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

1 Introduction

• Existing methods face limited cross-generator transfer, fixed reward signals that miss scene-specific priorities, and weak modeling of quality changes between refinement turns.The introduction describes existing approaches as generator- and scene-specific, limiting generalization across heterogeneous tasks.Source: §1 · p. 1
• AdaPilot formulates multi-turn generation as a visual-feedback MDP and uses visually encoded observations to support policy transfer across generators.The policy interacts through observations rather than relying on generator-specific internals.Source: §1 · p. 1
• Its scene-aware reward activates and renormalizes quality dimensions according to task semantics, while a process-level reward combines terminal quality with turn-wise monotonic improvement.Training freezes the generator and optimizes only the policy end-to-end.Source: §1 · p. 2
• AdaPilot consistently outperforms baselines across ten datasets and multiple frozen generators, including on held-out quality dimensions, while transferring without adaptation to unseen generators.The reported transfer evaluation shows consistent gains across all evaluated generators.Source: §1 · p. 2, Figure 2 · p. 2

Generator-Tuning Approaches

• Generator-tuning approaches optimize diffusion generators using policy gradients, reward backpropagation, online reinforcement learning, or preference and concept supervision.Examples include Flow-GRPO and DanceGRPO for online RL, and T2I-R1 for collaborative semantic- and token-level CoT.Source: §2 · p. 2

Generator-Frozen Approaches

• Generator-frozen approaches use prompt rewriting, iterative optimization, attention or spatial control, LLM planning, and output refinement; AdaPilot instead combines scene-conditioned multi-metric reward routing with turn-wise monotonicity supervision.The passage also describes GenAgent as using final-quality and reflection rewards for cross-tool generalization.Source: §2 · p. 2

3.1 Adaptive Multi-Turn Image Generation

• AdaPilot frames generation as an MDP in which an MLLM policy observes intermediate images, diagnoses defects, and iteratively refines prompts while interacting with a frozen generator.The state is accumulated context, and the agent output is the action at each turn.Source: §3.1 · p. 2

3.2 Scene-Aware Quality-Driven Optimization

• AdaReward combines format-gated terminal quality and monotonic improvement, balancing the final image score with a process-level trajectory score.Quality supervision is disabled when the interaction format is noncompliant or no image is generated.Source: §3.2 · p. 3
• Scene-adaptive assessment activates and renormalizes metric weights for each scene, using CLIP-T, HPS, Aesthetic, Grounding DINO, and OCR to assess distinct quality dimensions.The metrics correspond to text-image alignment, human preference, aesthetics, object accuracy, and text rendering, respectively.Source: §3.2 · pp. 3–4
• AdaReward rewards quality improvement at every generation step; any regression sets the monotonicity term to zero, while a fully improving trajectory receives its normalized gain.This process-level signal models quality evolution rather than assessing only the terminal image.Source: §3.2 · p. 4

3.3 Cross-Generator Policy Generalization

• AdaPilot accesses generators through prompts and encoded visual observations rather than internals or gradients, allowing policy reuse without retraining; at inference, an unseen generator replaces the training generator.The visual encoder maps outputs from different generators into a shared observation space, and the policy is reused without adaptation.Source: §3.3 · p. 4

• Table 1 compares baselines and AdaPilot across six in-distribution datasets covering diverse tasks and domains.Source: Table 1 · p. 5

4.1 Experimental Setup

• Evaluation uses ten datasets, nine baselines, and seven metrics, with Qwen2.5-VL-7B-Instruct trained using Qwen-Image-2512.Source: §4.1 · p. 5

4.2 Main Results (RQ1)

• AdaPilot outperforms all baselines across every evaluation dimension on six in-distribution datasets spanning diverse scene types.The largest improvements occur in text rendering and aesthetics; comparisons include single-turn, iterative, and multi-turn alternatives.Source: §4.2 · pp. 5–6

4.3 Out-of-Distribution Generalization (RQ2)

• On four OOD datasets, AdaPilot outperforms all baselines on every reported dimension, including held-out Realism and Physics.The reported tasks include dense-prompt, table, compositional, and instruction-following scenes.Source: §4.3 · p. 6, Table 2 · p. 6

4.4 Cross-Generator Transfer (RQ3)

• AdaPilot improves unseen generators’ in-distribution averages and their overall averages on unseen OOD data, with gains varying by generator and setting.Transfer is evaluated on FLUX.2, LongCat, Ovis, and Z-Image without target-specific retraining.Source: §4.4 · pp. 6–7, Figure 6 · p. 6, Figure 5 · p. 6

4.5 Ablation Study (RQ4)

• Removing reinforcement learning, scene-adaptive reward, or monotonic improvement lowers performance, with the full model leading every reported metric.The ablations report pronounced losses in text rendering, aesthetics, and realism when scene-adaptive or monotonic rewards are removed.Source: §4.5 · p. 7, Table 3 · p. 7

4.6 Turn-wise Quality Analysis (RQ5)

• Across three analyzed rounds, all datasets improve monotonically and finish above BoN; Lunara and TVBench recover from below-BoN first rounds.Text-rich and aesthetic-critical datasets show the largest turn-wise gains; results cover three rounds under a five-step cap.Source: §4.6 · p. 7, Figure 7 · p. 7

• AdaPilot combines scene-adaptive quality assessment, process-level evolution modeling, and shared-observation-space policy-generator decoupling for zero-shot transfer.Source: §5 · p. 7

• The optimal expected-quality supremum cannot decrease, since a multi-turn policy can terminate after its first image and reproduce a single-turn policy.Source: §A.1 · p. 10
• A feedback extension can strictly improve expected final image quality when an identifiable, correctable defect occurs with positive probability.The assumed gain is in expectation; individual stochastic generations need not improve strictly.Source: §A.1 · p. 10

A.2 Proof of Proposition 2

• Scene-adaptive supervision cannot reduce the shared-policy quality supremum and strictly improves it when fixed weights misalign strategy rankings with scene-specific requirements.The gap is zero when both scalarizations select the same supporting face.Source: §A.2 · p. 11

A.3 Proof of Proposition 3

• Zero-shot transfer preserves positive quality gain when training-generator gains exceed finite-horizon trajectory shift, with policy and generator parameters fixed.The condition extends to a collection of unseen generators when it holds uniformly.Source: §A.3 · p. 12

B Prompts of AdaPilot

• Phase 1 prompts the agent to generate from the description; after each image, Phase 2 elicits visual assessment and refinement or termination, subject to the turn budget.The latest valid image is returned at termination or when the budget is reached.Source: B Prompts of AdaPilot · p. 12

C AdaPilot Algorithm Details

• Each policy update samples complete trajectories with a frozen generator, computes rewards, then performs end-to-end policy optimization.Later turns sample one generation or termination action; invalid turn-specific outputs are retained for format-compliance evaluation.Source: C AdaPilot Algorithm Details · p. 12
• The generator and reward evaluators remain frozen, so no gradient passes through them, although their inference and evaluation costs remain.The update processes complete serialized multimodal trajectories, while the response mask limits loss aggregation to policy-generated positions.Source: C AdaPilot Algorithm Details · p. 13

D Dataset Details

• Ten public datasets cover compositional, dense-prompt, instruction-following, fine-grained, poster, text, table, product, financial-chart, and aesthetic generation tasks.Six datasets supply training prompts, with 640 instances sampled from each; testing uses 128 non-overlapping instances from each of the ten datasets, totaling 3,840 training and 1,280 test instances.Source: D Dataset Details · pp. 13–15
• The dataset suite includes benchmarks for dense prompt following, compositional generation, intricate instructions, poster design, and dense text rendering.TableVisBench assesses data accuracy, text rendering, relationships, and aesthetic quality; FinMME chart images are converted into textual prompts.Source: D Dataset Details · pp. 13–14

E Baseline Details

• Baselines span generator-tuning methods and frozen-generator methods, with each method evaluated under its specified generator and optimization paradigm.Frozen-generator examples include BoN reward-based selection and T2I-Copilot's three-agent prompt interpretation, model selection, and refinement.Source: E Baseline Details · p. 15

F Evaluation Metrics

• CLIP-T, HPS, GDino, Aesthetic, and OCR are training-reward metrics; Realism and Physics are held-out evaluation dimensions.Realism assesses photographic credibility, while Physics checks whether prompt-implied physical constraints are realized.Source: F Evaluation Metrics · pp. 15–16, Figure 9 · p. 16
• The training metrics assess text-image alignment, human preference, object detection, visual appeal, and rendered-text accuracy.GDino combines object-existence, spatial-relationship, and count scores; Aesthetic evaluates properties including clarity, style, color, and layout.Source: F Evaluation Metrics · p. 15
• OCR is scored by character-level F1 against prompt-specified text and is excluded when rendered text cannot be verified.The comparison uses edit-distance matches, insertions, deletions, and substitutions.Source: F Evaluation Metrics · p. 15

G Implementation Details

• Qwen2.5-VL-7B-Instruct is trained as the agent with Qwen-Image-2512 frozen; transfer is evaluated zero-shot on FLUX.2, LongCat, Ovis, and Z-Image.Methods are evaluated on the same test sets over three independent runs, with averaged results reported.Source: G Implementation Details · p. 16
• AdaReward activates different metric subsets by dataset and renormalizes active weights to sum to one; its mapping remains fixed during training.HQ-Poster uses all five training metrics, while other listed datasets use selected subsets.Source: G Implementation Details · p. 16

H Broader Impact

• The authors identify benefits in educational, scientific, and product scenes alongside risks involving forgery, deceptive marketing, impersonation, and bias amplification.They describe public-interface use, usage guidelines, provenance signals, and safety-oriented dimensions as mitigations or follow-up directions.Source: H Broader Impact · p. 17

I Case Study

• In the grooming case, AdaPilot better aligns comb teeth with hair strands and produces more natural object relations, texture, and scene coherence in close-up.The comparison says several baselines retain the setting but show weak comb-hair interaction, semantic shifts, or local distortions.Source: I Case Study · p. 17

J Failure Case Discussion

• The dense infographic remains unreliable because its source omits names and statistics, and the policy terminates despite recognizing unresolved text and chart defects.The discussion identifies generator text-rendering limits and the lack of explicit constraint verification as complementary limitations.Source: J Failure Case Discussion · p. 17

K Limitations

• Later rounds depend on accurate early quality assessments; an inaccurate gap analysis can steer refinements suboptimally and compound into degraded final quality.The interaction loop requires sequential image generation at each round.Source: K Limitations · p. 17
• A predefined scene-to-metric mapping may misalign rewards for novel scene types, while unfamiliar generator artifacts may weaken transfer and automated evaluators may diverge from human judgment.The shared visual encoder may fail to preserve task-relevant cues for artifacts outside its pretraining distribution.Source: K Limitations · p. 18

L Future Work

• Proposed remedies include self-correction for erroneous assessments and automatic discovery of quality dimensions to improve generalization to novel scene types.The latter would replace the predefined scene-metric mapping and reduce manual metric engineering.Source: L Future Work · p. 18
• Further directions are testing scene-adaptive policy learning on text-to-video and text-to-3D, and incorporating human preference feedback to better align optimization with perceptual quality.Feedback could be collected through online human evaluation or learned preference models.Source: L Future Work · p. 18

K Limitations

• In a dense sports infographic, refinement improves layout organization, but text and charts remain unreliable and the agent stops despite residual defects.This case illustrates the limits of the current interaction loop on a dense infographic task.Source: Figure 11 · p. 19