Maestro: Self-Improving Text-to-Image Generation via Agent Orchestration

Motivation

• Manual prompt refinement is costly, time-consuming, and expertise-intensive because prompt sensitivity and user under-specification can produce unsatisfactory images.Users repeatedly generate, evaluate, and revise prompts to approach their intended result.Source: §1 · p. 1

Maestro

• Maestro autonomously self-improves T2I generations through prompt refinement using only an initial user prompt and black-box MLLM and T2I calls.The system is designed to mimic human iterative refinement without requiring manual intervention.Source: §1 · p. 2

Self-critique

• Specialized MLLM critics identify image weaknesses and under-specification, produce interpretable edit signals, and let a verifier preserve the user’s original intent.This replaces scalar reward feedback with targeted multimodal reasoning for prompt editing.Source: §1 · p. 2

Pairwise objective

• An MLLM judge compares each current image against the best image so far in head-to-head tournaments, returning the incumbent best generation after a budget or patience criterion.The pairwise objective avoids relying on a single scalar quality score for subjective, multifaceted visual evaluation.Source: §1 · pp. 2–3

Results

• Experiments on complex T2I tasks with black-box models show significant image-quality gains over initial prompts and state-of-the-art automated methods using automatic metrics and human evaluations.Effectiveness increases when more capable MLLMs, such as Gemini 2.0, are used in the framework.Source: §1 · p. 3

Scope

• Black-box T2I prompt optimization is studied through train-time and test-time methods, excluding approaches requiring model modification or white-box access.This scope focuses on prediction API-only access to the underlying T2I model.Source: §2 · p. 3

Train-time approaches

• Train-time methods amortize optimization cost but depend on large-scale data, may specialize to particular generators or reward models, and cannot readily use test-time image feedback.Maintaining such dedicated optimizer models also requires retraining as foundation models advance.Source: §2 · p. 3

Test-time approaches

• Test-time methods iteratively generate, evaluate, and revise prompts, offering model agnosticism and flexibility without dedicated training or large-scale data collection.Some use explicit objectives such as DSG and CLIPScore, while others directly query MLLMs for revisions.Source: §2 · pp. 3–4

Maestro pipeline

• Maestro’s pipeline initializes decomposed visual questions and an improved prompt, then generates images, scores them against the questions, and uses a comparator to track the best result.Critique and improvement signals are verified against the visual questions before producing proposals for the next iteration.Source: Figure 1 · p. 4, §2 · p. 4

Limitations

• Existing approaches often paraphrase underspecified prompts, use scalar reward signals that miss multifaceted visual quality, or lack a mechanism for retaining the best candidate.Maestro combines richer critique with a pairwise objective to address these limitations.Source: §2 · pp. 4–5

3. Methodology

• Maestro is primarily a model-agnostic test-time optimization framework that requires only the initial prompt, while accepting higher runtime from iterative generation.The design can also incorporate train-time prompt optimizers during initialization.Source: §3 · pp. 5–6

3.1. Initialization (Blocks 1-3 in Fig. 1)

• Maestro initializes an LLM-rewritten prompt and decomposed visual questions from the user prompt, establishing both a stronger starting point and interpretable criteria for later evaluation.The rewrite enriches descriptions, while DVQs probe properties such as subject, text, style, lighting, and setting.Source: §3.1 · p. 6

3.2. Pairwise Objective and Comparator (Block 6)

• Maestro directly optimizes pairwise MLLM preferences by dueling each new candidate with the incumbent best image rather than aggregating DVQ scores into a scalar objective.The comparator can use repeated, position-swapped judgments and stops after a fixed budget or patience limit.Source: §3.1 · p. 7, §3.2 · pp. 7–8
• The system generates an image for each proposal, evaluates it against DVQs with an MLLM, and stores prompt-image-response tuples for iterative comparison.DVQ responses are probabilities of the “Yes” token for each image-question pair.Source: §3.2 · p. 7

3.3. Generating New Prompt Proposals (Blocks 7-9)

• Its dual-generator proposal stage combines targeted editing of failed DVQs with implicit holistic improvement from the best prompt-image pair.Targeted editing produces rationalizations and specific prompt revisions for unmet criteria, while implicit improvement is not restricted to predefined DVQs.Source: §3.3 · pp. 8–9

3.4. User Intent Grounding via Self-Verification (Block 10)

• Because richer edits can cause semantic drift, a dedicated self-verification block checks revised prompts against the original DVQs and corrects violations before further generation.Verification repeats until no change is proposed or a patience limit is reached, then routes proposals back to the T2I model.Source: §3.4 · pp. 9–10

Experimental setup

• Experiments use Imagen 3 on p2-hard and DSG-1K, filter out prompts that achieve perfect DSGScore through repeated sampling, and limit repeated-sampling methods to eight T2I queries.The evaluation therefore emphasizes cases where repeated sampling alone does not reach perfect prompt-image consistency.Source: §4 · pp. 10–12

Main results

• Human side-by-side evaluation compares Maestro with LM-BBO, the next strongest automatic baseline, using responses from three participants.The aggregated result is weighted by the number of responses submitted by each respondent.Source: Figure 6 · p. 12, §4 · p. 12
• Maestro consistently outperforms baselines across metrics, with superior AutoSxS win rates and stronger DSGScore than OPT2I.The comparison uses Gemini 2.0 Flash as judge and includes both pairwise and pointwise evaluation.Source: Table 1 · p. 11, §4 · p. 12
• Maestro’s automatic pairwise comparisons agree well with human judgments, supporting pairwise evaluation as aligned with human assessment of generation quality.The study notes that human evaluation did not include a tie option, while the automatic evaluation did.Source: §4 · p. 12

Qualitative analysis

• Qualitative trajectories show Maestro correcting underspecified or complex concepts, including Jeff Koons-style Thanos, Sierpiński pyramids, and Hubble-like telescope details.The examples illustrate improvements in instruction following, domain-specific fidelity, and text rendering.Source: §4 · pp. 13–14
• Maestro also refines already adequate generations by improving aesthetics, growth emphasis, calmness, window detail, and Tyrannosaurus rex features.These examples support improvements in nuanced alignment with user intent rather than only basic prompt satisfaction.Source: §4 · pp. 13–14

Ablation studies

• Ablations compare rewriting, iterative improvement, pairwise best-generation tracking, and full self-verification on subsampled p2-hard and DSG-1K data.The full algorithm is V+P+I+R, while the other variants remove pairwise comparison, iterative improvement, or verification components.Source: §4 · p. 13, Figure 7 · p. 13
• The full Maestro algorithm modestly surpasses P+I+R head-to-head and improves DSGScore more substantially, while each added component improves AutoSxS and DSGScore over the original prompt.The authors connect the larger DSGScore gain to self-verification grounding prompts in user intent.Source: §4 · p. 14

• Maestro improves image quality over initial prompts and state-of-the-art techniques by combining interpretable multi-agent critique with pairwise MLLM-as-a-judge evaluation.Source: §5 · p. 14
• The framework’s benefits scale with advances in MLLMs, supporting its use as an autonomous self-improvement approach for T2I generation.Source: §5 · p. 14
• Current evaluations focus on Imagen, although Maestro is designed to be model-agnostic and requires testing across diverse T2I systems.Source: §5 · p. 14
• Future work includes applying Maestro’s principles to audio or video synthesis and using improvement trajectories to train models beyond prompt refinement.Source: §5 · p. 15

Initialization

• DVQs are generated once using instructions and few-shot examples from Cho et al. (2024), then fixed across methods that use them.Source: A. Implementation Details · p. 18

Pairwise comparison

• Pairwise comparison runs 2n parallel judgments with reversed image order to reduce positional bias, retaining the preferred candidate as the best prompt so far.Source: A. Implementation Details · p. 18

Prompt generation

• Targeted editing answers decomposed visual questions and uses the resulting feedback to revise prompts, with posthoc rationalization chosen for fairer baseline comparison than chain-of-thought prompting.Source: A. Implementation Details · pp. 18–19
• Implicit improvement can emit NO_CHANGE when the current generation is satisfactory, and the same termination option is applied to LM-BBO.Source: A. Implementation Details · pp. 19–20
• Maestro generates one targeted-editing prompt and one implicit-improvement prompt per iteration, with a maximum of eight T2I calls across four steps.Source: A. Implementation Details · p. 20

Human preference studies

• The human preference study used three author colleagues, randomized image ordering, and comparisons between LM-BBO and Maestro without revealing optimized prompts.Source: A. Implementation Details · p. 20

Different Judge Models

• Relative method strength remains consistent across Gemini 1.5 Pro, Gemini 1.5 Flash, and Gemini 2.0 Flash judges.Source: §B.1 · p. 20, Figure 8 · p. 21
• Gemini 1.5 Flash produces a larger fraction of ties, likely because of weaker multimodal understanding and stronger positional bias.Source: §B.1 · p. 21

Different Optimizer Models

• Maestro’s win rate improves with Gemini 2.0 Flash as optimizer, exceeding both corresponding baseline methods and Maestro with the weaker optimizer.Source: §B.2 · p. 21, Figure 9 · p. 21

B.3. Additional Qualitative Results

• The optimized text prompts producing the images in Tables 2 and 3 are provided separately.Source: §B.3 · p. 21
• The examples cover diverse prompts, including a retrofuturistic movie poster, a smoke-and-wind figure holding a portal, and a superhero-mask combination.Source: §B.3 · p. 22
• Other prompts depict an armored anthropomorphic wolf, abstract bicycles, graffiti text, and the word “DRAW” made of pencils.Source: §B.3 · p. 22
• A bus scene illustrates an everyday composition involving a white-and-blue bus, brick ground, backlighting, and a person in a brown jacket.Source: §B.3 · p. 22
• Table 3 presents additional qualitative examples of image trajectories generated by Maestro.The table complements Table 2 in the main text.Source: §B.3 · p. 21, Table 3 · p. 22