Visual content generation has advanced from single-image to multi-image workflows, yet existing agents remain largely plan-driven and lack systematic reflection mechanisms to correct mid-trajectory visual errors. To address this limitation, we propose VisionCreator-R1, a native visual generation agent with explicit reflection, together with a Reflection-Plan Co-Optimization (RPCO) training methodology. Through extensive experiments and trajectory-level analysis, we uncover reflection-plan optimization asymmetry in reinforcement learning (RL): planning can be reliably optimized via plan rewards, while reflection learning is hindered by noisy credit assignment. Guided by this insight, our RPCO first trains on the self-constructed VCR-SFT dataset with reflection-strong single-image trajectories and planning-strong multi-image trajectories, then co-optimization on VCR-RL dataset via RL. This yields our unified VisionCreator-R1 agent, which consistently outperforms Gemini2.5Pro on existing benchmarks and our VCR-bench covering single-image and multi-image tasks.
Figures & tables
Figure 1: Comparison between without-Reflection, with Good-Reflection, Under-Reflection and Over-Reflection.
Figure 2: Performance - Reflection Quality.
Model
Single-Img
Ref. Quality (U/G/O)
Qwen-Image-Fast
0.505
–
Gemini2.5Pro
0.515
32.0%/55.4%/12.6%
Qwen3VL32B
0.500
41.3%/49.7%/9.0%
SFT-Reflection
0.499
58.2%/38.9%/2.9%
Reflection-s30
0.510
47.7%/48.9%/3.3%
Reflection-s130
0.521
33.7%/58.5%/7.8%
Table 1: Reflection experiment on Single-Img task in VCR-Bench. Ref. Quality (U/G/O): reflection quality with under-/good-/over-reflection percentages. And all models’ plan score on the single-image task is larger than 99.5%.
Model
Multi-Img
Ref. Quality (U/G/O)
Plan Score
SR
0.565
68.6%/21.6%/9.8%
0.8285
RPC
0.617
73.9%/16.5%/9.7%
0.8431
GPro
0.649
61.6%/31.8%/6.6%
0.9667
Table 2: Transfer from single-image reflection to multi-image tasks. Model: SR is Strong-Reflection model, RPC is Reflection-Plan Conflict model, GPro is Gemini2.5Pro.
Visual content creation tasks demand a nuanced understanding of design conventions and creative workflows-capabilities challenging for general models, while workflow-based agents lack specialized knowledge for autonomous creative planning. To overcome these challenges, we propose VisionCreator, a native visual-generation agentic model that unifies Understanding, Thinking, Planning, and Creation (UTPC) capabilities within an end-to-end learnable framework. Our work introduces four key contributions: (i) VisGenData-4k and its construction methodology using metacognition-based VisionAgent to generate high-quality creation trajectories with explicit UTPC structures; (ii) The VisionCreator agentic model, optimized through Progressive Specialization Training (PST) and Virtual Reinforcement Learning (VRL) within a high-fidelity simulated environment, enabling stable and efficient acquisition of UTPC capabilities for complex creation tasks; (iii) VisGenBench, a comprehensive benchmark featuring 1.2k test samples across diverse scenarios for standardized evaluation of multi-step visual creation capabilities; (iv) Remarkably, our VisionCreator-8B/32B models demonstrate superior performance over larger closed-source models across multiple evaluation dimensions. Overall, this work provides a foundation for future research in visual-generation agentic systems.
Jinxiang Lai, Zexin Lu, Jiajun He +11
Hong Kong University of Science and Technology · Tencent Hunyuan
Text-to-Image (T2I) models and Unified Multimodal Models (UMMs) have achieved remarkable progress in visual generation. However, their reliance on a single-pass generation paradigm limits their ability to handle complex prompts requiring iterative refinement. To enable multi-round Reflective Visual Generation (RVG), we formalize the Reason-Reflect-Rectify (R^3) loop as a core framework and introduce R^3-Bench, a benchmark of over 600 expert-annotated instances that quantifies iterative reasoning and rectification capabilities. Evaluation on R^3-Bench reveals a critical gap: while state-of-the-art models can identify generation errors, they fail to generate actionable rectification instructions. To bridge this gap, we propose R^3-Refiner, a dual-stage framework leveraging Group Relative Policy Optimization (GRPO) and a Hierarchical Reward Mechanism (HRM) to better align rectification with reflective reasoning. Experiments show that R^3-Refiner achieves significant improvements on R^3-Bench (+12.0% in Reflective Verdict Score, +9.0% in Rectification Score), and can be seamlessly integrated with various MLLMs to enhance the generation quality of different T2I models on GenEval++ and T2I-CompBench. Code is available at https://github.com/xiaomoguhz/R3-Bench.
Junjie Wang, Xinghua Lou, Jason Li +8
Harbin Institute of Technology, Shenzhen · University of Science and Technology of China · Shenzhen Loop Area Institute +5
Agentic visual generation produces experience that can improve the native understanding (agent) and generation roles of a unified multimodal backbone, but task success alone does not determine which role should learn. Generator updates can also make the agent's action supervision stale. We introduce RSIgen, a task-centric reciprocal self-improvement framework linking learning allocation, supervision refresh, and recursive curriculum construction. Dynamic Responsibility Routing (DRR) uses matched interventions on evidence, inputs, and supported actions to select learning recipients and construct compatible training targets. After generator learning, RSIgen re-evaluates the relative utility of tested actions with branch inputs held fixed to refresh agent supervision; the retained roles then construct the next curriculum. Across BAGEL, Qwen-Image, and Wan-Image-7B, RSIgen improves AgentGen-Bench Full Overall9 by 3.2--3.3 points over a three-cycle SearchGen adaptation with action costs. Wan controls support allocation gains at matched role-level training exposure, refresh gains at a matched adaptation budget, and benefits from updated curriculum sources while recipients continue learning. In blind evaluation on 200 real-user requests, pooled tie-adjusted preference scores against the same baseline reach 70.0% for task completion and 64.9% for visual quality.
Yidi Liu, Haozhe Wang, Jie Xiao +3
University of Science and Technology of China · The Hong Kong University of Science and Technology · Wan Team, Alibaba Group