AdaPilot: Towards Scene-Adaptive Policy Learning for Cross-Generator Text-to-Image Quality Optimization

Wenjin Liu; Fayuan Ke; Yue Lu; Zhe Cui; Anh Tuan Luu; Haoran Luo

cs.CVAug 25, 2026

AdaPilot: Towards Scene-Adaptive Policy Learning for Cross-Generator Text-to-Image Quality Optimization

Authors: Wenjin Liu, Fayuan Ke, Yue Lu, Zhe Cui, Anh Tuan Luu, Haoran Luo

Organizations: Nanyang Technological University, Singapore · Hithink Research, China

Abstract

Existing methods for improving text-to-image generation quality have progressed from generator fine-tuning and prompt optimization to reinforcement learning with multi-turn visual feedback. However, existing strategies are deeply coupled with specific generators and tasks, and the learned capabilities are difficult to generalize into a universal quality optimization policy. Therefore, we propose AdaPilot, which learns a scene-adaptive, cross-generator transferable quality optimization policy by formulating multi-turn image generation as a Markov Decision Process (MDP) and optimizing it via end-to-end reinforcement learning. Specifically, AdaPilot decouples the policy from generator internals to enable cross-generator transfer, introduces scene-aware rewards that adaptively align quality assessment dimensions with task semantics, and employs process-level rewards to model the evolution trajectory of image quality. Experimental results show AdaPilot outperforms baselines in generation quality and generalization. Separate cross-generator evaluations further show that a single policy transfers zero-shot to unseen generators while maintaining positive average gains across all evaluated generators. Our project is available at https://github.com/QwenQing/Ada_pilot.

Figures & tables

Figure 1: An illustration of the proposed AdaPilot.

Figure 2: Qualitative comparison with and without AdaPilot agent on Qwen-Image, GPT-Image-1, and Gemini-3-pro-Image-Preview. AdaPilot enhances the image generation quality across different generators without updating their parameters.

Figure 3: Overview of AdaPilot, a policy for scene-adaptive image quality optimization through multi-turn visual feedback.

Figure 4: The system prompts for the AdaPilot agent. The agent optimizes the text prompt in Phase 1 and assesses the generated image against the original description, iteratively refining the prompt in Phase 2 until the scene-specific criteria are fully met.

Dataset	Metric	Baseline	CoT	GRPO	SFT	BoN	F-GRPO	T2I-R1	Copilot	MPrompt	Ours ( $\Delta\uparrow$ )
Fine-T2I	GDino	80.63	74.47	74.41	73.58	80.15	77.95	80.57	75.86	73.60	84.25 (+3.62)
	Realism	41.47	40.43	40.04	38.48	42.58	30.43	24.41	41.41	40.25	46.46 (+4.99)
	HPS	71.38	74.85	71.62	76.59	71.57	17.26	68.48	70.08	66.38	77.90 (+6.52)
	Physics	38.10	35.63	35.24	35.16	36.42	26.84	26.56	37.10	30.17	45.83 (+7.73)
AutoPP1M	CLIP-T	95.84	96.74	96.98	96.27	97.21	87.54	97.06	94.37	96.74	97.31 (+1.47)
	GDino	90.39	87.21	84.11	85.62	86.73	82.03	83.74	88.17	87.47	91.75 (+1.36)

Table 1: Comparison of the baselines and AdaPilot on six in-distribution datasets covering diverse tasks and domains. F-GRPO denotes Flow-GRPO, and Copilot denotes T2I-Copilot.

\Delta\uparrow

means the gap between AdaPilot and the baseline image generator (Qwen-Image), where higher values indicate better performance. Bold values mean the best performance. All values are in %.

Dataset	Metric	Baseline	CoT	GRPO	SFT	BoN	F-GRPO	T2I-R1	Copilot	MPrompt	Ours ( $\Delta\uparrow$ )
DPG-Bench	CLIP-T	94.89	94.43	92.92	94.92	94.54	81.93	92.95	95.24	94.71	95.72 (+0.83)
	GDino	82.68	79.42	76.11	78.37	79.46	67.47	77.42	84.35	83.10	86.67 (+3.99)
	Realism	44.40	45.51	41.60	44.34	44.73	28.08	26.76	42.38	36.46	52.36 (+7.96)
	Physics	42.21	41.94	37.40	38.29	41.13	23.81	26.40	43.76	22.89	51.40 (+9.19)
TVBench	OCR	39.14	35.19	42.14	43.32	39.87	4.61	0.09	37.76	42.64	44.60 (+5.46)
	Aesthetic	15.82	6.25	8.98	6.25	13.67	9.38	5.47	12.65	7.59	24.02 (+8.20)

Table 2: Comparison of the baselines and the proposed AdaPilot on four out-of-distribution datasets.

\Delta\uparrow

is AdaPilot’s gain over Qwen-Image, and bold marks the best result. All values are in %.

Figure 5: Averages over applicable evaluation metrics for four unseen image generators on four out-of-distribution datasets and overall, comparing results with and without AdaPilot to evaluate zero-shot cross-generator transfer.

Method	TextAtlas5M		AutoPP1M		DPG-Bench		CompBench
	OCR $\uparrow$	Aesthetic $\uparrow$	HPS $\uparrow$	Realism $\uparrow$	HPS $\uparrow$	Realism $\uparrow$	HPS $\uparrow$	CLIP-T $\uparrow$
w/o Agent	33.03	31.64	52.04	43.95	55.98	44.40	65.90	82.50
w/o RL	23.09	30.17	55.15	40.45	60.72	39.14	68.31	82.59
w/o Scene-Adaptive Reward	48.30	37.89	52.83	46.88	57.86	45.90	67.81	85.78
w/o Monotonic Reward	34.30	35.24	55.90	48.24	58.23	47.63	70.06	84.19
AdaPilot (Full)	58.35	48.63	57.37	54.49	60.77	52.36	72.06	86.67

Table 3: Ablation study. w/o RL is without reinforcement learning; w/o Agent is without the AdaPilot Agent; w/o Scene-Adaptive Reward is without Scene-Adaptive Quality Assessment; w/o Monotonic Reward is without the Monotonic Improvement term.

Figure 6: Average values of unseen generators on in-distribution datasets with and without AdaPilot agent.

Figure 7: Seven-metric averages for three of five rounds. Stars denote pre-feedback Round 0 and dashed lines denote BoN.

Appendix figures & tables5 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 8: An Example of AdaPilot’s Multi-turn Visual-Feedback Generation Trajectory.

Figure 9: Overview of the seven evaluation metrics employed in our proposed AdaPilot, including the evaluation model and prompt for each metric. CLIP-T, HPS, GDino, Aesthetic, and OCR are utilized in the training reward, while Realism and Physics serve as the held-out metrics.

Method	Paradigm	Agent / Model	Generator	Learning Rate	Max Steps	Epochs
Baseline	-	-	Qwen-Image-2512	-	1	-
Chain-of-Thought	Frozen	Qwen2.5-VL-7B	Qwen-Image-2512	-	1	-
Supervised Fine-Tuning	Frozen	Qwen2.5-VL-7B	Qwen-Image-2512	$1\times 10^{-4}$	1	2
Group Relative Policy Optimization	Frozen	Qwen2.5-VL-7B	Qwen-Image-2512	$1\times 10^{-6}$	1	1
Best-of-N ( $N$ =5)	Frozen	-	Qwen-Image-2512	-	1	-
Flow-GRPO	Tuning	-	Stable Diffusion 3.5 Medium	$1\times 10^{-4}$	1	-

Appendix

Table 4: Hyperparameter settings for AdaPilot and baselines.

Figure 10: Case studies of the baselines and AdaPilot, which include Qwen-Image (baseline generator), CoT, SFT, GRPO, BoN, Flow-GRPO, T2I-R1, T2I-Copilot, MinorityPrompt, and AdaPilot (ours).

Figure 11: Failure case of AdaPilot on a dense sports infographic. Iterative refinement improves layout organization, but the text and charts remain unreliable, and the agent terminates despite identifying residual defects.

T2I Generation

Text-to-image generation creates images from natural-language prompts using generative models, with research focused on fidelity, control, efficiency, and safety.

See all 299 papers

APT: Feedback-Driven Agentic Post-Training for Text-to-Image Generation
Oct 13, 2026Sirui Zhang, Fangyun Wei, Dong Chen +5T2I Generation RL for Image Generation
University of Science and Technology of China · Microsoft Research Asia · University of Sydney +2
STAR: SpatioTemporal Adaptive Reward Allocation for Text-to-Image RL Post-Training
Jun 16, 2026Jinjie Shen, Wei Deng, Xian Hu +2Diffusion Model Alignment Compositional T2I Generation
Generation Navigator: A State-Aware Agentic Framework for Image Generation
May 18, 2026Jinming Liu, Ruoyu Feng, Yuqi Wang +2T2I Generation Agentic Image Generation
Shanghai Jiao Tong University · Independent · Eastern Institute of Technology, Ningbo, China

Oct 13, 2026J/K move · Enter open · S save

APT: Feedback-Driven Agentic Post-Training for Text-to-Image Generation

Sirui Zhang, Fangyun Wei, Dong Chen, Jinjing Zhao+4

University of Science and Technology of China · Microsoft Research Asia · University of Sydney+2

T2I Generation RL for Image Generation

Open paper AskPDF arXiv

Post-training is crucial for improving text-to-image models, yet existing pipelines typically rely on manually curated training data and fixed training distributions. We introduce APT (Agentic Post-Training), a fully automated framework that adaptively post-trains a base model according to its evolving capability profile. APT coordinates three agents to construct evaluation sets from user requirements, diagnose model-specific strengths and failure modes, and generate reinforcement learning data with an adaptive prompt distribution that emphasizes underperforming capabilities. By repeatedly evaluating, diagnosing, and updating the generator, APT lets the current model's behavior shape its next training tasks, forming a feedback loop inspired by recursive self-improvement. For general-purpose settings without user-specified evaluation scenarios, we further develop an evaluation harness that enables broad and diverse evaluation set generation. Experiments across multiple base models, benchmarks, and generation architectures demonstrate that APT is model-agnostic, benchmark-agnostic, architecture-agnostic, and fully automated.

Abstract

T2I Generation

Explore similar work

APT: Feedback-Driven Agentic Post-Training for Text-to-Image Generation

STAR: SpatioTemporal Adaptive Reward Allocation for Text-to-Image RL Post-Training

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