Precise Text-to-Image (T2I) generation has achieved great success but is hindered by the limited relational reasoning of static text encoders and the error accumulation in open-loop sampling. Without real-time feedback, initial semantic ambiguities during the Ordinary Differential Equation trajectory inevitably escalate into stochastic deviations from spatial constraints. To bridge this gap, we introduce AFS-Search (Agentic Flow Steering and Parallel Rollout Search), a training-free closed-loop framework built upon FLUX.1-dev. AFS-Search incorporates a training-free closed-loop parallel rollout search and flow steering mechanism, which leverages a Vision-Language Model (VLM) as a semantic critic to diagnose intermediate latents and dynamically steer the velocity field via precise spatial grounding. Complementarily, we formulate T2I generation as a sequential decision-making process, exploring multiple trajectories through lookahead simulations and selecting the optimal path based on VLM-guided rewards. Further, we provide AFS-Search-Pro for higher performance and AFS-Search-Fast for quicker generation. Experimental results show that our AFS-Search-Pro greatly boosts the performance of the original FLUX.1-dev, achieving state-of-the-art results across three different benchmarks. Meanwhile, AFS-Search-Fast also significantly enhances performance while maintaining fast generation speed.
Figures & tables
Figure 1: From a visual perspective, our AFS-Search provides a closed-loop generation paradigm to achieve precise spatial grounding generation.
Figure 2: Motivation of our AFS-Search. Open-loop generation follows a fixed, feed-forward sampling trajectory without intermediate feedback or correction while closed-loop generation introduces real-time visual feedback.
Figure 3: The framework of AFS-Search. The pipeline operates in four phases: (1) Prompt Optimization. A VLM rewrites the user prompt into an explicit instruction. (2) Generating Initial Structure. The FLUX.1-dev model generates an intermediate state up to a bifurcation point. (3) Parallel Rollout Search. A VLM Critic diagnoses the intermediate state to guide search. The system explores three branches: a Baseline Branch , an Exploration Branch , and a Corrective Branch by AFS. (4) Output & Feedback: The optimal trajectory is selected based on VLM scores. If the scores are lower than threshold, a global redesign loop is triggered.
Figure 4: Motivation for Parallel Rollout Search. While the standard open-loop trajectory (Base Branch) yields a sub-optimal alignment (Score: 8.5), our search mechanism actively explores alternative futures. By comparing the Corrective Branch (guided by AFS) and Exploration Branch against the baseline, the agent identifies and selects the optimal trajectory (Score: 9.5) that best matches the prompt.
Figure 5: VLM’s scoring mechanism. A VLM-driven evaluation framework (ranging from -10 to +10) that balances prompt adherence (50%), relational logic (30%), and visual integrity (20%). It employs a granular penalty-bonus system to enforce semantic precision and reward aesthetic quality. Full prompt is in Supplyment 0.A .
Figure 6: Illustration of the pipeline of AFS . Given an intermediate preview x^0 , the VLM diagnoses the defect and generates a spatial mask M . We formulate a contrastive energy function and project PRS gradient back to the velocity field vt . This time-scaled gradient modulation ensures optimal trajectory steering toward the target concept while strictly confining the intervention within the masked region.
Method
Attribute Binding
Object Relationship
Complex ↑
Average ↑
Time (s) ↓
color ↑
shape ↑
Texture ↑
Spatial ↑
Non-Spatial ↑
General T2I Models
DALL-E 2
0.5750
0.5464
0.6374
0.1283
0.3043
0.3696
0.4268
-
SDXL
0.6369
0.5408
0.5637
0.2032
0.3179
0.4091
0.4453
4.6
PixArt- α
0.6886
0.5582
0.7044
0.2082
0.3179
0.4117
0.4815
6.0
FLUX
0.7736
0.5112
0.6325
0.2747
0.3077
0.3622
0.4770
11.7
Table 1: Quantitative comparison on T2I-CompBench. Red indicates the best performance, Blue indicates the second best, and “-” indicates close-sourced.
Figure 7: Visual Comparison of T2I-CompBench experiment. We conducted tests from three dimensions: attribute binding, spatial relationship, and complexity, and obtained good results in all, demonstrating the validity of our method.
Figure 8: Results on GenEval. We conducted tests from Objects, counting, color, position and overall, and obtained good results in all, proving the validity of our method.
Figure 9: Visual Comparison of R2I-Bench experiment. We further try to explore the reasoning capabilities of our method. Our agentic framework helps original FLUX.1-dev think and react across the whole pipeline.
Method
Causal
Logical
Commonsense
Compositional
Mathematical
Average
Lumina-Image 2.0
0.40
0.56
0.49
0.65
0.13
0.45
Sana-1.5
0.21
0.49
0.49
0.67
0.13
0.40
Lumina-T2I
0.18
0.38
0.38
0.49
0.13
0.31
Omnigen
0.34
0.51
0.43
0.60
0.18
0.41
EMU3
0.41
0.61
0.44
0.62
0.09
0.43
Janus-Pro-7B
0.36
0.46
0.45
0.64
0.07
0.40
Table 2: Quantitative comparison on the R2I-Bench benchmark. Bold indicates the best performance.
Method
Attribute Binding
Object Relationship
Complex ↑
Average ↑
color ↑
shape ↑
Texture ↑
Spatial ↑
Non-Spatial ↑
AFS-Search (ours)
0.8847
0.6292
0.7609
0.6250
0.5305
0.6185
0.6748
FLUX
0.7736
0.5112
0.6325
0.2747
0.3077
0.3622
0.4770
w/o Optimization
0.8251
0.5614
0.6843
0.3654
0.3521
0.5011
0.5482
w/o Parallel Rollout Search
0.7932
0.5421
0.6587
0.3315
0.3359
0.4132
0.5124
w/o Self-Correction
0.8012
0.5631
0.6496
0.4031
0.3468
0.4463
0.5350
Table 3: Ablation study on T2I-CompBench verifying the effectiveness of each component, proving that our core components improve the performance of FLUX.1-dev.
Figure 10: Ablation Study of the Search Strategy. (a) Search Timing: Shows that searching at multiple stages is more effective than searching at any single time point. (b) Simulation Steps: Demonstrates that looking ahead with more simulation steps is essential for selecting the best path and avoiding errors.
Appendix figures & tables3 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 11: Additional Visual Results.
Figure 12: Conceptual Illustration of Vector Field Warping . (Left) Standard Flow: The trajectory follows a straight line but ends in a region with incorrect attributes (e.g., wrong color). (Right) AFS-Steered Flow: The velocity field is locally warped by the semantic gradient. The trajectory is bent toward the target basin while maintaining the overall linear transport properties of the flow.
Figure 13: Extra experiments on hyperparameters . The success rate is calculated by Qwen-vl-max, and the scoring mechanism is provided in above prompt.
Text-to-image generation creates images from natural-language prompts using generative models, with research focused on fidelity, control, efficiency, and safety.
Despite rapid advances in text-to-image generation, faithfully realizing user intent remains challenging, often requiring manual multi-turn trial and error. To automate this process, existing systems rely on either simple prompt rewriting or closed-loop agents driven by hand-crafted rules, rather than learning to adapt actions to the evolving generation process. In this paper, we reformulate image generation as a state-conditioned action-making problem and propose Generation Navigator, a multi-turn T2I agent that learns to dynamically steer the generation trajectory and output the next action. However, training this agent via reinforcement learning introduces a critical credit assignment challenge: naively rewarding a trajectory based solely on a single state assigns equal credit to all actions in the rollout, ignores the quality dynamics across turns, and fails to distinguish actions that improve the trajectory from those that degrade it or waste turns without progress. We resolve this with PRE-GRPO (Peak-Retention-Efficiency Group Relative Policy Optimization), a trajectory-level reinforcement learning objective that explicitly rewards discovering a high-quality image (Peak), avoiding subsequent quality degradation across turns (Retention), and minimizing unnecessary turns (Efficiency). Experiments show substantial improvements across benchmarks, reaching a WISE score of 0.90 and 79.06% reasoning accuracy on T2I-ReasonBench.
Jinming Liu, Ruoyu Feng, Yuqi Wang +2
Shanghai Jiao Tong University · Independent · Eastern Institute of Technology, Ningbo, China
Reinforcement learning (RL) enables language models to improve text-to-image (T2I) generation by rewriting prompts for a frozen image generator. Existing methods typically apply a single signal derived from the image to individual subword tokens, although objects, counts, attributes, and relations form coherent semantic spans. We propose SCOPE, an objective based on Group Relative Policy Optimization (GRPO) that aligns update weighting with this structure. The rewriter jointly generates an enhanced prompt and its semantic partition. SCOPE pools shifts in policy log probabilities within each span to derive a shared coefficient for its token losses while retaining a single advantage for each completion. It requires neither separate span rewards nor additional image renders. Across Qwen-Image, SD3, and FLUX.1-dev, SCOPE achieves average relative gains over vanilla GRPO of 9.48% on GeneVal and 7.62% on T2I-CompBench. Ablation studies show that both semantic boundaries and pooling within spans contribute to these gains, supporting semantic structure as a practical inductive bias for prompt engineering with RL.
Yufei Jiao, Jinfa Huang, Zeyue Xue +5
Xiamen University · University of Rochester · The University of Hong Kong +1
Reinforcement learning (RL) is promising for spatial text-to-image generation, but existing work largely focuses on policy optimization, assuming accurate rewards and effective online rollouts. We observe in-domain gains with out-of-domain degradation, exposing two limitations: unreliable spatial rewards in open-world settings and rollouts lacking basic-to-complex progression. We propose a curriculum-guided RL framework. First, we introduce SpatialCritic, an agentic reward model equipped with a spatial perception harness that combines reusable verification guidance with specialized vision tools. We then construct spatial preference datasets through targeted counterfactual perturbations and perform agent-harness alignment to improve tool selection and evidence integration efficiently. Second, curriculum learning organizes shared foundational capabilities and their compositional dependencies in a capability graph for mastery-guided rollouts, periodic review, and targeted remediation. On SpaRW-Eval, SpatialCritic achieves 95.1% verification accuracy, outperforming 320B GLM-5.3-Flash with a 40x smaller 8B backbone and raising average baseline scores from 78.4 to 83.9 on SD3.5-Medium and 83.2 to 86.5 on BAGEL. Project page: https://xiaofeng-tan.github.io/projects/spatialcritic/