pith. sign in

arxiv: 2606.26947 · v1 · pith:YXKTIKVJnew · submitted 2026-06-25 · 💻 cs.CV · cs.AI

Scaling Multi-Reference Image Generation with Dynamic Reward Optimization

Wenwang Huang , Yusen Fu , Junjie Wang , Mengfei Huang , Yulin Li , Gan Liu , Jing Cai , Yancheng He
show 1 more author
Zhuotao Tian
This is my paper

Pith reviewed 2026-06-26 05:04 UTC · model grok-4.3

classification 💻 cs.CV cs.AI
keywords multi-reference image generationOmniRef-BenchDyRefDifficulty-aware Advantage ReweightingDiscriminative Reward Scalingpersonalized image generationreward optimizationdiffusion models
0
0 comments X

The pith

A two-stage framework with dynamic reward techniques improves open-source models on complex multi-reference image generation.

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

The paper establishes a benchmark called OmniRef-Bench for testing multi-reference image generation with complex combinations of reference images. It proposes DyRef, which uses supervised fine-tuning in the first stage and then Difficulty-aware Advantage Reweighting along with Discriminative Reward Scaling in the second stage to handle large numbers of mixed-type references. This matters because existing models struggle as the number of references grows, and the method aims to scale performance accordingly. The approach shows gains on the new benchmark and on single-image editing tasks.

Core claim

DyRef is a two-stage training framework where supervised fine-tuning provides basic MRIG capability and a second stage with Difficulty-aware Advantage Reweighting (DAR) and Discriminative Reward Scaling (DRS) dynamically adjusts optimization to improve handling of many mixed-type reference images, leading to better performance on OmniRef-Bench and single-image editing benchmarks.

What carries the argument

DyRef two-stage framework with Difficulty-aware Advantage Reweighting (DAR) that adjusts the optimization objective and Discriminative Reward Scaling (DRS) that enlarges intra-group reward differences.

If this is right

  • Mainstream open-source models can better handle complex MRIG scenarios with increasing numbers of mixed-type reference images.
  • The framework generalizes to improve performance on single-image editing benchmarks.
  • DAR enables dynamic adjustment for large numbers of references.
  • DRS supports more effective policy optimization through larger reward differences.

Where Pith is reading between the lines

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

  • Applying similar dynamic reward methods could benefit other multi-input generation tasks in vision.
  • The introduction of OmniRef-Bench may drive development of models that maintain performance with more references.
  • Testing on even larger reference sets could reveal scalability limits of the approach.

Load-bearing premise

The performance improvements are caused by the DAR and DRS techniques rather than other unmentioned changes in training or evaluation.

What would settle it

An experiment that applies the same first-stage fine-tuning but omits DAR and DRS, then measures if performance on OmniRef-Bench still improves compared to the base model.

Figures

Figures reproduced from arXiv: 2606.26947 by Gan Liu, Jing Cai, Junjie Wang, Mengfei Huang, Wenwang Huang, Yancheng He, Yulin Li, Yusen Fu, Zhuotao Tian.

Figure 1
Figure 1. Figure 1: Versatile samples of the proposed DyRef. Given complex samples that involve mixed reference image types and varying numbers of reference images, DyRef consistently generates high-quality results in accordance with user instructions. The Proposed OmniRef-Bench. To better assess MRIG performance, we introduce OmniRef-Bench, a personalized image generation benchmark comprising intricate combinations of divers… view at source ↗
Figure 2
Figure 2. Figure 2: Performance of DyRef. On both MRIG and single-image editing tasks, DyRef (Ours) consistently outperforms open-source state-of-the-art methods and achieves performance comparable to leading closed-source model Nano Banana Pro. Our Solution. To address this issue, we propose DyRef, a two-stage training framework. In the first stage, we use Supervised Fine-Tuning (SFT) to equip the model with preliminary capa… view at source ↗
Figure 3
Figure 3. Figure 3: Motivation of our DyRef. (a) Qualitative results: open source models yield high-quality results with a limited set of references, but suffer from significant artifacts and semantic loss as the number and complexity of reference images increase. (b) Quantitative results: on the OmniRef-Bench, CLIP-I scores between generated images and target images confirm that the performance of mainstream open-source mode… view at source ↗
Figure 4
Figure 4. Figure 4: Data construction pipeline. Our data construction pipeline is structured into two phases. (a) We obtain the target image using T2I generation. (b) We acquire the corresponding reference images of each type (except style) based on the target image, and we obtain the stylized target image by updating the original target image using an external style reference image from the OmniConsistency dataset [32]. 3. M… view at source ↗
Figure 5
Figure 5. Figure 5: Framework of DyRef. In Stage I, SFT equips the model with the basic capability to handle complex MRIG tasks. In Stage II, (a) DRS enlarges the reward differences across samples for better training. (b) DAR enhances the model’s focus on samples with a large number of mixed-type reference images. objective is defined as: min v ∫︁ 1 0 E [︁ ∥(z1 − z0) − vθ (zt , t)∥ 2 ]︁ dt. (1) 3.2. Stage II: Difficulty-aware… view at source ↗
Figure 6
Figure 6. Figure 6: Qualitative comparison with different models or methods on OmniRef-Bench. Each reference image is framed with a distinct color. Its corresponding phrase in the text prompt is also displayed in the same color. substantially improves performance on complex MRIG and single-image editing tasks, demonstrating the effectiveness and generalization capability of our approach. 13 [PITH_FULL_IMAGE:figures/full_fig_… view at source ↗
Figure 7
Figure 7. Figure 7: Overview of OmniRef-Bench. (a) Some illustrative examples from OmniRef-Bench. (b) Sunburst chart showing the dataset distribution. Subj, Bg, and Light are the abbreviations for Subject, Background, and Lighting, respectively. The inner rim categorizes samples by reference type combinations. The outer rim further partitions these categories by the number of reference images. Notably, while the "Subj" task s… view at source ↗
Figure 8
Figure 8. Figure 8: The prompts for reference image (or target image) generation of each type. The placeholders <subject> and <pose> are replaced with specific entity names and posture descriptions during generation. Note that pose-related instruction in subject reference generation is exclusively applied when the subject is identified as a human. T2I Generation for Target Image. Inspired by UNO [26], we first obtain raw subj… view at source ↗
Figure 9
Figure 9. Figure 9: Examples of diverse reference pointers for background, pose, lighting, and style. The placeholder "X" denotes the index of the reference image. We employ multiple linguistic variants for each reference type to enhance prompt diversity and robustness. 23 [PITH_FULL_IMAGE:figures/full_fig_p023_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Example of superficial alignment captured by objective metrics. The lighting objective metric erroneously assigns a high score due to content leakage from the reference image, while the MLLM evaluator identifies this flaw. Lighting Reference. Using the target images as reference, we leverage Seedream4.5 to generate new images that match the illumination effects of the target images. The generated images a… view at source ↗
Figure 11
Figure 11. Figure 11: Example case of the inconsistent scoring and lack of fine-grained spatial perception of MLLM-based evaluation. While the objective metric yields consistent scores for structurally similar poses, the MLLM evaluator exhibits significant variance (scoring 9 vs. 5) and provides inaccurate reasoning on the fine-grained pose alignment. Background reference. We use Gemini 3 Flash to check the image quality of th… view at source ↗
Figure 12
Figure 12. Figure 12: Effect of γ and t on OmniRef-Bench. MLLM Evaluation Average and Objective Metrics Average denote the mean scores computed across all evaluation dimensions of MLLM evaluation and objective metrics in OmniRef-Bench. slight fluctuations. This may be attributed to variations in the initial noise and the additional noise introduced during training to ensure sample diversity. In addition, Gemini 3 Flash may als… view at source ↗
Figure 13
Figure 13. Figure 13: Motivation of Difficulty-aware Advantage Reweighting [PITH_FULL_IMAGE:figures/full_fig_p029_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Multi-reference image generation results of different methods. We compare our method with a range of state-of-the-art open-source and close-source models across diverse representative cases. 36 [PITH_FULL_IMAGE:figures/full_fig_p036_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Multi-reference image generation results of different methods. We compare our method with a range of state-of-the-art open-source and close-source models across diverse representative cases. 37 [PITH_FULL_IMAGE:figures/full_fig_p037_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: Multi-reference image generation results of different methods. We compare our method with a range of state-of-the-art open-source and close-source models across diverse representative cases. 38 [PITH_FULL_IMAGE:figures/full_fig_p038_16.png] view at source ↗
Figure 17
Figure 17. Figure 17: Failure cases of DyRef on the multi-reference image generation task. 39 [PITH_FULL_IMAGE:figures/full_fig_p039_17.png] view at source ↗
Figure 18
Figure 18. Figure 18: System prompt for subject instance generation. For every subject category in Objects365 [27], we leverage DeepSeek-v3 [28] to generate 20 concrete subject instances. 40 [PITH_FULL_IMAGE:figures/full_fig_p040_18.png] view at source ↗
Figure 19
Figure 19. Figure 19: System prompt for T2I prompt generation. We leverage Gemini 3 Flash [29] to integrate logically-consistent combinations of subject instances into a coherent scene description, which is then formulated as the final T2I prompt. 41 [PITH_FULL_IMAGE:figures/full_fig_p041_19.png] view at source ↗
Figure 20
Figure 20. Figure 20: System prompt for MLLM-based evaluation. The MLLM is required to provide both assessment scores and concise rationales to enhance reliability and facilitate debugging. Notably, a score of −1 is assigned to a reference type as a placeholder if the reference image is absent; such scores are excluded from the final score aggregation. 42 [PITH_FULL_IMAGE:figures/full_fig_p042_20.png] view at source ↗
Figure 21
Figure 21. Figure 21: User prompt for MLLM-based evaluation. These prompts incorporate granular scoring rubrics for each dimension, which appear as placeholders in this prompt, and are elaborated in figs. 22 to 27. 43 [PITH_FULL_IMAGE:figures/full_fig_p043_21.png] view at source ↗
Figure 22
Figure 22. Figure 22: Evaluation rubrics for prompt following. 44 [PITH_FULL_IMAGE:figures/full_fig_p044_22.png] view at source ↗
Figure 23
Figure 23. Figure 23: Evaluation rubrics for subject consistency. Evaluation Rubrics: Style Consistency [Style Consistency] Goal: Determine if the Target adopts the visual style of the Style Reference. Focus: Art medium, brushwork, color palette, texture, and overall visual atmosphere should be considered, while the subject content should be ignored. Scoring (0–10, Conservative Scoring): • −1: Reference missing / Not Applicabl… view at source ↗
Figure 24
Figure 24. Figure 24: Evaluation rubrics for style consistency. 45 [PITH_FULL_IMAGE:figures/full_fig_p045_24.png] view at source ↗
Figure 25
Figure 25. Figure 25: Evaluation rubrics for lighting consistency. Evaluation Rubrics: Background Consistency [Background Consistency] Goal: Determine whether the target image preserves the structural integrity and layout of the Background Reference. Focus: Scene layout, object placement, perspective, and depth should be considered. If a style reference is also provided, texture or color changes are allowed, but the underlying… view at source ↗
Figure 26
Figure 26. Figure 26: Evaluation rubrics for background consistency. 46 [PITH_FULL_IMAGE:figures/full_fig_p046_26.png] view at source ↗
Figure 27
Figure 27. Figure 27: Evaluation rubrics for aesthetic. 47 [PITH_FULL_IMAGE:figures/full_fig_p047_27.png] view at source ↗
read the original abstract

While personalized image generation has achieved remarkable progress, multi-reference image generation (MRIG) remains a challenging task. Most existing benchmarks fail to adequately evaluate complex MRIG scenarios, hindering further progress in this area. To better assess model performance on complex MRIG tasks, we introduce OmniRef-Bench, a benchmark that covers complex combinations of reference image types and a large number of reference images. Evaluations on OmniRef-Bench show that mainstream open-source models struggle in complex MRIG scenarios, and their performance deteriorates significantly as the number of mixed-type reference images increases. To address this issue, we propose DyRef, a two-stage training framework. In the first stage, supervised fine-tuning equips the model with the basic capability to handle complex MRIG tasks. In the second stage, we introduce Difficulty-aware Advantage Reweighting (DAR) and Discriminative Reward Scaling (DRS). DAR dynamically adjusts the optimization objective to improve performance when handling a large number of mixed-type reference images. DRS enlarges intra-group reward differences for more effective policy optimization. Experiments demonstrate that DyRef significantly improves the performance of open-source models on OmniRef-Bench and single-image editing benchmarks, demonstrating the effectiveness and generalization capability of our approach.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper introduces OmniRef-Bench to evaluate complex multi-reference image generation (MRIG) with mixed reference types and large reference counts. It reports that existing open-source models degrade as reference complexity increases. The authors propose DyRef, a two-stage framework: supervised fine-tuning (SFT) to acquire basic MRIG capability, followed by Difficulty-aware Advantage Reweighting (DAR) and Discriminative Reward Scaling (DRS) for dynamic optimization. Experiments are stated to show that DyRef yields significant gains on OmniRef-Bench and single-image editing benchmarks, with claims of effectiveness and generalization.

Significance. A new benchmark targeting under-evaluated complex MRIG scenarios would be useful if accompanied by reproducible baselines. If the reported gains are shown to arise specifically from DAR and DRS rather than additional training or data exposure, the two-stage dynamic-reward approach could offer a practical route to scaling open-source models on multi-reference tasks. The manuscript does not yet supply the quantitative tables, error bars, or controlled ablations needed to evaluate this potential.

major comments (2)
  1. [Experiments] Experiments section: The central claim that DAR and DRS produce the observed gains on OmniRef-Bench is load-bearing, yet no ablation is reported that holds total training steps, data, and base model fixed while toggling only the second-stage reward components. Without such controls it remains possible that gains arise from extra optimization rather than the proposed dynamic mechanisms.
  2. [Experiments] §4 (or equivalent results section): No quantitative metrics, standard deviations, or per-reference-count breakdowns are referenced to support the abstract statement that performance 'deteriorates significantly' with increasing mixed-type references; the absence of these numbers prevents verification of the deterioration trend that motivates DyRef.
minor comments (2)
  1. [Abstract] The abstract would benefit from at least one concrete metric (e.g., FID or CLIP score delta) and the number of references at which deterioration is observed.
  2. [Method] Notation for DAR and DRS reward terms should be defined explicitly with equations in the method section to allow readers to assess whether the reweighting is parameter-free or introduces new hyperparameters.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback on the experimental section. We address the major comments point by point below and will revise the manuscript to incorporate the requested controls and quantitative details.

read point-by-point responses

  1. Referee: [Experiments] Experiments section: The central claim that DAR and DRS produce the observed gains on OmniRef-Bench is load-bearing, yet no ablation is reported that holds total training steps, data, and base model fixed while toggling only the second-stage reward components. Without such controls it remains possible that gains arise from extra optimization rather than the proposed dynamic mechanisms.

    Authors: We agree that a controlled ablation isolating the contribution of DAR and DRS is essential. In the revised manuscript, we will add an ablation where total training steps, data volume, and base model are held fixed, directly comparing SFT alone against SFT followed by the DAR/DRS stage to demonstrate that the gains derive specifically from the dynamic reward mechanisms. revision: yes

  2. Referee: [Experiments] §4 (or equivalent results section): No quantitative metrics, standard deviations, or per-reference-count breakdowns are referenced to support the abstract statement that performance 'deteriorates significantly' with increasing mixed-type references; the absence of these numbers prevents verification of the deterioration trend that motivates DyRef.

    Authors: We acknowledge the absence of these supporting numbers in the current version. The revised results section will include detailed quantitative tables reporting metrics with standard deviations and per-reference-count breakdowns across mixed reference types to substantiate the deterioration trend on OmniRef-Bench. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical claims rest on benchmark comparisons without self-referential derivations

full rationale

The paper describes a two-stage empirical training procedure (SFT followed by DAR/DRS reward optimization) and reports performance gains on OmniRef-Bench and editing tasks. No equations, uniqueness theorems, fitted parameters renamed as predictions, or self-citation chains appear in the provided text that would reduce any claimed result to its own inputs by construction. The central claims are falsifiable via external benchmarks and are not tautological, so the derivation chain is self-contained.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

The work is an empirical ML engineering paper; the abstract contains no explicit mathematical axioms, free parameters, or newly postulated physical entities. All claims rest on the unstated assumption that standard diffusion or autoregressive image models plus the described reward adjustments are sufficient to capture the target distribution.

pith-pipeline@v0.9.1-grok · 5769 in / 1305 out tokens · 18373 ms · 2026-06-26T05:04:27.059019+00:00 · methodology

discussion (0)

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

Reference graph

Works this paper leans on

75 extracted references · 21 linked inside Pith

  1. [1]

    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

  2. [2]

    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

  3. [3]

    Qwen-image technical report, 2025

    Chenfei Wu, Jiahao Li, Jingren Zhou, Junyang Lin, Kaiyuan Gao, Kun Yan, Sheng ming Yin, Shuai Bai, Xiao Xu, Yilei Chen, Yuxiang Chen, Zecheng Tang, Zekai Zhang, Zhengyi Wang, An Yang, Bowen Yu, Chen Cheng, Dayiheng Liu, Deqing Li, Hang Zhang, Hao Meng, Hu Wei, Jingyuan Ni, Kai Chen, Kuan Cao, Liang Peng, Lin Qu, Minggang Wu, Peng Wang, Shuting Yu, Tingkun...

    Pith/arXiv arXiv 2025

  4. [4]

    Step1x-edit: A practical framework for general image editing.arXiv preprint arXiv:2504.17761, 2025

    Shiyu Liu, Yucheng Han, Peng Xing, Fukun Yin, Rui Wang, Wei Cheng, Jiaqi Liao, Yingming Wang, Honghao Fu, Chunrui Han, et al. Step1x-edit: A practical framework for general image editing.arXiv preprint arXiv:2504.17761, 2025

    Pith/arXiv arXiv 2025

  5. [5]

    FLUX.2: Frontier Visual Intelligence.https://bfl.ai/blog/flux-2, 2025

    Black Forest Labs. FLUX.2: Frontier Visual Intelligence.https://bfl.ai/blog/flux-2, 2025

    2025

  6. [6]

    Dreamo: A unified framework for image customization

    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. InProceedings of the SIGGRAPH Asia 2025 Conference Papers, pages 1–12, 2025

    2025

  7. [7]

    Ominicontrol: Minimal and universal control for diffusion transformer

    Zhenxiong Tan, Songhua Liu, Xingyi Yang, Qiaochu Xue, and Xinchao Wang. Ominicontrol: Minimal and universal control for diffusion transformer. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 14940–14950, 2025

    2025

  8. [8]

    Dreamomni3: Scribble-basededitingandgeneration.arXivpreprintarXiv:2512.22525, 2025

    Bin Xia, Bohao Peng, Jiyang Liu, Sitong Wu, Jingyao Li, Junjia Huang, Xu Zhao, Yitong Wang, Ruihang Chu,BeiYu,etal. Dreamomni3: Scribble-basededitingandgeneration.arXivpreprintarXiv:2512.22525, 2025

    arXiv 2025

  9. [9]

    Uniref-image-edit: Towards scalable and consistent multi- reference image editing.arXiv preprint arXiv:2602.14186, 2026

    Hongyang Wei, Bin Wen, Yancheng Long, Yankai Yang, Yuhang Hu, Tianke Zhang, Wei Chen, Haonan Fan, Kaiyu Jiang, Jiankang Chen, et al. Uniref-image-edit: Towards scalable and consistent multi- reference image editing.arXiv preprint arXiv:2602.14186, 2026

    arXiv 2026

  10. [10]

    Dreamomni2: Multimodalinstruction-basededitingandgeneration

    Bin Xia, Bohao Peng, Yuechen Zhang, Junjia Huang, Jiyang Liu, Jingyao Li, Haoru Tan, Sitong Wu, ChengyaoWang, YitongWang, etal. Dreamomni2: Multimodalinstruction-basededitingandgeneration. arXiv preprint arXiv:2510.06679, 2025

    arXiv 2025

  11. [11]

    3sgen: Unified subject, style, and structure-driven image generation with adaptive task-specific memory.arXiv preprint arXiv:2512.19271, 2025

    Xinyang Song, Libin Wang, Weining Wang, Zhiwei Li, Jianxin Sun, Dandan Zheng, Jingdong Chen, Qi Li, and Zhenan Sun. 3sgen: Unified subject, style, and structure-driven image generation with adaptive task-specific memory.arXiv preprint arXiv:2512.19271, 2025

    arXiv 2025

  12. [12]

    Multibanana: A challenging benchmark for multi-reference text-to-image generation

    Yuta Oshima, Daiki Miyake, Kohsei Matsutani, Yusuke Iwasawa, Masahiro Suzuki, Yutaka Matsuo, and Hiroki Furuta. Multibanana: A challenging benchmark for multi-reference text-to-image generation. arXiv preprint arXiv:2511.22989, 2025. 14 Scaling Multi-Reference Image Generation with Dynamic Reward Optimization (ECCV2026)

    arXiv 2025

  13. [13]

    Generative ai aids personalized product aesthetic generation and evaluation based on style themes.Advanced Engineering Informatics, 68:103756, 2025

    Yao Wang, Jingsen Zhang, Chengyi Shen, Huiling Yu, and Shijian Luo. Generative ai aids personalized product aesthetic generation and evaluation based on style themes.Advanced Engineering Informatics, 68:103756, 2025

    2025

  14. [14]

    Ctr-driven advertising image generation with multimodal large language models

    Xingye Chen, Wei Feng, Zhenbang Du, Weizhen Wang, Yanyin Chen, Haohan Wang, Linkai Liu, Yaoyu Li, Jinyuan Zhao, Yu Li, et al. Ctr-driven advertising image generation with multimodal large language models. InProceedings of the ACM on Web Conference 2025, pages 2262–2275, 2025

    2025

  15. [15]

    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

    arXiv 2025

  16. [16]

    Multiref: Controllable image generation with multiple visual references

    Ruoxi Chen, Dongping Chen, Siyuan Wu, Sinan Wang, Shiyun Lang, Peter Sushko, Gaoyang Jiang, Yao Wan, and Ranjay Krishna. Multiref: Controllable image generation with multiple visual references. In Proceedings of the 33rd ACM International Conference on Multimedia, pages 13325–13331, 2025

    2025

  17. [17]

    Contextgen: Contextual layout anchoring for identity- consistent multi-instance generation.arXiv preprint arXiv:2510.11000, 2025

    Ruihang Xu, Dewei Zhou, Fan Ma, and Yi Yang. Contextgen: Contextual layout anchoring for identity- consistent multi-instance generation.arXiv preprint arXiv:2510.11000, 2025

    arXiv 2025

  18. [19]

    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, 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.arXiv preprint arXiv:2506.18871, 2025

    Pith/arXiv arXiv 2025

  19. [20]

    Nanobanana: Gemini 2.5 flash image model

    Google DeepMind. Nanobanana: Gemini 2.5 flash image model. https://developers. googleblog.com/en/introducing-gemini-2-5-flash-image/, 2025

    2025

  20. [21]

    Seedream 4.0: Toward next-generation multimodal image generation.arXiv preprint arXiv:2509.20427, 2025

    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

    Pith/arXiv arXiv 2025

  21. [22]

    Psr: Scaling multi-subject personalized image generation with pairwise subject-consistency rewards.arXiv preprint arXiv:2512.01236, 2025

    Shulei Wang, Longhui Wei, Xin He, Jianbo Ouyang, Hui Lu, Zhou Zhao, and Qi Tian. Psr: Scaling multi-subject personalized image generation with pairwise subject-consistency rewards.arXiv preprint arXiv:2512.01236, 2025

    Pith/arXiv arXiv 2025

  22. [23]

    Dreambench++: A human-aligned benchmark for personalized image generation

    Yuang Peng, Yuxin Cui, Haomiao Tang, Zekun Qi, Runpei Dong, Jing Bai, Chunrui Han, Zheng Ge, Xiangyu Zhang, and Shu-Tao Xia. Dreambench++: A human-aligned benchmark for personalized image generation. InThe Thirteenth International Conference on Learning Representations, 2025. URL https://dreambenchplus.github.io/

    2025

  23. [24]

    Imgedit: A unified image editing dataset and benchmark.arXiv preprint arXiv:2505.20275, 2025

    Yang Ye, Xianyi He, Zongjian Li, Bin Lin, Shenghai Yuan, Zhiyuan Yan, Bohan Hou, and Li Yuan. Imgedit: A unified image editing dataset and benchmark.arXiv preprint arXiv:2505.20275, 2025

    Pith/arXiv arXiv 2025

  24. [25]

    Laion-5b: An open large-scale dataset for training next generation image-text models.Advances in neural information processing systems, 35:25278–25294, 2022

    Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models.Advances in neural information processing systems, 35:25278–25294, 2022. 15 Scaling Multi-Reference Image G...

    2022

  25. [26]

    Less-to-more gen- eralization: Unlocking more controllability by in-context generation

    Shaojin Wu, Mengqi Huang, Wenxu Wu, Yufeng Cheng, Fei Ding, and Qian He. Less-to-more gen- eralization: Unlocking more controllability by in-context generation. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 18682–18692, 2025

    2025

  26. [27]

    Objects365: A large-scale, high-quality dataset for object detection

    Shuai Shao, Zeming Li, Tianyuan Zhang, Chao Peng, Gang Yu, Xiangyu Zhang, Jing Li, and Jian Sun. Objects365: A large-scale, high-quality dataset for object detection. InProceedings of the IEEE/CVF international conference on computer vision, pages 8430–8439, 2019

    2019

  27. [28]

    Deepseek-v3technicalreport.arXivpreprintarXiv:2412.19437, 2024

    Aixin Liu, Bei Feng, Bing Xue, Bingxuan Wang, Bochao Wu, Chengda Lu, Chenggang Zhao, Chengqi Deng,ChenyuZhang,ChongRuan,etal. Deepseek-v3technicalreport.arXivpreprintarXiv:2412.19437, 2024

    Pith/arXiv arXiv 2024

  28. [29]

    Gemini: a family of highly capable multimodal models.arXiv preprint arXiv:2312.11805, 2023

    Gemini Team, Rohan Anil, Sebastian Borgeaud, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, Katie Millican, et al. Gemini: a family of highly capable multimodal models.arXiv preprint arXiv:2312.11805, 2023

    Pith/arXiv arXiv 2023

  29. [30]

    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

  30. [31]

    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. [32]

    Omniconsistency: Learning style-agnostic consis- tency from paired stylization data

    Yiren Song, Cheng Liu, and Mike Zheng Shou. Omniconsistency: Learning style-agnostic consis- tency from paired stylization data. 2025. URLhttps://api.semanticscholar.org/CorpusID: 278905729

    2025

  32. [33]

    Learning transferable visual models from natural language supervision

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. InInternational conference on machine learning, pages 8748–8763. PmLR, 2021

    2021

  33. [34]

    Dinov2: Learning robust visual features without supervision.arXiv preprint arXiv:2304.07193, 2023

    Maxime Oquab, Timothée Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision.arXiv preprint arXiv:2304.07193, 2023

    Pith/arXiv arXiv 2023

  34. [35]

    Measuring style similarity in diffusion models.arXiv preprint arXiv:2404.01292, 2024

    Gowthami Somepalli, Anubhav Gupta, Kamal Gupta, Shramay Palta, Micah Goldblum, Jonas Geiping, Abhinav Shrivastava, and Tom Goldstein. Measuring style similarity in diffusion models.arXiv preprint arXiv:2404.01292, 2024

    arXiv 2024

  35. [36]

    Vbench: Comprehensive benchmark suite for video generativemodels

    Ziqi Huang, Yinan He, Jiashuo Yu, Fan Zhang, Chenyang Si, Yuming Jiang, Yuanhan Zhang, Tianxing Wu, Qingyang Jin, Nattapol Chanpaisit, et al. Vbench: Comprehensive benchmark suite for video generativemodels. InProceedingsoftheIEEE/CVFConferenceonComputerVisionandPatternRecognition, pages 21807–21818, 2024

    2024

  36. [37]

    Monocular human pose estimation: A survey of deep learning-based methods.Computer vision and image understanding, 192:102897, 2020

    Yucheng Chen, Yingli Tian, and Mingyi He. Monocular human pose estimation: A survey of deep learning-based methods.Computer vision and image understanding, 192:102897, 2020

    2020

  37. [38]

    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, Liang Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models.Iclr, 1(2):3, 2022. 16 Scaling Multi-Reference Image Generation with Dynamic Reward Optimization (ECCV2026)

    2022

  38. [39]

    Flow matching for generative modeling.arXiv preprint arXiv:2210.02747, 2022

    Yaron Lipman, Ricky TQ Chen, Heli Ben-Hamu, Maximilian Nickel, and Matt Le. Flow matching for generative modeling.arXiv preprint arXiv:2210.02747, 2022

    Pith/arXiv arXiv 2022

  39. [40]

    Flow-grpo: Trainingflowmatchingmodelsviaonlinerl.arXivpreprintarXiv:2505.05470, 2025

    Jie Liu, Gongye Liu, Jiajun Liang, Yangguang Li, Jiaheng Liu, Xintao Wang, Pengfei Wan, Di Zhang, and WanliOuyang. Flow-grpo: Trainingflowmatchingmodelsviaonlinerl.arXivpreprintarXiv:2505.05470, 2025

    Pith/arXiv arXiv 2025

  40. [42]

    URLhttps://arxiv.org/abs/2602.12529

    arXiv

  41. [43]

    Uso: Unified style and subject-driven generation via disentangled and reward learning.arXiv preprint arXiv:2508.18966, 2025

    Shaojin Wu, Mengqi Huang, Yufeng Cheng, Wenxu Wu, Jiahe Tian, Yiming Luo, Fei Ding, and Qian He. Uso: Unified style and subject-driven generation via disentangled and reward learning.arXiv preprint arXiv:2508.18966, 2025

    arXiv 2025

  42. [44]

    Siglip 2: Multilingual vision-language encoders with improved semantic understanding, localization, and dense features

    Michael Tschannen, Alexey Gritsenko, Xiao Wang, Muhammad Ferjad Naeem, Ibrahim Alabdulmohsin, Nikhil Parthasarathy, Talfan Evans, Lucas Beyer, Ye Xia, Basil Mustafa, et al. Siglip 2: Multilingual vision-language encoders with improved semantic understanding, localization, and dense features. arXiv preprint arXiv:2502.14786, 2025

    Pith/arXiv arXiv 2025

  43. [45]

    Focal loss for dense object detection

    Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dollár. Focal loss for dense object detection. InProceedings of the IEEE international conference on computer vision, pages 2980–2988, 2017

    2017

  44. [46]

    Umo: Scaling multi- identity consistency for image customization via matching reward.arXiv preprint arXiv:2509.06818, 2025

    Yufeng Cheng, Wenxu Wu, Shaojin Wu, Mengqi Huang, Fei Ding, and Qian He. Umo: Scaling multi- identity consistency for image customization via matching reward.arXiv preprint arXiv:2509.06818, 2025

    arXiv 2025

  45. [47]

    Mosaic: Multi-subject personalized generation via correspondence-aware alignment and disentanglement.arXiv preprint arXiv:2509.01977, 2025

    Dong She, Siming Fu, Mushui Liu, Qiaoqiao Jin, Hualiang Wang, Mu Liu, and Jidong Jiang. Mosaic: Multi-subject personalized generation via correspondence-aware alignment and disentanglement.arXiv preprint arXiv:2509.01977, 2025

    arXiv 2025

  46. [48]

    Null-text inversion for editing real images using guided diffusion models

    Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6038–6047, 2023

    2023

  47. [49]

    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. InProceedings of the IEEE/CVF international conference on computer vision, pages 3836–3847, 2023

    2023

  48. [50]

    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

  49. [51]

    In-context edit: Enabling instructional image editing with in-context generation in large scale diffusion transformer.arXiv preprint arXiv:2504.20690, 2025

    Zechuan Zhang, Ji Xie, Yu Lu, Zongxin Yang, and Yi Yang. In-context edit: Enabling instructional image editing with in-context generation in large scale diffusion transformer.arXiv preprint arXiv:2504.20690, 2025

    Pith/arXiv arXiv 2025

  50. [52]

    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. 17 Scaling Multi-Reference Image Generation with Dynamic Reward Optimization (ECCV2026)

    Pith/arXiv arXiv 2025

  51. [53]

    Unireal: Universal image generation and editing via learning real- world dynamics

    Xi Chen, Zhifei Zhang, He Zhang, Yuqian Zhou, Soo Ye Kim, Qing Liu, Yijun Li, Jianming Zhang, Nanxuan Zhao, Yilin Wang, et al. Unireal: Universal image generation and editing via learning real- world dynamics. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 12501–12511, 2025

    2025

  52. [54]

    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, and Jidong Jiang. Focusdpo: Dynamic preference optimization for multi-subject personalized image generation via adaptive focus. arXiv preprint arXiv:2509.01181, 2025

    arXiv 2025

  53. [55]

    Training language models to follow instructions with human feedback.Advances in neural information processing systems, 35:27730–27744, 2022

    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

  54. [56]

    Skywork unipic 2.0: Building kontext model with online rl for unified multimodal model.arXiv preprint arXiv:2509.04548, 2025

    Hongyang Wei, Baixin Xu, Hongbo Liu, Size Wu, Jie Liu, Yi Peng, Peiyu Wang, Zexiang Liu, Jingwen He, Yidan Xietian, et al. Skywork unipic 2.0: Building kontext model with online rl for unified multimodal model.arXiv preprint arXiv:2509.04548, 2025

    arXiv 2025

  55. [57]

    Alphapose: Whole-body regional multi-person pose estimation and tracking in real-time.IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022

    Hao-Shu Fang, Jiefeng Li, Hongyang Tang, Chao Xu, Haoyi Zhu, Yuliang Xiu, Yong-Lu Li, and Cewu Lu. Alphapose: Whole-body regional multi-person pose estimation and tracking in real-time.IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022

    2022

  56. [58]

    Decoupled weight decay regularization.arXiv preprint arXiv:1711.05101, 2017

    Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization.arXiv preprint arXiv:1711.05101, 2017

    Pith/arXiv arXiv 2017

  57. [59]

    Adam: A method for stochastic optimization.(No Title), 2014

    Kingma Diederik. Adam: A method for stochastic optimization.(No Title), 2014

    2014

  58. [60]

    Gdpo: Group reward-decoupled normalization policy optimization for multi-reward rl optimization.arXiv preprint arXiv:2601.05242, 2026

    Shih-Yang Liu, Xin Dong, Ximing Lu, Shizhe Diao, Peter Belcak, Mingjie Liu, Min-Hung Chen, Hongxu Yin, Yu-Chiang Frank Wang, Kwang-Ting Cheng, et al. Gdpo: Group reward-decoupled normalization policy optimization for multi-reward rl optimization.arXiv preprint arXiv:2601.05242, 2026

    Pith/arXiv arXiv 2026

  59. [61]

    Qwen2.5-vl technical report.arXiv preprint arXiv:2502.13923, 2025

    Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, JunTang, HumenZhong, YuanzhiZhu, MingkunYang, ZhaohaiLi, JianqiangWan, PengfeiWang, Wei Ding, Zheren Fu, Yiheng Xu, Jiabo Ye, Xi Zhang, Tianbao Xie, Zesen Cheng, Hang Zhang, Zhibo Yang, Haiyang Xu, and Junyang Lin. Qwen2.5-vl technical report.arXiv pre...

    Pith/arXiv arXiv 2025

  60. [62]

    Auto-encoding variational bayes.arXiv preprint arXiv:1312.6114, 2013

    Diederik P Kingma and Max Welling. Auto-encoding variational bayes.arXiv preprint arXiv:1312.6114, 2013

    Pith/arXiv arXiv 2013

  61. [63]

    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 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13294–13304, 2025

    2025

  62. [64]

    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. InForty-first international conference on machine learning, 2024

    2024

  63. [65]

    using the scene of image X as background

    Black Forest Labs. Flux.https://github.com/black-forest-labs/flux, 2024. 18 Scaling Multi-Reference Image Generation with Dynamic Reward Optimization (ECCV2026) Appendix Overview This material provides supplementary details to the main paper, including the following sections: A Related Work. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ...

    2024

  64. [66]

    You will be given an<asset category> , you need to create an asset (brief subject prompt) based on the <asset category>

  65. [67]

    Elon Musk in pajamas

    These descriptions can refer only to appearance descriptions/or to certain brands. e.g., “Elon Musk in pajamas”, “a tiger in a black hat”, “A Mercedes sports car”, “A blonde”, “A door red on the left and green on the right”

  66. [68]

    Avoid adding separate accessories or objects

    Focus on the given<asset category>ONLY. Avoid adding separate accessories or objects

  67. [69]

    Do not repeat each asset, you need to use your logic and common sense of life to create

  68. [70]

    love and power

    No more than 12 words for each asset. Example [asset category]:Book Output: [asset1]:A book with a green cover [asset2]:comic book [asset3]:math book [asset4]:An open book [asset5]:Rotten books [asset6]:The book with “love and power” on the cover [asset7]:A book with a blue key on it ... (Up to[asset20]) User: [asset category]:{category} Figure 18:System ...

  69. [71]

    Merge and prioritize attributes across all subjects; resolve conflicts logically

  70. [72]

    secondary)

    Clarify relationships, composition, and focal hierarchy (primary subject vs. secondary)

  71. [73]

    Add tasteful art direction: lighting, environment, background, camera, lens, shot type, color palette, mood, time of day, material details, and post-processing

  72. [74]

    Preserve each subject’s core identity; distribute attributes sensibly (don’t duplicate or contradict)

  73. [75]

    For each person subject, specify adiversifiedpose ranging from static to dynamic (e.g., sitting with legs crossed, jumping with arms outstretched)

  74. [76]

    A woman with neon pink hair

    Input format: •subject_phrases: a list of short noun phrases describing subjects or objects. •variants: number of alternative prompts to produce (default: 1). Output format:Produce exactly N variants (N = variants). For each variant, output a single cohesive prompt ready for a text-to-image model. Output requirements:Return ONLY the prompt text itself. Ex...

  75. [77]

    Scoring(0–10,Conservative Scoring): •0: Completely irrelevant to the text prompt, or completely ignores the core reference requirement

    A failure to match a Reference Image (Style, Subject, Background, Lighting) should result in a score lower than 6. Scoring(0–10,Conservative Scoring): •0: Completely irrelevant to the text prompt, or completely ignores the core reference requirement. •2: The main subject is present, but the content or reference requirement is seriously violated. • 4: The ...