Research map / Image generation / Text-to-image diffusion

Generation Navigator: A State-Aware Agentic Framework for Image Generation

NeurIPS 2026 · Image generation › Text-to-image diffusion · approach: Layout-to-image control · 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 Generation Navigator, a multi-turn T2I agent whose MLLM navigator learns state-conditioned Stop/Refine/Regenerate actions via SFT on 103K trajectories and PRE-GRPO, a trajectory-level RL objective rewarding peak quality, retention, and efficiency.

Problem

Text-to-image generation often fails to satisfy user intent in one pass, and existing prompt-rewriting or fixed-workflow closed-loop agents do not learn state-dependent actions; vanilla RL rewards also assign credit poorly across multi-turn generation trajectories.

Approach

Reformulate T2I generation as a state-conditioned action-making problem where a learned MLLM navigator chooses Stop, Refine, or Regenerate plus a revised prompt, trained with PRE-GRPO whose trajectory reward combines peak score, final-score retention, turn efficiency, and format correctness.

Key results

• Generation Navigator reaches 79.06% overall reasoning accuracy on T2I-ReasonBench, versus 57.70% for FLUX.2-Klein-9B and 73.57% for StruVis.
• On WISE, the method achieves an overall score of 0.90, above SeedDream4.0 (0.82) and GPT-4o (0.80).
• A no-CoT navigator reaches 76.0% accuracy at 21.7 s per sample, 2.4x faster than the full-CoT version (77.4%, 52.5 s).

Surprising findings

• Training-free agent and SFT variants fall below the one-shot baseline on GenEval, attributed to a narrow action distribution.
• Vanilla GRPO with a best-score reward over-uses Regenerate (65.15%), while the pilot study shows Regenerate and Refine each win on different cases.

Limitations

• Reliance on a single MLLM reviewer for reward and scoring introduces possible reviewer bias, though robustness across reviewers is tested.
• Long chain-of-thought navigator adds substantial latency (about 60% of overhead).

Experimental setup

Datasets: T2I-ReasonBench WISE GenEval

Benchmarks: T2I-ReasonBench WISE GenEval

Metrics: reasoning accuracy image quality WISE overall score GenEval score latency

Baselines: GPT-4o Gemini-2.0 SeedDream4.0 Qwen-Image FLUX.2-Klein-9B GenAgent Think-then-Generate IRG StruVis Vanilla GRPO

Builds on: GRPO Qwen3-VL FLUX.2-Klein PromptEnhancer

Closest papers by meaning

• Gen-Searcher: Reinforcing Agentic Search for Image Generation (NeurIPS 2026 · similarity 0.90)
• Preference Adaptive and Sequential Text-to-Image Generation (ICML 2025 · similarity 0.87)
• T2I-R1: Reinforcing Image Generation with Collaborative Semantic-level and Token-level CoT (NeurIPS 2025 · similarity 0.87)
• VisionCreator-R1: A Reflection-Enhanced Native Visual-Generation Agentic Model (NeurIPS 2026 · similarity 0.87)
• GoT-R1: Unleashing Reasoning Capability of Autoregressive Visual Generation with Reinforcement Learning (ICLR 2026 · similarity 0.87)
• Grounded Answers for Multi-agent Decision-making Problem through Generative World Model (NeurIPS 2024 · similarity 0.86)