Text-to-image models have made remarkable progress in visual fidelity and instruction following, yet complex compositional generation remains challenging. Existing agentic approaches address this limitation through iterative generation and editing, adapting their actions based on interaction history. However, they typically fix the continuation state (i.e., the image from which refinement proceeds) to the latest image. Since iterative editing is not necessarily monotonic, the agent may continue refining from the latest image even when it has already degraded. In this work, we introduce GenRevisit, an agentic framework that extends action selection to joint continuation-state and action selection. The agent can not only use interaction history to adjust editing instructions but also select an image candidate to edit, jointly deciding where to continue from and how to refine. We train GenRevisit with SFT followed by agentic reinforcement learning. To address the coarse credit assignment of final outcome reward in GRPO, which assigns the same trajectory-level advantage to all actions, we further propose Normalized Action Credit Assignment (NACA) to redistribute positive advantage according to each action's contribution to reducing the remaining objective. On GenEval2, GenRevisit achieves GM scores of 84.3 and 74.4 with Qwen-Image-2512 and Z-Image-Turbo, surpassing the strongest baselines using the same generators by 12.7 and 10.9 points respectively. GenRevisit achieves state-of-the-art performance on GenEval, GenEval++, and GenEval2 among baselines.
Figures & tables
Figure 1: An illustrative comparison between latest-state refinement and our GenRevisit.
Figure 2: Overview of GenRevisit. Panel (a) presents the overall workflow. At each decision step, the multimodal planner selects the next action based on the decision context and, when editing, jointly selects a source image as the continuation state and a corresponding edit specification. The resulting image is verified and recorded in the interaction history to inform subsequent decisions, while submission returns the historical GM-best image. Panel (b) illustrates an editing example. The planner selects I1 , which is neither the latest candidate I2 nor the GM-best candidate I0 , and corrects the backpack count while preserving the other constraints. The resulting candidate I3 passes all constraint-level checks and becomes the new GM-best image.
Table 1: Main results on GenEval2. GM averages prompt-level geometric means of atom scores; Skill Avg. averages the five skill-wise AM scores. Bold and underlined scores mark the best and second-best results among (i) all direct-generation models and Qwen-Image-2512 workflows, or (ii) Z-Image-Turbo workflows alone. † marks the best external agentic baseline within each generator block; ↑ / ↓ shows absolute score-point changes relative to it.
Method
Agent
Generator
Overall ↑
Direct generation
FLUX.1-dev [ 41 , 13 ]
–
FLUX.1-dev
67.0
Qwen-Image [ 34 , 13 ]
–
Qwen-Image
87.0
Qwen-Image-2512 † [ 9 ]
–
Qwen-Image-2512
74.0
Z-Image-Turbo † [ 9 ]
–
Z-Image-Turbo
82.0
Agentic generation and refinement
Table 2: Cross-benchmark evaluation on GenEval and GenEval++.
Figure 3: Qualitative comparison of GenRevisit and baseline models.
Variant
Overall
Skill-wise AM
GM
Δ GM
Skill Avg.
Object
Attribute
Position
Count
Verb
Qwen-Image-2512
29.0
–
62.0
97.9
65.2
58.1
66.9
22.1
Untuned Qwen3-VL-8B-Instruct †
49.7
+20.7
74.6
97.9
88.9
64.4
72.4
49.6
SFT only
81.5
+31.8
85.3
98.9
96.1
91.7
88.8
51.1
SFT + GRPO
80.3
−1.2
85.1
98.7
96.0
91.1
88.4
51.4
SFT + NACA
84.3
+4.0
86.7
99.4
97.2
92.5
90.3
53.9
Table 3: Ablation study on GenEval2.
Appendix figures & tables24 assets
Supplementary material from the paper’s appendix.
Appendix
Role
Model or component
Use
Planner
Qwen3-VL-8B-Instruct
Language parameters are updated; vision tower and multimodal projector are frozen.
Teacher
GPT-5.5
Collects multi-round demonstrations with executed actions and verifier feedback.
Generator
Qwen-Image-2512
Produces a new candidate without an editing input.
Editor
Qwen-Image-Edit-2511
Edits the candidate named in the action.
Verifier
Qwen3-VL-8B-Instruct
Frozen VQA evaluator; provides atomic feedback and Soft-TIFA scores.
Guidance
Fixed skill library
Returns requested guidance; does not choose the next action or image.
Appendix
Table A.1: Model roles and trainable components. Backend routing is controlled by the environment, not predicted in an action.
Component
Information
Task
Original prompt, fixed atomic requirements and questions, maximum image attempts.
Candidate history
Every completed candidate, its bound image, parent identifier, GM, constraint results, and the action that produced it.
Control
Current rankings, visible and editable candidate identifiers, legal submissions, and remaining image budget.
Retrieved guidance
Actual returned skill content and its associated target constraints.
Query control
Available skills, completed and pending queries, and remaining query allowance.
Action error, when present
The last unresolved format or validation error, used to request a new complete action.
Appendix
Table A.2: Decision-time information in the Joint input. All entries are environment-provided inputs.
Interface
Visible candidates
Editable candidates
Submission
Joint
All completed images
All completed images
GM-best
Best-only
All completed images
Current GM-best
GM-best
Latest-only
All completed images
Current latest
GM-best
Appendix
Table A.3: Candidate permissions under the GM-control interface. Rankings and permissions are recomputed after each completed image. This table defines the interfaces, not additional trained variants.
Setting
Value
Planner initialization
Qwen3-VL-8B-Instruct
Prompt membership
800 train / 100 validation / 100 test
Teacher prompt pool
1,000 prompts across the three splits
Learning rate / global batch size / epochs
10−5 / 32 / 2
Selected action targets
3,473 train / 437 validation / 413 test
Targets per exported record
One canonical assistant action
Appendix
Table A.4: SFT settings and supervision interface, aligned with the main-paper configuration.
Setting
Value
Training protocol
Initialization
Joint selective SFT checkpoint
Rounds / prompt groups per round
5 / 64
Rollouts per group
4
Collected trajectories
256 per round; 1,280 total
Optimizer steps
8 per round; 40 total
Appendix
Table A.5: RL settings for the five-round training chain. Hardware is specified only for the final round where it was confirmed.
Action
Selection rule
Initial generation
Retain the required first generation.
Editing
Retain when gchild>gedit+εSFT .
Subsequent generation
Retain when gchild>garchive+εSFT .
Submission
Retain the actually executed submission when it selects the current GM-best and uses a valid reason.
Guidance retrieval
Require a real returned response. The first subsequent executed image action must itself be selected, and its target constraints must intersect the query’s target constraints.
Appendix
Table B.1: Selection rules for positive SFT targets. All comparisons use the record before the candidate action and its actual execution outcome.
Split
Generate
Edit
Retrieve
Submit
Total
Train
961
1,076
584
852
3,473
Validation
108
140
103
86
437
Test
110
114
99
90
413
Total
1,179
1,330
786
1,028
4,323
Appendix
Table B.2: Selected targets in the current Joint SFT export. Counts are action responses, not distinct prompts or raw trajectories.
Case
Handling
Normal submission
Reward is the utility of the submitted image.
Policy-invalid trajectory
Reward is −1 and participates in group statistics. Retained sampled responses are trained if their group is not filtered.
Image backend, transport, or verifier failure
Retry or recover; do not substitute a score or treat the failure as a policy violation.
Image budget exhausted, followed by valid submission
Score the submitted image normally; budget exhaustion alone adds no penalty.
No valid submission
Use the recorded termination class to distinguish policy invalidity from incomplete execution or infrastructure failure; do not invent a submission.
Zero-variance group
Remove the group from optimizer data, including KL.
Appendix
Table C.1: Termination and failure handling during RL data preparation.
Image action in a nonpositive-advantage trajectory
Ai
Retained
Retrieval, submission, or retained invalid response
Ai
Retained
Context, tool results, and verifier feedback
Not a target
Excluded
Appendix
Table C.2: Credit and KL for responses in retained optimizer data. The first two rows assume effective positive progress is available.
GenEval
Method
Single obj. ↑
Two obj. ↑
Counting ↑
Colors ↑
Position ↑
Color attr. ↑
Overall ↑
Direct generation
FLUX.1-dev [ 41 ]
0.9900
0.8100
0.7500
0.8000
0.2100
0.4800
0.6700
Qwen-Image [ 34 ]
0.9900
0.9200
0.8900
0.8800
0.7600
0.7700
0.8700
Agentic generation and refinement
T2I-Copilot † [ 2 , 14 ]
0.9900
0.9100
0.6800
0.8600
0.5500
0.4600
0.7400
Appendix
Table E.1: Cross-benchmark generalization. Bold scores indicate the best result in each column.
Method
Image backend
Category-wise accuracy
Overall
Color
Count
Color/ Count
Color/ Pos
Pos/ Count
Pos/ Size
Multi- Count
Direct generation
FLUX.1-dev [ 13 ]
FLUX.1-dev
0.4000
0.6000
0.2500
0.2500
0.0750
0.4000
0.3000
0.3250
BAGEL [ 38 ]
BAGEL
0.3250
0.6000
0.2500
0.3250
0.2500
0.4750
0.3750
0.3710
Qwen-Image [ 13 ]
Qwen-Image
0.8750
0.7250
0.7250
0.6000
0.4750
0.7250
0.5500
0.6680
Echo-4o [ 38 ]
BAGEL
0.8000
0.5750
0.5500
0.7750
0.6250
0.8000
0.6250
0.6790
Appendix
Table E.2: Comparison on GenEval++. The seven categories follow the benchmark taxonomy. Higher is better; bold marks the highest displayed result in each column.
Figure F.1: Joint and Best-only from a shared prefix. Source labels identify the actual editing inputs. Joint’s final parent is a_002 , not J:a_003 ; Best-only submits shared a_001 , not B:a_004 . Suffix identifiers are branch-local. The final Joint instruction appears in Appendix G.2 .
Figure F.2: An archived historical-access versus latest-only pair. Both branches use two additional calls. Full-history has parent relations a001 → H:a003 and a001 → H:a004 ; Latest-only follows a002 → L:a003 → L:a004 . The detail panels reproduce the same-region crops from the archived figure and refer to the last generated images, not necessarily to the submissions. Dotted lines denote detail views, not edits.
Figure F.3: A successful continuation from below the current best. The only executed edge shown is a001 → a004 . a000 is a comparison candidate, not an intermediate image in that edit. Discrete labels report the saved verifier outcomes. This is one observed trajectory rather than a paired continuation from both possible starting images.
Image action
Pass; GM
Li,t
qi,t
wi,t
GRPO
NACA
Generate
5/6; 62.03
218
.824812
1.898143
1.730187
3.284142
a000→a001
5/6; 82.54
342
.046820
.107747
1.730187
.186422
a001→a002
5/6; 54.16
376
0
0
1.730187
0
a001→a003
6/6; 95.64
380
.989549
2.277251
1.730187
3.940071
Appendix
Table F.1: Recorded image-action credit. GM is on a 0–100 scale. The printed coefficients are rounded; token-weighted conservation applies to the stored full-precision values.
Figure F.4: NACA changes action credit within a single positive-advantage rollout. Solid bars show NACA action advantages; the dashed line is the common GRPO broadcast value. Each bar represents an image action, not a token count. The zero bar removes positive policy-gradient credit, not the response’s KL term. Retrieval and submission are not shown and keep their trajectory advantage.
Figure F.5: Selected outputs on designed prompts. Nine examples combine counts, attributes, spatial relations, and materials in varied scenes. The prompt key summarizes each request. Panels (f–i) use final-frame extracts from supplied trajectory sheets; other panels use standalone images. Tile sizes organize the display and do not encode scores.
Figure F.6: Selected GenEval2 outputs from the SFT policy. Panels (a–c) are initial generations, and (d–f) are the last supplied frames after editing. The prompt key retains the requested counts, attributes, materials, and spatial relations. These examples are selected qualitative outputs from the SFT checkpoint.
Figure F.7: Selected GenEval++ outputs. Nine images span the benchmark’s seven compositional categories. These are selected qualitative examples from GenRevisit. Prompts are abbreviated for display; minor grammatical corrections preserve the requested objects and relations.
Figure F.8: Executed revisits on a designed prompt. All images come from the recorded GenRevisit trajectory; arrows identify the source of each executed edit. Intermediate errors remain visible alongside the submitted result.
Aspect
Teacher collection
Joint student / SFT
Input packaging
Task, planner context, candidate observations, GM continuation control
Dedicated structured decision-time input
Candidate score key
primary_score
gm
Control key
gm_continuation_control
control
Editing instruction
Four semantic parts requested in the collection guidance
One instruction string; no extra four-part schema
Candidate access
All completed images available for joint editing
All completed images visible and editable
Output
One structured action
One structured action
Appendix
Table G.1: Teacher and student input contracts. Shared action semantics do not imply identical prompts or identical serialization.
Table G.2: Task-relevant arguments of the four actions. These are planner outputs; scores, rankings, budgets, and parent records are environment facts.
Skill
Focus
Counting and instance layout
Count and instance layout
Spatial relations
Spatial relations
Attribute binding
Attribute-to-entity binding
Local edit preservation
Local edits and preservation
Action and pose relations
Action, pose, and relation errors
Object identity and presence
Object identity and presence
Appendix
Table G.3: Documented skill catalog. Each entry supplies guidance rather than a new action type.
Constraint
Fixed verification question
Status
Probability
c_001
How many flamingos are in the image?
Pass
0.999979
c_002
Are the flamingos green?
Pass
0.999994
c_003
Are the flamingos yellow?
Pass
0.808171
c_004
Are there any flamingos in the image?
Pass
0.999774
Appendix
Table G.4: Recorded verifier feedback on the submitted Joint flamingo image. Probabilities are rounded only for display.
Compositional image generation studies how to combine objects, attributes, spatial relations, and visual references into coherent, controllable images.
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
Recent text-to-image (T2I) generators can synthesize realistic images, but still struggle with compositional prompts involving multiple objects, counts, attributes, and relations. We introduce EPIC (Efficient Predicate-Guided Inference-Time Control), a training-free inference-time refinement framework for compositional T2I generation. EPIC casts refinement as predicate-guided search: it parses the original prompt once into a fixed visual program of object variables and typed predicates, covering checkable conditions such as object presence, counts, attributes, and relations. Each generated or edited image is verified against this program using visual evidence extracted from that image. An image is judged to satisfy the prompt only when all predicates are satisfied; otherwise, failed predicates decide the next step, routing local failures to targeted editing and global failures to resampling while the fixed visual program remains unchanged. On GenEval2, EPIC improves prompt-level accuracy from 34.16% for single-pass generation with the base generator to 71.46%. Under the same generator/editor setting and maximum image-model execution budget, EPIC outperforms the strongest prior refinement baseline by 19.23 points while reducing realized cost by 31% in image-model executions, 72% in MLLM calls, and 81% in MLLM tokens per prompt.
Sunung Mun, Sunghyun Cho, Jungseul Ok
Graduate School of Artificial Intelligence, POSTECH · Department of Computer Science & Engineering, POSTECH
Text-to-image generation has advanced rapidly, but existing models still struggle with faithfully composing multiple objects and preserving their attributes in complex scenes. We propose coDrawAgents, an interactive multi-agent dialogue framework with four specialized agents: Interpreter, Planner, Checker, and Painter that collaborate to improve compositional generation. The Interpreter adaptively decides between a direct text-to-image pathway and a layout-aware multi-agent process. In the layout-aware mode, it parses the prompt into attribute-rich object descriptors, ranks them by semantic salience, and groups objects with the same semantic priority level for joint generation. Guided by the Interpreter, the Planner adopts a divide-and-conquer strategy, incrementally proposing layouts for objects with the same semantic priority level while grounding decisions in the evolving visual context of the canvas. The Checker introduces an explicit error-correction mechanism by validating spatial consistency and attribute alignment, and refining layouts before they are rendered. Finally, the Painter synthesizes the image step by step, incorporating newly planned objects into the canvas to provide richer context for subsequent iterations. Together, these agents address three key challenges: reducing layout complexity, grounding planning in visual context, and enabling explicit error correction. Extensive experiments on benchmarks GenEval and DPG-Bench demonstrate that coDrawAgents substantially improves text-image alignment, spatial accuracy, and attribute binding compared to existing methods.
Chunhan Li, Qifeng Wu, Jia-Hui Pan +7
Lingnan University. · China University of Petroleum · CMU +4