# Scaling Multi-Reference Image Generation with Dynamic Reward Optimization

Source: [Scaling Multi-Reference Image Generation with Dynamic Reward Optimization](<https://www.sota2.com/research/paper/6a3e0791206542251d7cb496>)

Collection index: [llms.txt](<https://www.sota2.com/research/llms.txt>)

Data updated: 2026-06-27T05:00:11.922Z

Authors: Wenwang Huang, Yusen Fu, Junjie Wang, Mengfei Huang, Yulin Li, Gan Liu, Jing Cai, Yancheng He, Zhuotao Tian

Published: 2026-06-25T12:21:13.000Z

## 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.

## Recorded benchmark results

A research result applies to its stated dataset, split, metric and evaluation configuration. Compare like-for-like settings, preserve the metric's unit and direction, and distinguish paper-reported results from independent reproduction. The best result recorded here is not a claim of exhaustive, current global coverage.

- [Multi-Reference Image Generation on DreamBench++](<https://www.sota2.com/research/sota/multi-reference-image-generation-on-dreambench-plus-plus.md>): Multi-reference image generation; Dreambench++; CP (Animal): 68.

- [Image Editing on DreamBench++](<https://www.sota2.com/research/sota/image-editing-on-dreambench-plus-plus.md>): Image Editing; Dreambench++; DINOv1 Score: 0.6.

- [Multi-reference Image Generation on MultiBanana (Text and Tone Metrics)](<https://www.sota2.com/research/sota/multi-reference-image-generation-on-multibanana-text-and-tone-metrics.md>): Multi-reference image generation; MultiBanana; Text Quality Score: 3.82.

- [Multi-reference Image Generation on MultiRef](<https://www.sota2.com/research/sota/multi-reference-image-generation-on-multiref.md>): Multi-reference image generation; MultiRef; BBox Score: 8.71.

- [Single-Image Editing on ImgEdit Benchmark](<https://www.sota2.com/research/sota/single-image-editing-on-imgedit-benchmark.md>): Single-image editing; ImgEdit Benchmark; Extract Score: 4.07.

- [Multi-reference image generation on OmniRef-Bench 1.0 (test)](<https://www.sota2.com/research/sota/multi-reference-image-generation-on-omniref-bench-1-0-test.md>): Multi-reference image generation; OmniRef-Bench 1.0 (test); Score: 8.38.

- [Multi-reference image generation on User Study 50 samples](<https://www.sota2.com/research/sota/multi-reference-image-generation-on-user-study-50-samples.md>): Multi-reference image generation; User Study 50 samples; Subjective Score: 8.38.

- [Subject-Consistent Image Generation on OmniContext](<https://www.sota2.com/research/sota/subject-consistent-image-generation-on-omnicontext.md>): Subject-consistent image generation; OmniContext; Fidelity (Single, Character): 8.59.

- [Multi-reference image generation on MultiBanana](<https://www.sota2.com/research/sota/multi-reference-image-generation-on-multibanana.md>): Multi-reference image generation; MultiBanana; Fidelity (Single Object): 6.34.

- [Multi-Reference Image Generation on OmniRef-Bench](<https://www.sota2.com/research/sota/multi-reference-image-generation-on-omniref-bench.md>): Multi-reference image generation; OmniRef-Bench; Objective Subject Score: 0.63.

## Original sources and implementation

[Original paper](<https://arxiv.org/abs/2606.26947>)

[PDF](<https://arxiv.org/pdf/2606.26947.pdf>)
