VisionCreator-R1: A Reflection-Enhanced Native Visual-Generation Agentic Model

2026-03-11 · Source: cs.CV updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Visual Generation · Depth: Expert, extended

Summary

VisionCreator-R1 is a new native visual generation agent developed by Tencent Hunyuan and Hong Kong University of Science and Technology, designed to overcome limitations in existing plan-driven agents by incorporating explicit reflection mechanisms. The model utilizes a Reflection–Plan Co-Optimization (RPCO) training methodology, which addresses an identified asymmetry in reinforcement learning where planning is reliably optimized, but reflection learning is hindered by noisy credit assignment in multi-image tasks. RPCO first trains on the VCR-SFT dataset, combining reflection-strong single-image trajectories and planning-strong multi-image trajectories, then co-optimizes on the VCR-RL dataset via RL. VisionCreator-R1 consistently outperforms Gemini2.5Pro on existing benchmarks and the new VCR-bench, which covers single-image, multi-image, and image-to-image tasks, achieving an overall score of 7.23 on GEdit-Bench and significant gains on multi-image tasks in VCR-Bench.

Key takeaway

For Computer Vision Engineers developing advanced visual generation agents, integrating explicit reflection mechanisms is critical for overcoming error accumulation in complex, multi-image workflows. You should consider adopting a "decoupled-then-fused" training strategy, starting with supervised fine-tuning on diverse, high-quality datasets before applying multi-task reinforcement learning, to effectively co-optimize planning and reflection capabilities in long-horizon stochastic environments.

Key insights

Explicit reflection and co-optimization of planning and reflection are crucial for robust visual generation agents.

Principles

• Reflection optimization is difficult in long-horizon RL due to noisy credit assignment.
• Planning rewards are stable; reflection rewards are highly stochastic.
• Decoupled-then-fused training improves reflection and planning synergy.

Method

The Reflection–Plan Co-Optimization (RPCO) methodology involves initial supervised fine-tuning on mixed datasets, followed by multi-task reinforcement learning to synergistically optimize planning and reflection capabilities.

In practice

• Use VLM-based judges for reflection reward computation.
• Combine reflection-strong and planning-strong data for SFT initialization.
• Implement multi-dimensional reward systems for comprehensive RL guidance.

Topics

• Visual Generation Agents
• Reflection Mechanisms
• Reinforcement Learning Optimization
• Multi-Image Generation
• RPCO Training Methodology

Code references

• black-forest-labs/flux

Best for: Research Scientist, Computer Vision Engineer, AI Researcher, AI Scientist, Deep Learning Engineer

Related on AIssential

• InterleaveThinker: Reinforcing Agentic Interleaved Generation — Multi-agent orchestration enables existing image generators to perform complex interleaved text-image sequence generation.
• Think Twice, Act Once: Verifier-Guided Action Selection For Embodied Agents — Explicit test-time verification with a specialized, generatively trained verifier significantly boosts embodied agent robustness.
• Image Generators are Generalist Vision Learners — Instruction-tuning image generators unlocks powerful generalist visual understanding capabilities, mirroring LLM emergent behaviors.
• Interactor: Agentic RL oriented Iterative Creation for Ad Description Generation in Sponsored Search — Interactor uses agentic RL with generative reward models for iterative, high-quality ad description generation.
• 🎙️"We Are the Only Ones Who Would Build It" — Open-weight models, combined with reinforcement learning, accelerate AI progress, enhance safety, and democratize access to frontier capabilities.
• Train the Agent, Not the Expert: Learning to Harness Heterogeneous Experts for Multi-Turn Visual Reasoning — VisHarness trains an agent to orchestrate specialized visual experts for general visual reasoning, enhancing performance and adaptability.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.CV updates on arXiv.org.