Look Back to Move Forward: Learning to Revisit for Agentic Compositional Image Generation

1 INTRODUCTION

• Latest-only refinement can force edits to start from a degraded image, spending limited budget restoring constraints already satisfied by an earlier candidate.The introduction motivates selecting a continuation state as well as an instruction.Source: §1 · p. 1
• GenRevisit makes continuation-state selection explicit, letting its planner choose a historical image and an edit instruction tailored to remaining errors while preserving satisfied constraints.The planner can revisit a stronger candidate and open new refinement branches instead of being restricted to the latest image.Source: §1 · p. 2
• SFT initializes the planner from multi-round trajectories, retaining failed actions and feedback in history while excluding failed actions from the SFT loss.RL with GRPO then improves the planner through its own interactions, using terminal reward based on final-image constraint satisfaction.Source: §1 · p. 2
• 84.3 GM on GenEval2 exceeded the strongest agentic baseline by 12.7 points; Z-Image-Turbo gained 10.9 without additional training, and GenEval/GenEval++ supported generalization.The introduction also reports state-of-the-art performance on GenEval, GenEval++, and GenEval among baselines.Source: §1 · p. 2
• In matched-budget teacher continuations, joint selection beat latest-only and best-only rules by 5.0 and 2.8 GM points; joint-demo training scored 81.5 GM, 4.0 above best-only training.NACA redistributes positive trajectory advantage across generation and editing actions according to the fraction of the remaining utility gap each action closes.Source: §1 · p. 2

2 RELATED WORK

• Compositional benchmarks move from predefined dimensions toward prompt-specific atomic verification, which GenRevisit uses as fine-grained feedback on individual constraints across attempts.The passage contrasts GenEval, T2I-CompBench, and ConceptMix [4, 13, 14] with TIFA, Davidsonian Scene Graph, and GenEval 2 [15] [16] [5].Source: §2 · p. 3
• Iterative generation progresses from rewriting and editing to explicit feedback loops and state-dependent action selection; GenRevisit additionally selects a historical source state for each operation.Historical candidates become selectable branching points for jointly choosing source state and subsequent operation.Source: §2 · p. 3

3 PRELIMINARIES

• The planner interacts with external tools to satisfy a prompt's compositional requirements within a total budget of image-generation and editing calls.The environment executes each action and updates the decision context with observations and feedback.Source: §3 · p. 3
• At step t, interaction state s_t is the decision context, including the prompt, available action/image/feedback history, and remaining budget; actions specify an operation and its parameters.The state is not an individual image candidate.Source: §3 · p. 3
• A trajectory is the sequence of interaction states and actions, and it terminates when an output image is selected or another prescribed condition is met.Generation and editing calls are bounded by budget B, which need not equal the number of decision steps T.Source: §3 · p. 3
• Historical image candidates can serve as continuation states distinct from the full interaction state; GenRevisit jointly selects a source and edit, then returns the historical GM-best image on submission.Figure 2 illustrates choosing I1 rather than latest candidate I2 or GM-best candidate I0, with the resulting I3 passing constraint checks.Source: §3 · p. 3, Figure 2 · p. 4

4.1 HISTORY-CONDITIONED INTERACTION

• For editing, the planner jointly selects a historical image as the continuation state and an instruction specifying which constraints to target and preserve.Its decision context includes atomic constraints, historical candidates, prior actions and verification feedback, plus guidance from previous skill queries.Source: §4.1 · pp. 4–5
• Each generated or edited image is verified against fixed constraints, added to history with its action and feedback, and can start a new branch without restoring budget.Generation and editing each consume one image call; submission returns the historical GM-best image.Source: §4.1 · p. 5

4.2 SUPERVISED FINE-TUNING FOR COLD START

• SFT targets are selected from teacher trajectories by action outcomes: improving edits receive imitation loss, while degraded attempts and feedback remain in context.The state conditions each target, and editing responses include both the image choice and edit specification.Source: §4.2 · p. 5

4.3 REINFORCEMENT LEARNING WITH NORMALIZED ACTION CREDIT

• RL uses terminal image utility that prioritizes passed constraints and uses GM to break ties; nonpositive advantages remain unchanged.Image utility combines passed-constraint count with Soft-TIFA GM, using α = 0.25.Source: §4.3 · pp. 5–6
• NACA redistributes positive advantage to effective-progress generation and editing actions according to the fraction of the remaining utility gap they close.Unlike standard GRPO's shared trajectory advantage, the normalized local-progress weights are designed to distinguish actions within positive-advantage trajectories.Source: §4.3 · p. 6

5.2 EVALUATIONS

• 84.3 GM with Qwen-Image-2512 exceeds the strongest same-generator agentic baseline by 12.7 points; Z-Image-Turbo reaches 74.4 GM, exceeding its baseline by 10.9 points.Qwen-Image-2512 also reaches 86.7 Skill Avg. and leads four of five skill categories; the Z-Image-Turbo policy is used without additional training.Source: Table 1 · p. 7, §5.2 · p. 7

5.1 SETUP

• Evaluation uses GenEval2 GM as its primary metric, also reports GenEval and GenEval++ overall scores, and tests Z-Image-Turbo transfer without additional training.The planner uses greedy decoding and at most five generation or editing calls per prompt; the final submitted image is evaluated.Source: §5.1 · p. 7

5.2 EVALUATIONS

• GenRevisit achieves the highest overall scores on GenEval (0.9844) and GenEval++ (0.8929), with counting scores of 0.9875 and 1.000, respectively.The results suggest the learned policy can guide compositional generation beyond GenEval2.Source: §5.2 · pp. 7–8, Table 2 · p. 8
• Qualitative comparisons show more consistent satisfaction of count, color, attribute, and relation constraints, including teapot colors and flower, candle, and cat arrangements.The comparisons are presented as complementing quantitative results and illustrating compositional alignment while maintaining visual coherence.Source: Figure 3 · p. 8, §5.2 · p. 8

5.3 ABLATION STUDY

• On GenEval2, direct generation scores 29.0 GM, the untuned planner with the workflow scores 49.7, and SFT reaches 81.5 GM.The untuned planner uses constraint-level feedback, interaction history, and historical images as editing sources; SFT’s largest skill-wise gains over it are Position (+27.3) and Count (+16.4).Source: §5.3 · p. 8
• 84.3 GM with NACA exceeds standard GRPO’s 80.3 by 4.0 points and SFT’s 81.5 by 2.8 points, with the highest scores across all five evaluated skills.These results support redistributing positive trajectory credit according to individual image actions’ progress.Source: §5.3 · p. 9, Table 3 · p. 9

5.4 ANALYSIS

• Joint selection occurs in three patterns: returning to an earlier higher-scoring image, continuing from a lower-scoring latest image, or choosing an earlier constraint-preserving candidate.The patterns reflect whether candidates preserve already satisfied constraints or satisfy previously unmet ones.Source: §5.4 · p. 9
• In controlled continuations, Joint reaches 55.2 mean final GM versus 52.4 for Best-only and 50.2 for Latest-only; Joint-trained students score 81.5 versus 77.5 GM.Joint improves over Best-only by 2.8 points (paired 95% CI: [0.46, 5.15]) in teacher continuations and by 4.0 points (paired 95% CI: [2.81, 5.18]) for student policies.Source: §5.4 · p. 9

6 CONCLUSION

• GenRevisit jointly selects continuation states and actions, with experiments showing improved compositional generation across benchmarks and cross-generator transfer without additional planner training.The conclusion frames source selection and editing as decisions to learn together.Source: §6 · p. 9

A EXPERIMENTAL SETUP AND INTERACTION PROTOCOL

• The training chain uses Joint selective SFT followed by five rounds of NACA reinforcement learning.The appendix separates planner decisions from information and rules supplied by the environment.Source: A EXPERIMENTAL SETUP AND INTERACTION… · p. 14

A.1 MODELS AND FROZEN COMPONENTS

• The planner starts from Qwen3-VL-8B-Instruct, while separate frozen Qwen-Image-2512 and Qwen-Image-Edit-2511 backends handle generation and editing.The planner specifies image content and, for edits, which candidate to modify; it does not choose the model, provider, sampling configuration, or seed.Source: §A.1 · p. 14

A.2 DECISION INPUTS, ACTIONS, AND CANDIDATE ACCESS

• The Joint input includes completed candidates, their images and evaluations, parent identifiers, and executed actions; an edit specifies the candidate, targets, preservation constraints, and instruction.Image pixels are bound to candidate and artifact identifiers, and the fields of an edit form one structured action.Source: §A.2 · p. 14
• The planner may edit any completed candidate, while GM-best controls submission; candidate rankings, SFT selection, and RL utility serve distinct roles.There is no prescribed share of historical edits or requirement to avoid the current best.Source: §A.2 · p. 15
• The first generation counts toward the five-image budget, and submitting the fixed GM-best candidate becomes mandatory when the budget is exhausted or it passes all constraints.Budget exhaustion takes precedence if the last image also passes all constraints; otherwise, the planner may stop early when no mandatory stopping condition applies.Source: §A.2 · p. 15
• Executed failed or ineffective edits remain in the history as evidence, while rejected raw output does not become persistent free-form memory.The environment rebuilds persistent state from validated actions, execution artifacts, and evaluation events.Source: §A.2 · p. 15

A.3 TRAINING SETTINGS

• Selective SFT establishes structured joint decisions and initializes five rounds of reinforcement learning.The SFT and RL settings are presented as aligned with the main-paper configuration and five-round training chain, respectively.Source: §A.3 · p. 16, Table A.4 · p. 16, Table A.5 · p. 16
• Only the planner’s language parameters are optimized; its visual components, image backends, and verifier remain frozen.The SFT checkpoint initializes the RL rounds.Source: §A.3 · p. 16
• The 320 prompt-group instances need not be distinct prompts, and action-segment counts vary with sampled trajectories rather than a fixed gradient-accumulation count.Groups with effectively zero reward variance are excluded from optimizer data.Source: §A.3 · p. 16

A.4 BENCHMARK AND EVALUATION PROTOCOLS

• GenEval2 reports final-image quality using Soft-TIFA GM, distinct from the passed-constraint proportion; evaluation is capped at five image calls and scores the submitted image.GenEval and GenEval++ results are reported separately from GenEval2 Soft-TIFA GM, and cross-generator evaluation changes the generation backend without additional planner training.Source: §A.4 · p. 17
• 84.3 GM is reported for the Qwen-backed submissions on all 800 prompts, and official rescoring agrees numerically with the online scores.The absolute difference between online and official mean scores on the GM×100 scale is 0.00000082; the maximum per-image difference is 0.00000437.Source: §A.4 · p. 17

B TEACHER DATA AND SELECTIVE SUPERVISION

• Teacher supervision is selected after execution, while each training input contains only information available before its target action.This separates useful evidence for target selection from behavior the student should imitate.Source: B TEACHER DATA AND SELECTIVE SUPERVISION · p. 18

B.1 PROMPT SELECTION AND TEACHER COLLECTION

• GPT-5.5 collects multi-round trajectories from 1,000 prompts, and teachers can jointly select an image and instruction from historical candidates when editing.The teacher and student use different input wrappers; the teacher’s four-part editing guidance is semantic structure within one instruction string, not four action fields.Source: §B.1 · p. 18

B.2 ACTION-LEVEL TARGET SELECTION

• An edit can be an SFT target if it improves its own input even when it remains below the archive best; degrading actions are not targets, but their feedback stays available for later decisions.Constraint-level regressions are recorded separately and do not automatically veto a GM-positive target; a degrading bridge action is not made a target merely because a later descendant succeeds.Source: §B.2 · p. 18

B.3 DECISION-TIME INPUTS AND LOSS MASKS

• Each selected target is paired with its reconstructed pre-action input, and only the assistant response is supervised; failed actions remain context rather than receiving negative imitation loss.For edits, the supervised response includes candidate choice, target and preservation sets, and the complete instruction.Source: §B.3 · p. 19

B.4 DATA SPLITS AND CORPUS STATISTICS

• The export contains 6,858 image bindings and 3,418 unique image artifacts across selected records, while training uses 800 distinct prompts; validation and test next-token losses are not final image-quality evaluations.The same image may appear in multiple decision contexts, and action-response counts are not prompt counts.Source: §B.4 · p. 19

C REINFORCEMENT LEARNING DETAILS

• The planner is trained for five NACA RL rounds starting from the Joint SFT checkpoint.The section specifies rollout state, round-specific policy references, rewards, action credit, and optimization loss.Source: C REINFORCEMENT LEARNING DETAILS · p. 20

C.1 ROLLOUT INITIALIZATION AND POLICY REFERENCES

• Each GRPO group contains four independent trajectories sharing a prompt and initial five-image budget, with each trajectory generating its own first image.The initial image counts toward the budget and NACA credit; the stated seed formula uses the rollout candidate index, not within-trajectory attempt number.Source: §C.1 · p. 20
• Each RL round collects trajectories from the preceding checkpoint, uses that checkpoint as the frozen KL reference, then optimizes for one epoch and eight steps.Only planner language parameters are updated; the vision tower, multimodal projector, image backends, and verifier remain frozen.Source: §C.1 · p. 20

C.2 TERMINAL REWARDS AND GROUP ADVANTAGES

• The terminal utility combines passed constraints with Soft-TIFA geometric mean, while invalid-policy, infrastructure-failure, and zero-variance cases receive distinct handling.The submitted image determines terminal reward; policy-invalid trajectories receive -1, infrastructure failures are retried or recovered, and zero-variance groups are excluded from optimization.Source: §C.2 · pp. 20–21

C.3 REFERENCE UTILITY AND LOCAL PROGRESS

• An edit’s local progress is measured against its selected source image, while a new generation is compared with the highest-utility available candidate.Progress measures the fraction of remaining utility gap closed; nonincreasing utility receives zero progress, and positive progress need not surpass the historical best.Source: §C.3 · p. 21

C.4 TOKEN-WEIGHTED CREDIT ALLOCATION

• NACA reallocates positive trajectory advantage across eligible image actions using normalized local progress, while preserving token-weighted credit and leaving nonpositive advantages unchanged.Zero-credit image actions retain their KL term; retrieval and submission responses outside the eligible image actions retain their trajectory advantage.Source: §C.4 · pp. 22–23

D.2 FEASIBLE-SET MONOTONICITY

• Historical-source access expands the feasible policy set, so optimal reachable value cannot decrease under the same controlled setting, though a trained policy need not realize that optimum.The comparison requires source-set inclusion over reachable histories while holding observations, tools, submission rules, and budgets fixed; strict improvement is not guaranteed.Source: D OPTIMIZATION POTENTIAL OF HISTORICAL… · p. 25, §D.2 · p. 25, §D.3 · p. 25

E.1 CONTROLLED TEACHER CONTINUATIONS

• On 200 unsolved teacher prefixes, Joint reaches mean final GM 55.2 versus 52.4 for Best-only and 50.2 for Latest-only, with a 2.8-point paired improvement over Best-only.Joint also raises the share of prefixes producing a candidate above the shared-prefix best from 57.0% to 65.0%.Source: §E.1 · p. 26

E.2 LEARNING FROM JOINT DEMONSTRATIONS

• 81.5 GM versus 77.5 shows that students trained on Joint demonstrations outperform Best-only-trained students on GenEval2, a 4.0-point paired improvement.Both policies use the same prompts, initial images, and evaluation protocol.Source: §E.2 · p. 26

E.3 HISTORICAL-SOURCE ACCESS AFTER REGRESSION

• 71.67% versus 60.00% shows historical-source access improves new-best discovery on degraded prefixes, an 11.67-percentage-point increase.Success requires generating a new candidate above the frozen prefix best; returning an existing historical image does not count.Source: §E.3 · p. 26

E.4 CROSS-BENCHMARK CATEGORY RESULTS

• 0.9844 on GenEval and 0.8929 on GenEval++ are the overall scores whose category breakdowns appear in Tables E.1 and E.2.The tables report cross-benchmark category results; E.2 follows the benchmark’s seven-category taxonomy.Source: §E.4 · p. 26, Table E.1 · p. 27, Table E.2 · p. 27

F CASE STUDIES AND QUALITATIVE COMPARISONS

• The case-study records distinguish teacher continuations, student behavior, and training-credit diagnostics; GM is on a 0–100 scale and arrows denote executed edits.Separate submission labels mark final outputs.Source: F CASE STUDIES AND QUALITATIVE… · p. 28

F.1 JOINT AND BEST-ONLY CONTINUATIONS

• 94.81 GM and 4/4 passed constraints for Joint contrast with Best-only’s submitted existing image at 82.65 GM and 3/4.Joint’s final edit used a candidate that was neither latest nor GM-best; the figure identifies the actual editing inputs.Source: §F.1 · p. 28, Figure F.1 · p. 28

F.2 HISTORICAL CONTINUATION AND LATEST-ONLY REFINEMENT

• 99.91 GM with all nine constraints passed contrasts with Latest-only’s final generated image, which still fails the target relation.Full-history edits the prefix best on both calls; Latest-only submits its original prefix best instead of its last generated image.Source: §F.2 · p. 29, Figure F.2 · p. 29

F.3 CURRENT QUALITY AND CONTINUATION POTENTIAL

• 99.27 GM with all five constraints passed follows an edit from a001, which had lower GM than the available candidate a000.The example shows one successful lower-scoring start, not that edits from a000 would fail or that either candidate has greater expected continuation value.Source: §F.3 · p. 30, Figure F.3 · p. 30

F.4 ACTION CREDIT ON A RECORDED TRAJECTORY

• In the recorded trajectory, an edit that lowered GM without adding passed constraints receives zero positive action credit, while the final revisit receives the largest multiplier.The four image actions preserve 1,316 token-weighted advantage coefficients; this is an allocation example, not a separate policy-performance estimate.Source: §F.4 · p. 31, Table F.1 · p. 31, Figure F.4 · p. 31

F.5.1 DESIGNED-PROMPT RESULTS

• Nine designed-prompt examples combine counts, attributes, spatial relations, and materials; tile sizes do not encode scores.Panels f–i use final-frame extracts from supplied trajectory sheets, while the other panels use standalone images.Source: Figure F.5 · p. 32, §F.5.1 · p. 32

F.5.2 GENEVAL2 RESULTS

• Selected SFT outputs show initial generations in panels a–c and last supplied frames after editing in panels d–f.The prompt key retains requested counts, attributes, materials, and spatial relations.Source: Figure F.6 · p. 33

F.5.3 GENEVAL++ RESULTS

• Figure F.7 shows nine selected GenRevisit outputs spanning the benchmark’s seven compositional categories, with prompts abbreviated for display.The figure notes minor grammatical corrections that preserve the requested objects and relations.Source: Figure F.7 · p. 34

F.5.4 AN EXECUTED REVISIT TRAJECTORY

• The recorded policy revisits the initial candidate after an unsuccessful edit, then continues from a later candidate; arrows show editing dependencies rather than image chronology.The figure displays intermediate errors alongside the submitted result.Source: §F.5.4 · p. 35, Figure F.8 · p. 35

G.1 TEACHER AND STUDENT PROMPT CONTRACTS

• Teacher and student each return one executable structured response using candidate images and feedback to choose an editing input and instruction, which need not target the latest or GM-best image.The teacher asks for operation, spatial grounding, preservation requirements, and forbidden changes within one instruction; the student prompt and schema do not repeat these as separate requirements.Source: §G.1 · p. 36, Table G.1 · p. 36

G.2 ACTION FIELDS AND EXAMPLES

• Each action response specifies one action and its arguments; the interface enforces legal candidate and constraint identifiers, permissions, and action-specific fields.The four mutually exclusive actions are generation, editing, skill query, and submission; queries wait for real returned observations, and rejected responses request a complete legal action.Source: §G.2 · p. 36, Table G.2 · p. 36

G.3 GUIDANCE RETRIEVAL AND VERIFIER INTERFACES

• Skill retrieval is limited to applicable guidance returned by the fixed skill directory, while the verifier records fixed-question results rather than treating planner declarations as proof.Action-pose guidance requires an evaluated verb failure or uncertainty; unresolved queries block another query, and internal transition labels compare each child with its declared parent.Source: §G.3 · p. 38

G.4 VERIFIER REQUESTS AND SCORE COMPUTATION

• Per-constraint verifier scores sum probabilities of expected-answer variants for the first token, with pass at p_k ≥0.50, fail at p_k ≤0.20, and uncertain otherwise.Per-prompt GM weights VQA questions equally, and the benchmark score is the arithmetic mean of per-prompt GMs multiplied by 100.Source: §G.4 · p. 39

G.5 VERBATIM JOINT PROMPTS AND INPUT WRAPPERS

• The Joint prompts direct planners to compare editable sources together with concrete interventions against achieved GM-best, weighing repair evidence and risk rather than candidate age alone.The selected source must ground its targets, preservation choices, counts, and spatial anchors; prior instructions and observed outcomes inform route reassessment.Source: TEACHER SYSTEM MESSAGE · pp. 41–42
• The joint policy selects both an editable candidate and a concrete intervention, using image and feedback history to choose a useful source rather than defaulting to the latest or GM-best image.When images exist, all completed candidates are eligible edit sources; the policy adapts the intervention to the chosen source and avoids blindly repeating failed interventions.Source: TEACHER SYSTEM MESSAGE · pp. 42–43

G.6 VERBATIM SKILL LIBRARY

• The skill library provides targeted guidance for exact-count repair, static spatial relations, attribute-to-entity binding, and narrow edits with explicit targets and anchors.The listed guidance emphasizes preserving correct instances, relations, and non-target attributes while making localized changes.Source: COUNTING AND INSTANCE LAYOUT (VERSION… · p. 44, SPATIAL RELATION LAYOUT (VERSION 2.0.0) · p. 45, ATTRIBUTE ENTITY BINDING (VERSION 1.0.0) · p. 45
• The supplied material contains skill-library guidance, not GenEval++ result values or benchmark comparisons.It covers action-pose-relation and object-identity-presence instructions.Source: ACTION POSE RELATION (VERSION 2.1.0) · p. 47, OBJECT IDENTITY PRESENCE (VERSION 1.0.0) · p. 49

• The planner controls instructions and candidate choices, not image-model parameters; a five-image budget limits alternatives, and revisiting cannot guarantee a useful edit.Source: H LIMITATIONS · p. 50
• Constraint feedback comes from a fixed verifier, while GM, passed-constraint count, and perceptual quality measure different properties; compositional scores alone do not establish aesthetic preference.Source: H LIMITATIONS · p. 50
• GM-based SFT can omit temporarily degrading actions later followed by success, while NACA tracks local utility rather than causal contribution to the final submission.Source: H LIMITATIONS · p. 50
• NACA leaves nonpositive trajectory advantages unchanged and conserves advantage coefficients, not gradient norms; token-sum optimization's response-length effects remain.Source: H LIMITATIONS · p. 50
• Mandatory stopping and submission eligibility are environment-enforced; submitting the permitted GM-best does not establish unconstrained final-image ranking, and teacher successes do not establish student learning.Source: H LIMITATIONS · p. 50