Research map / Vision-language and multimodal / Multimodal LLMs

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

NeurIPS 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Reinforcement learning for visual reasoning · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes VisionCreator-R1, a native visual generation agent with explicit reflection trained via RPCO (SFT on mixed reflection/planning trajectories followed by multi-task RL), outperforming Gemini2.5Pro on VCR-Bench and GEdit-Bench.

Problem

Existing visual generation agents are plan-driven and lack systematic reflection to correct mid-trajectory visual errors, and reflection learned in single-image settings does not transfer to multi-image workflows under RL.

Approach

Reflection-Plan Co-Optimization (RPCO) decouples reflection learning on single-image tasks, then mixes reflection-strong single-image and planning-strong multi-image trajectories in SFT, then runs multi-task RL with plan and reflection rewards; a theoretical variance-asymmetry argument motivates the staging.

Key results

• VisionCreator-R1 reaches 0.700 on multi-image VCR-Bench vs. 0.649 for Gemini2.5Pro, and 0.532 on single-image vs. 0.515.
• Plan score rises from 0.8285 (Strong-Reflection) and 0.8431 (Reflection-Plan Conflict) to 0.9746 for VisionCreator-R1 (Gemini2.5Pro 0.9667).
• On GEdit-Bench, overall score G_O is 7.23, the highest among compared models (Step1X-Edit 6.97, Qwen-Image-Fast 7.03).

Surprising findings

• Direct RL transfer of single-image reflection to multi-image tasks degrades reflection quality (good-reflection drops from 21.6% to 16.5%) even while task score improves.
• SFT-only reflection training leads to a strong under-reflection bias (58.2%) and degraded single-image performance.

Limitations

• Reflection quality does not improve in the final multi-image model (63.5%/31.0%/5.5% vs. Reflection-Plan SFT), and the theoretical asymmetry argument is largely heuristic with approximations.
• Main evaluation relies on Gemini2.5Pro and Qwen3-VL-32B as VLM judges; human evaluation is only pairwise against Gemini2.5Pro.

Experimental setup

Datasets: VCR-SFT VCR-RL GEdit-Bench

Benchmarks: VCR-Bench GEdit-Bench

Metrics: checkpoint-based task score reflection quality (under/good/over) plan score G_SC G_PQ G_O human pairwise preference

Baselines: Gemini2.5Pro Qwen3VL32B Qwen-Image-Fast Gemini 2.0 BAGEL FLUX.1 Kontext Pro Step1X-Edit

Builds on: GRPO VisionCreator Qwen-Image-Edit-2509-Lightning

Closest papers by meaning

• Generation Navigator: A State-Aware Agentic Framework for Image Generation (NeurIPS 2026 · similarity 0.87)
• COVR: Collaborative Optimization of VLMs and RL Agent for Visual-Based Control (AAAI 2026 · similarity 0.86)
• Gen-Searcher: Reinforcing Agentic Search for Image Generation (NeurIPS 2026 · similarity 0.86)
• MaskSense: Confronting the Visual Exploration Trap in Masked Image Generation (NeurIPS 2026 · similarity 0.86)
• FLIP: Flow-Centric Generative Planning as General-Purpose Manipulation World Model (ICLR 2025 · similarity 0.85)
• Benchmarking and Evolving Reason-Reflect-Rectify for Reflective Visual Generation (ICML 2026 · similarity 0.85)