zhuang, Joyce Lee, Yufei Guo, et al. Improving image generation with better captions.Computer Science. https://cdn. openai. com/papers/dall-e-3. pdf, 2(3):8, 2023

    2023

  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. In The Twelfth International Conference on Learning Representations, 2024. URL https://openreview.net/forum?id=YCWjhGrJFD

    2024

  3. [3]

    End-to-end object detection with transformers

    Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213–229. Springer, 2020

    work page 2020

  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]

    Unireal: Universal image generation and editing via learning real-world dynamics.arXiv preprint arXiv:2412.07774, 2024

    Xi Chen, Zhifei Zhang, He Zhang, Yuqian Zhou, Soo Ye Kim, Qing Liu, Yijun Li, Jianming Zhang, Nanxuan Zhao, Yilin Wang, Hui Ding, Zhe Lin, and Hengshuang. Unireal: Universal image generation and editing via learning real-world dynamics.arXiv preprint arXiv:2412.07774, 2024

    Pith/arXiv arXiv 2024

  6. [6]

    Emerging properties in unified multimodal pretraining.arXiv preprint arXiv:2505.14683, 2025

    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

    Pith/arXiv arXiv 2025

  7. [7]

    Scaling rectified flow transformers for high-resolution image synthesis

    Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In Forty-firstinternational conference on machine learning, 2024

    2024

  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]

    Pulid: Pure and lightning id customization via contrastive alignment.Advances in neural information processing systems, 37:36777–36804, 2024

    Zinan Guo, Yanze Wu, Chen Zhuowei, Peng Zhang, Qian He, et al. Pulid: Pure and lightning id customization via contrastive alignment.Advances in neural information processing systems, 37:36777–36804, 2024

    2024

  10. [10]

    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

  11. [11]

    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

    2022

  12. [12]

    In-context lora for diffusion transformers.arXiv preprint arXiv: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.arXiv preprint arXiv:2410.23775, 2024

    Pith/arXiv arXiv 2024

  13. [13]

    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. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7476–7485, 2024

    2024

  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. InProceedings of the AAAI Conference on Artificial Intelligence, volume 39, pages 3707–3714, 2025

    work page 2025

  15. [15]

    The hungarian method for the assignment problem.Naval research logistics quarterly, 2(1-2): 83–97, 1955

    Harold W Kuhn. The hungarian method for the assignment problem.Naval research logistics quarterly, 2(1-2): 83–97, 1955

    work page 1955

  16. [16]

    Flux: Official inference repository for flux.1 models, 2024

    Black Forest Labs. Flux: Official inference repository for flux.1 models, 2024. URL https://github.com/ black-forest-labs/flux. Accessed: 2025-02-07

    work page 2024

  17. [17]

    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

  18. [18]

    Dreamo: A unified framework for image customization.arXiv preprint arXiv:2504.16915, 2025

    Chong Mou, Yanze Wu, Wenxu Wu, Zinan Guo, Pengze Zhang, Yufeng Cheng, Yiming Luo, Fei Ding, Shiwen Zhang, Xinghui Li, et al. Dreamo: A unified framework for image customization.arXiv preprint arXiv:2504.16915, 2025. 13

    arXiv 2025

  19. [19]

    Introducing gpt-4.1 in the api, 2025

    OpenAI. Introducing gpt-4.1 in the api, 2025. URLhttps://openai.com/index/gpt-4-1. Accessed: 2025-04-14

    work page 2025

  20. [20]

    Training language models to follow instructions with human feedback

    Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. Advances in neural information processing systems, 35:27730–27744, 2022

    2022

  21. [21]

    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

  22. [22]

    Sdxl: Improving latent diffusion models for high-resolution image synthesis.arXiv preprint arXiv:2307.01952, 2023

    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.arXiv preprint arXiv:2307.01952, 2023

    Pith/arXiv arXiv 2023

  23. [23]

    Movie gen: A cast of media foundation models.arXiv preprint arXiv:2410.13720, 2024

    Adam Polyak, Amit Zohar, Andrew Brown, Andros Tjandra, Animesh Sinha, Ann Lee, Apoorv Vyas, Bowen Shi, Chih-Yao Ma, Ching-Yao Chuang, et al. Movie gen: A cast of media foundation models.arXiv preprint arXiv:2410.13720, 2024

    Pith/arXiv arXiv 2024

  24. [24]

    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

  25. [25]

    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

  26. [26]

    Photorealistic text-to-image diffusion models with deep language understanding.Advances in neural information processing systems, 35:36479–36494, 2022

    Chitwan Saharia, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding.Advances in neural information processing systems, 35:36479–36494, 2022

    2022

  27. [27]

    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 Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8228–8238, 2024

    2024

  28. [28]

    Instantid: Zero-shot identity-preserving generation in seconds.arXiv preprint arXiv:2401.07519, 2024

    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

    Pith/arXiv arXiv 2024

  29. [29]

    Ms-diffusion: Multi-subject zero-shot image personalization with layout guidance.arXiv preprint arXiv:2406.07209, 2024

    Xierui Wang, Siming Fu, Qihan Huang, Wanggui He, and Hao Jiang. Ms-diffusion: Multi-subject zero-shot image personalization with layout guidance.arXiv preprint arXiv:2406.07209, 2024

    Pith/arXiv arXiv 2024

  30. [30]

    Elite: Encoding visual concepts into textual embeddings for customized text-to-image generation

    Yuxiang Wei, Yabo Zhang, Zhilong Ji, Jinfeng Bai, Lei Zhang, and Wangmeng Zuo. Elite: Encoding visual concepts into textual embeddings for customized text-to-image generation. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 15943–15953, 2023

    2023

  31. [31]

    Omnigen2: Exploration to advanced multimodal generation.arXiv preprint arXiv:2506.18871, 2025

    Chenyuan Wu, Pengfei Zheng, Ruiran Yan, Shitao Xiao, Xin Luo, Yueze Wang, Wanli Li, Xiyan Jiang, Yexin Liu, Junjie Zhou, et al. Omnigen2: Exploration to advanced multimodal generation.arXiv preprint arXiv:2506.18871, 2025

    Pith/arXiv arXiv 2025

  32. [32]

    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

  33. [33]

    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. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 13294–13304, 2025

    2025

  34. [34]

    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

  35. [35]

    Imagereward: Learning and evaluating human preferences for text-to-image generation

    Jiazheng Xu, Xiao Liu, Yuchen Wu, Yuxuan Tong, Qinkai Li, Ming Ding, Jie Tang, and Yuxiao Dong. Imagereward: Learning and evaluating human preferences for text-to-image generation. Advances in Neural Information Processing Systems, 36:15903–15935, 2023

    2023

  36. [36]

    Dancegrpo: Unleashing grpo on visual generation.arXiv preprint arXiv:2505.07818, 2025

    Zeyue Xue, Jie Wu, Yu Gao, Fangyuan Kong, Lingting Zhu, Mengzhao Chen, Zhiheng Liu, Wei Liu, Qiushan Guo, Weilin Huang, et al. Dancegrpo: Unleashing grpo on visual generation.arXiv preprint arXiv:2505.07818, 2025. 14

    Pith/arXiv arXiv 2025

  37. [37]

    Gpt-imgeval: A comprehensive benchmark for diagnosing gpt4o in image generation.arXiv preprint arXiv:2504.02782, 2025

    Zhiyuan Yan, Junyan Ye, Weijia Li, Zilong Huang, Shenghai Yuan, Xiangyang He, Kaiqing Lin, Jun He, Conghui He, and Li Yuan. Gpt-imgeval: A comprehensive benchmark for diagnosing gpt4o in image generation.arXiv preprint arXiv:2504.02782, 2025

    Pith/arXiv arXiv 2025

  38. [38]

    Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models.arXiv preprint arXiv: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.arXiv preprint arXiv:2308.06721, 2023

    Pith/arXiv arXiv 2023

  39. [39]

    Openstory: A large-scale open-domain dataset for subject-driven visual storytelling

    Zilyu Ye, Jinxiu Liu, JinJin Cao, Zhiyang Chen, Ziwei Xuan, Mingyuan Zhou, Qi Liu, and Guo-Jun Qi. Openstory: A large-scale open-domain dataset for subject-driven visual storytelling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7953–7962, 2024

    work page 2024

  40. [40]

    Does reinforcement learning really incentivize reasoning capacity in llms beyond the base model?arXiv preprint arXiv:2504.13837, 2025

    Yang Yue, Zhiqi Chen, Rui Lu, Andrew Zhao, Zhaokai Wang, Shiji Song, and Gao Huang. Does reinforcement learning really incentivize reasoning capacity in llms beyond the base model?arXiv preprint arXiv:2504.13837, 2025

    Pith/arXiv arXiv 2025

  41. [41]

    UXO Team

    Yifu Zhang, Peize Sun, Yi Jiang, Dongdong Yu, Fucheng Weng, Zehuan Yuan, Ping Luo, Wenyu Liu, and Xinggang Wang. Bytetrack: Multi-object tracking by associating every detection box. InEuropean conference on computer vision, pages 1–21. Springer, 2022. 15 UMO: Scaling Multi-Identity Consistency for Image Customization via Matching Reward Appendix F Detaile...

    work page 2022

This paper was first reviewed by deepseek-v4-flash on August 4, 2026.