PaperlayerView exact revision on arXiv

Source-linked AI summary

Scaling Properties of Text Conditioning in Visual Generation

Zilong Chen, Chaorui Deng, Kunchang Li, Hongyi Yuan, Haoqi Fan

arXiv:2607.29679v1cs.CV

TL;DR

Text-to-image models have lacked clear evidence on how caption information content affects training. This paper measures text-conditioning scaling through structured-language metrics, then combines structured supervision with prompter training to improve generation, leading evaluated open-weight models on all but one reported metric and matching or surpassing strongest closed systems on most.

  • Problem

    Evidence is limited on whether caption information content, rather than natural-language length, systematically affects text-to-image training and generation.

  • Method

    The paper calibrates GPG and ED against diffusion loss, constructs image-derived structured prompts, and trains an LLM prompter with distillation and refinement.

  • Results

    The resulting system leads every evaluated open-weight model on all but one reported metric and matches or surpasses the strongest closed systems on most evaluations.

  • Takeaways & Limitations

    Caption information content is a measurable scaling axis, while structured prompts and prompter training improve visual coherence, prompt fidelity, and compositional generation.

  • Takeaways & Limitations

    The fitted scaling relations may shift with different judges or extractors, and structured prompting adds prompter inference latency.

Abstract

from arXiv · show

We study empirical scaling properties for text conditioning in visual generation. Such properties have rarely been measured because diffusion loss does not scale with the number of tokens in natural-language prompts. Surprisingly, we find that the converged diffusion loss scales with the amount of structured language in the prompt. To quantify structured language, we adapt two complementary measures: a white-box likelihood metric (GPG) and a black-box attribute metric (ED). Across controlled training runs, the converged diffusion loss decreases approximately linearly with GPG and follows a power law with ED. Guided by these scaling properties, we improve \emph{diffusability} by constructing structured prompts with semantic and geometric annotations derived from images, and improve \emph{promptability} by training a prompter through supervised fine-tuning, cold-start, and verifier-gated on-policy distillation. The resulting system outperforms all evaluated open-weight models on nearly every compositional, reasoning, and world-knowledge benchmark, while matching or surpassing the strongest closed-weight models on most evaluations.

1 INTRODUCTION

The paper studies how structured, informative text conditioning affects diffusion training and introduces structured prompts plus a trained prompter to improve visual generation. Together, these methods produce an end-to-end system that leads evaluated open-weight models on all but one metric and matches or surpasses the strongest closed systems on most.

  • Motivation: The introduction identifies a basic asymmetry: language models receive training signal from the text stream, whereas text-to-image models do not receive an equivalent signal from prompt tokens.This motivates studying scaling properties of text conditioning rather than assuming that text-to-image generation follows language-model scaling directly.
  • Scaling properties of text conditioning: The study measures caption informativeness with complementary GPG and ED metrics and relates both to converged diffusion loss under a fixed training recipe.Across sweeps of natural-language and structured-prompt formats and detail levels, image data, architecture, initialization, and compute remain fixed.
  • Raising promptability through training and inference-time refinement: A prompter must infer unspecified visual details while preserving the user’s request because test-time generation lacks a paired image or oracle annotation.The paper calls this capability promptability and reports that zero-shot quality improves with prompter scale and, except at the smallest scale, chain-of-thought inference.
  • End-to-end results: The end-to-end system leads every evaluated open-weight model on all but one reported metric and matches or surpasses the strongest closed systems on most evaluations.The widest margins occur on composition- and reasoning-heavy benchmarks, with gains also extending to prompt fidelity, visual coherence, and compositional detail.
  • Raising diffusability through structured prompts: Structured prompts organize image-grounded content into named fields, increasing measured informativeness and lowering diffusion loss without architectural changes.The contribution explicitly frames this intervention as raising diffusability through structured prompts.

2 RELATED WORK

Prior work improves text-to-image generation through larger models, richer captions, structured controls, and inference-time prompt enhancement. This work differs by studying caption diffusability and prompter promptability within a fixed structured schema and diffuser.

  • Text-to-image scaling, long captions, and structured control: Text-to-image progress has relied on diffusion and latent-diffusion models, transformer backbones, longer captions, and structured control.These approaches include layout conditioning, attention manipulation, and richer caption representations.
  • Text-to-image scaling, long captions, and structured control: This work places boxes, depth, and relations in ordinary structured text rather than intervening on a fixed diffuser through auxiliary spatial conditions or internal attention manipulation.The comparison distinguishes the proposed schema from methods such as GLIGEN and Attend-and-Excite.
  • Recaptioning and caption quality: Detailed synthetic recaptioning and richer or principled captions improve prompt following, while long JSON captions outperform short captions under matched training settings.The cited systems include DALL-E 3, PixArt-α, PixArt-Σ, CogView3, RECAP, and FIBO.
  • LLM prompters, RFT, and inference-time methods: Inference-time prompt enhancers use regional decomposition, multimodal reasoning, reinforcement fine-tuning, chain-of-thought rewriting, or input-side scaling.Examples include RPG, PromptEnhancer, and related input-side inference-time methods.
  • LLM prompters, RFT, and inference-time methods: With the schema and backbone fixed, the work studies prompter promptability under LLM scale, chain-of-thought, verifier-gated self-distillation, and an agentic refine–render–judge loop.Unlike generator-side reward or preference optimization, the verifier filters prompter rollouts while RFT holds the SP-trained diffuser fixed.

3 METHOD

The method replaces unreliable lengthening of free-form captions with structured prompts that organize image-grounded information across global, object, geometric, and relational fields. It measures caption informativeness through GPG and ED, then uses calibrated scaling properties to improve diffusability and promptability through structured supervision and trained prompting.

  • Structured prompts: Lengthening free-form natural-language captions does not reliably increase useful supervision, motivating structured prompts as the method’s shared interface.The NL intervention leaves ED and converged diffusion loss essentially unchanged despite nudging GPG upward, whereas restoring structured-prompt fields improves informativeness and lowers loss.
  • Structured prompts: Structured prompts organize image-grounded information into global fields, per-element attributes and geometry, and cross-element relationships.Per-element entries include identity, attributes, actions, bounding-box position, optional depth, and photography.
  • Scaling analysis: GPG and ED provide complementary white-box and black-box measures of caption informativeness for analyzing converged diffusion training loss.GPG is a total masked conditional log-likelihood gain, while ED matches caption attributes against an image-grounded reference set and reports F0.5(PA, RA).
  • Scaling analysis: Measured informativeness, rather than caption length, predicts matched-budget converged diffusion loss across the calibrated caption configurations.The scaling relations have residual standard deviations of approximately 6×10^-4 for GPG and 7.8×10^-4 for ED.
  • Training interventions: Image-to-structured-prompt annotation supplies full-schema supervision for diffusion training, while a trained LLM prompter improves promptability alongside the diffuser’s improved diffusability.The method combines structured supervision from paired-image annotations with a trained LLM prompter, and evaluates their combined system and corresponding ablations later.

4 EXPERIMENTS

Experiments show that structured prompting improves compositional and alignment performance over representative systems and matched controls. Prompter training and inference-time refinement provide complementary gains, with training delivering stronger single-shot results and refinement mainly improving structure.

  • Comparison with existing systems: The system leads or ties nearly every reported metric among evaluated open-weight models and matches or surpasses evaluated closed systems on most metrics.Complex prompts more faithfully realize spatial layouts, object counts, and attribute bindings.
  • Matched NL control: 72.5 GenEval2 GM and 85.2 CoReBench are achieved by our system, versus 52.8 and 74.7 for the official Qwen-Image prompt enhancer.Both systems use the same Qwen-Image architecture, isolating the contribution of structured prompts and associated training.
  • Promptability: transferring LLM progress: Zero-shot structured-prompt filling improves alignment and GSB over matched NL rewriting but lowers DPG-Bench and structure within single-turn comparisons.The experiment holds the schema and Qwen-Image diffuser fixed while varying the prompt-producing backend and inference procedure.
  • Promptability: training the prompter: 4.860→6.273 is the largest single-stage structure gain from SFT; verifier-gated OPSD then produces the strongest endpoint, with structure reaching 7.600.Cold-start further improves structure and GSB, while verifier-reward GRPO and ungated OPSD improve structure, alignment, and GSB over Cold-start.
  • Returns from additional rounds: 1.247 and 0.660 are the one-to-eight-round structure gains for Base and the trained prompter, compared with alignment gains of 0.366 and 0.266.Iterative feedback primarily repairs object decomposition, relations, and layout while schema, diffuser, judge, and prompter weights remain fixed.
  • Inference-time refinement: 54.7% GSB at Tmax =8 exceeds the strongest coding agent’s 44.7%, while increasing the trained prompter’s Tmax from four to eight changes structure from 8.213 to 8.260.Average rounds rise only from 2.04 to 2.31, indicating limited benefit from substantially longer refinement trajectories.

5 CONCLUSION … A.1 DEFAULT SCORING PROTOCOL

The paper identifies caption information content, measured by GPG and ED, as a scalable axis for text-to-image learning and frames captioning as a Diffusability×Promptability interface. Structured prompts improve diffusability, while prompter training improves promptability and end-to-end visual quality without changing the diffusion architecture, subject to measurement and schema limitations.

  • 5 CONCLUSION: Caption information content, rather than caption length, predicts converged diffusion loss through calibrated GPG and ED scaling relations.These relations enable comparisons among caption representations after calibration.
  • 5 CONCLUSION: Structured prompts raise diffusability by exposing and organizing image-grounded variables in addressable fields.Prompter scaling and training raise promptability by translating user requests into detailed, coherent structured instances.
  • Appendix: Agentic inference-time scaling increases refine–render–judge rounds through larger Tmax values under a fixed schema, Qwen-Image backbone, online Gemini judge, and prompter weights.Final outputs are evaluated offline by GPT-5.4 for structure/alignment and order-swapped GSB net preference against the zero-shot Qwen3.5-397B-A17B Base prompter at Tmax = 1.
  • 5 CONCLUSION: Matched natural-language retraining indicates that structured-representation gains are not explained by additional training alone.The resulting system improves visual coherence, prompt fidelity, and compositional generation without changing the diffusion architecture.
  • 5 CONCLUSION: The two metrics depend on image-conditioned measurement and may shift under different vision-language judges or attribute extractors.GPG queries a vision-language judge at scoring time, whereas ED uses one-time offline extraction of attribute tuples from each image.
  • 5 CONCLUSION: The hand-designed schema remains a limitation, with automatic schema discovery and extension to video identified as future directions.Agreement between GPG and ED reduces but does not eliminate the risk of scoring-interface-specific trends.
  • Appendix: The appendix includes an E.5 GSB pairwise preference protocol as part of the evaluation materials.The referenced protocol appears in the appendix section listing.
  • A.1 DEFAULT SCORING PROTOCOL: Headline GPG scores use Qwen3.5-397B-A17B frozen as the VLM judge, with images resized to a 1024-pixel longer side while preserving aspect ratio.Captions are mapped to a deterministic scored sequence; SP canonicalization removes specified global metadata keys, while NL captions retain their text with template-only spans marked separately.

A.2 GPG ROBUSTNESS ACROSS JUDGES

GPG’s linear scaling relationship is evaluated with the complete judge interface held fixed, rather than with likelihood queries alone. Across compatible judges and coordinate representations, the ordering is robust, while an unadapted bbox format can invert it.

  • Interface control: GPG robustness depends on fixing the complete judge interface, including the VLM, tokenizer, conversation template, image processor, and coordinate convention.Headline comparisons hold this interface fixed rather than varying only the likelihood query.
  • Interface control: The default judge is Qwen3.5-397B-A17B because its normalized bbox convention directly matches the caption format’s 0–999 coordinates.This avoids an additional coordinate conversion in the headline measurement.
  • Cross-judge robustness: GPG’s ordering remains robust across compatible judges, including smaller alternatives and the cross-family InternVL3-8B check.Figure 16 reports GPG →MSE fits normalized separately per judge and compares fit quality by judge size.
  • Coordinate compatibility: r = +0.23 occurs when Qwen2.5-VL-7B receives unadapted 0–999 bbox tokens, inverting the ordering; a native-bbox adapter resolves the representation mismatch.The raw case cannot parse the bbox tokens, whereas the adapter maps them to native pixel-space coordinates.

A.3 FULL 15-SETTING DATA

Table 7 provides the complete 15-setting GPG, ED, and converged diffusion-loss data underlying Eqs. (3) and (4), measured consistently across all runs. It also documents the shared measurement protocols, unified token budget, and ablation-suite baseline accounting.

  • A.3 FULL 15-SETTING DATA: All settings use the same 30,000 paired image UIDs for both caption-side measurements.GPG uses Qwen3.5-397B-A17B with content masking and canonicalize-JSON; ED follows the extraction and matching protocol in Appendix B.
  • A.3 FULL 15-SETTING DATA: MSE is measured at the unified budget of 2.84×10^10 cumulative image tokens reached by every run.This common budget supports the per-setting comparisons reported in Table 7.
  • A.3 FULL 15-SETTING DATA: Table 7 reports all 15 settings’ GPG, ED, and converged diffusion loss, sorted by GPG.The full-schema ablation baseline coincides with Structured L10 and is counted once.

A.4 MONOTONICITY ANALYSIS

Across 15 settings sorted by GPG, MSE decreases overall, despite small local reversals whose causes are not assigned.

  • A.4 MONOTONICITY ANALYSIS: Across the 15 settings, MSE decreases overall when sorted by GPG, with a few small local reversals.The total reversal magnitude is approximately 0.0016 MSE units; approximately 0.0012 comes from structured L5 relative to nearby NL settings.
  • A.4 MONOTONICITY ANALYSIS: The remaining violations are |∆MSE|≤2×10−4, matching trailing-window read-out variation and not assigned to a specific cause.The reversal is reported descriptively rather than attributed to a specific cause.

A.5 FIT SENSITIVITY ACROSS SETTINGS AND TRAINING BUDGETS … B.3 SHARED-UID RESAMPLING SENSITIVITY

The appendix tests whether the reported scaling relations remain stable across resampling, training budgets, metric constructions, and paired-image resampling. It also clarifies that the mutual-information framing motivates, but does not theoretically determine, the empirical relations.

  • A.5 FIT SENSITIVITY ACROSS SETTINGS AND TRAINING BUDGETS: Over B =2,000 resamples, fit magnitudes remained high: mean±SD |r| = 0.984 ± 0.008 for GPG and |r|=0.973 ± 0.012 for ED.Each setting was trained once, so these analyses do not estimate seed-level optimization uncertainty or population confidence intervals.
  • A.5 FIT SENSITIVITY ACROSS SETTINGS AND TRAINING BUDGETS: Across six matched budget cuts, refits preserved the linear GPG relation and ED power law, while fitted slope and exponent steepened mildly during training.Budget read-outs used trailing-window mean training MSE, with corrected token axes for the two restarted runs.
  • A.5 FIT SENSITIVITY ACROSS SETTINGS AND TRAINING BUDGETS: Trailing-window block-mean variation spanned 1.0–5.0 × 10−4, with median 3.5×10−4, comparable to residuals of 6×10−4 for GPG and 7.8 × 10−4 for ED.These measurements contextualize residual scale but do not separate lack-of-fit from run-level optimization variability.
  • A.6 MUTUAL-INFORMATION MOTIVATION AND THE EMPIRICAL STATUS OF THE RELATIONS: The mutual-information identity motivates relating image-grounded caption information to conditional loss, but GPG is only an operational, judge-dependent estimate.The framing does not derive the functional form observed in the experiments.
  • A.6 MUTUAL-INFORMATION MOTIVATION AND THE EMPIRICAL STATUS OF THE RELATIONS: Equation (3) is an empirical, recipe-specific calibration: GPG measured by a frozen VLM tracks converged loss across tested caption settings, without universal slope interpretation.The fitted slope depends on the judge, caption construction, model, objective, optimization, and fixed training budget.
  • B.1 IMAGE-GROUNDED SOURCE EXTRACTION: The ED image-grounded source set used one offline Gemini 3 Pro call per image, producing a mean of 102 OARG tuples over the shared 30,000-image pool.GPT-5.4 separately extracted caption-side tuples and performed symmetric paraphrase-tolerant matching against cached image-side tuples.
  • B.2 WHY PRECISION-WEIGHTED F0.5?: R2 decreased from 0.943 for F0.5 to 0.313 for F2, although the fitted log–log slope stayed negative and the nested SP ladder remained monotonic.Variant MAE more than doubled, indicating that the precision-oriented score was markedly more predictive.
  • B.3 SHARED-UID RESAMPLING SENSITIVITY: A shared-UID bootstrap over 30,000 paired images gave Spearman ρ = −0.979 ± 0.010, with central 95% range [−0.993, −0.957] between ED and converged MSE.The corresponding log–log Pearson range was [−0.978, −0.946], and the exponent range was [−0.220, −0.191]. These ranges condition on cached proposer, extractor, and matcher outputs.

B.4 ROBUSTNESS ACROSS SOURCE EXTRACTORS

ED’s relationship with converged MSE and the ordering of experimental cells remain robust across source extractors, although absolute ED scales shift between backends. Accordingly, ED is used as a relative ruler with one fixed extractor, comparing rankings and slopes rather than absolute values.

  • Cross-extractor robustness: ED remains a strong negative predictor of converged MSE across extractors, with ρ(ED, MSE) from −0.86 to −0.90.The power-law slope also keeps its sign and rough magnitude.
  • Cross-extractor robustness: Cross-backend cell rankings remain consistent, with rank ρ=0.80–0.92 and per-tuple Cohen κ=0.75–0.77.These results indicate agreement in relative ordering despite different extraction backends.
  • Cross-extractor robustness: Absolute ED levels are not backend-invariant: absolute-agreement ICC is 0.11, whereas consistency ICC is 0.64.Different extractors shift ED scales while preserving ordering.
  • Methodological choice: The paper therefore uses ED as a relative ruler with a single fixed Gemini 3 Pro extractor, comparing rankings and slopes across backends.Absolute ED values are never compared between extractors.
  • Supporting measurements: Complete per-setting ED values appear with GPG and converged MSE in Table 7, supporting the power-law fit of Eq. (4).Grid-k denotes verbal locations on a k×k grid, while ablation rows mask the named L10 field.

C STRUCTURED-PROMPT SCHEMA AND IMPLEMENTATION DETAILS

The system uses a structured prompt with required scene-skeleton fields, optional relational and global-control fields, and an inference-only ratio control. Its implementation combines multi-stage image annotation, matched diffusion training, and a serially trained LoRA prompter.

  • C.1 STRUCTURED-PROMPT SCHEMA: The schema requires intent, macro scene, and elements fields, while relationships and atmosphere, photography, style, and lighting are optional controls.The prompter also emits an inference-only ratio field that selects the output canvas and is removed before diffusion conditioning.
  • C.2 ANNOTATION PIPELINE: Seed-VL builds the scene skeleton and global fields, separating foreground from background and ordering entities by compositional importance.All annotation models run frozen at the specified inference settings.
  • C.2 ANNOTATION PIPELINE: Per-element annotation combines concise captions and named visual-dimension keys with Sapiens pose overlays, while raw 133-keypoint predictions remain outside the structured prompt.The pose overlay helps resolve body-side orientation and joint geometry.
  • C.2 ANNOTATION PIPELINE: DepthAnything V2 and SAM 2.1 provide depth, masks, and occlusion cues that support geometric relations, while semantic relations are inferred during VLM reconciliation.Relative depth is quantized from 0 (nearest) to 255 (farthest) when available.
  • C.2 ANNOTATION PIPELINE: A second VLM pass assembles the stage-1–3 outputs into a schema-constrained L10 record, followed by deterministic degradation masks for L10-to-L9-to-L5 variants.One degradation set is generated per L10 annotation and reused across training cells.
  • C.3 DIFFUSION-BACKBONE TRAINING: BAGEL supports the scaling-property sweep, whereas Qwen-Image-2512 supports promptability and main-result experiments; both use rectified-flow or flow-matching diffusion objectives.BAGEL is trained per caption condition, while Qwen-Image is trained once on the structured-prompt and natural-language corpus and then held fixed across prompter ablations and main results.
  • C.3 DIFFUSION-BACKBONE TRAINING: Each BAGEL scaling cell is compared at a common cumulative image-token budget of 2.84×10^10 tokens, defining the matched-budget converged MSE used in the fits.Qwen-Image is instead trained once for 500,000 steps.
  • C.5.1 TRAINING STAGES: The prompter is a rank-128 LoRA adapter on frozen Qwen3.5-397B-A17B weights, trained serially through SFT, Cold-start, and RFT.Stage 1 uses ∼333k examples and ∼0.97B tokens, with roughly one third devoted to user-prompt-to-structured-prompt JSON and reverse image-to-JSON tasks.

C.4 SCHEMA FIELD ABLATION

At a common training budget, scene context is the most influential L10 schema field group: removing it causes the largest observed degradation in GPG and MSE. Bounding boxes produce the second-largest MSE increase.

  • Schema field ablation: At 2.84 × 10^10 cumulative image tokens, removing scene context reduces GPG by 42 and raises MSE by 35.3 × 10−4.This was measured across six retrained ablation runs at a common training budget.
  • Schema field ablation: Scene context is the most influential field group in the controlled ablation setting.Its removal causes a substantially larger loss increase than any other tested ablation.
  • Schema field ablation: Removing bounding boxes produces the second-largest MSE increase, at 14.4 × 10−4.This follows the scene-context ablation in effect size among the reported field-group removals.

D ADDITIONAL BENCHMARK RESULTS · D.1 DPG-BENCH PER-CATEGORY BREAKDOWN · D.2 GENEVAL PER-SKILL BREAKDOWN

The additional results document evaluation protocols and provenance, then break down DPG-Bench by category and GenEval by skill. These breakdowns cover compositional scene understanding, entity and attribute reasoning, spatial relations, counting, colors, and binding.

  • D ADDITIONAL BENCHMARK RESULTS: Evaluation protocols and score provenance are specified for the reported benchmark results, including leaderboard or original-paper baselines and the Nano Banana DPG-Bench score source.WISE evaluations use the legacy WiScore protocol with GPT-4o-2024-05-13 for Nano Banana, Qwen-Image∗, matched NL, and Ours.
  • D.1 DPG-BENCH PER-CATEGORY BREAKDOWN: Table 15 reports DPG-Bench performance across global scene description, entity presence, attribute binding, relations between entities, other dense-prompt aspects, and overall score.Published category breakdowns are shown for baselines, while Qwen-Image∗, matched NL, and Ours use matched per-category evaluations.
  • D.1 DPG-BENCH PER-CATEGORY BREAKDOWN: The DPG-Bench breakdown organizes evaluation around dense-prompt capabilities rather than only an aggregate score.Its categories explicitly include global scene description, entity presence, attribute binding, and inter-entity relations.
  • D.1 DPG-BENCH PER-CATEGORY BREAKDOWN: External code and model dependencies are listed by pipeline role together with available upstream license information.Unresolved dependency entries are marked explicitly.
  • D.2 GENEVAL PER-SKILL BREAKDOWN: Table 16 reports GenEval accuracy by single-object and two-object presence, counting, colors, position, color-attribute binding, and overall accuracy.The table presents published per-skill breakdowns for baselines and the authors’ evaluation for Ours.
  • D.2 GENEVAL PER-SKILL BREAKDOWN: Qwen-Image∗ uses its official prompt enhancement for the GenEval evaluation, consistent with Table 2.This protocol choice is stated alongside the per-skill comparison.

D.3 WISE PER-CATEGORY BREAKDOWN

Table 17 breaks down WiScore by category under the legacy WISE protocol, covering six domains and evaluating consistency, realism, and aesthetic quality with GPT-4o-2024-05-13. It compares multiple model groups using a weighted overall score derived from domain-level prompt scores.

  • The legacy WISE protocol uses GPT-4o-2024-05-13 to score consistency, realism, and aesthetic quality.
  • WISE evaluates six domains: cultural, temporal, spatial, biology, physics, and chemistry.
  • Models are grouped into dedicated T2I diffusion models, unified multimodal LLMs, proprietary systems, and open systems with prompt rewriting.
  • The official overall WiScore aggregates prompt-level scores using domain weights of 40%, 16.7%, 13.3%, 10%, 10%, and 10%.
  • Published baselines follow the official legacy leaderboard, while Nano Banana, Qwen-Image∗, matched NL, and Ours use the same legacy evaluator and scoring rule.

D.4 T2I-COREBENCH PER-CATEGORY BREAKDOWN … G ADDITIONAL QUALITATIVE EXAMPLES

The appendix reports per-category T2I-CoReBench evaluation details and documents the structured prompting, judging, and cold-start filtering prompts used by the system. These prompts enforce reasoned visual planning, explicit uncertainty handling, structured JSON outputs, and independent quality-control criteria.

  • D.4 T2I-COREBENCH PER-CATEGORY BREAKDOWN: Table 18 expands T2I-CoReBench comparisons across four Composition and eight Reasoning categories using the Gemini 2.5 Flash evaluator.Published leaderboard baselines are compared with Qwen-Image∗, matched NL, and Ours under the same matched evaluation harness.
  • E.1 COLD-START TEACHER SYSTEM PROMPT: The Cold-start teacher converts short image descriptions into a reasoned visual blueprint followed by a single structured JSON object.The required analysis proceeds through knowledge, reasoning, aspect ratio, composition, and spatial-constraint stages before emitting the JSON.
  • E.1 COLD-START TEACHER SYSTEM PROMPT: The teacher prompt requires forward-chaining decisions, separating general knowledge from image-specific design choices and preserving uncertainty for underspecified attributes.Concrete choices must be derived from reasons, while unspecified attributes are framed as reasonable visual implementations rather than necessities.
  • E.1 COLD-START TEACHER SYSTEM PROMPT: The structured blueprint specifies scene ratio, intent, atmosphere, style, lighting, elements, spatial relationships, and photography fields.Elements include normalized-grid positions, depth, captions, dynamic attributes, and action keys, with scene elements following the same structure.
  • E.2 AESTHETIC JUDGE SYSTEM PROMPT: The aesthetic judge evaluates visual quality independently of prompt alignment and structural accuracy, reports issues, and passes images with score >= 6.Its rubric covers composition, lighting, color harmony, overall impression, aspect-ratio fit, and technical defects.
  • E.3 COLD-START FILTERING SYSTEM PROMPTS: Validator v4 uses five independent Gemini calls to assess alignment, introduction quality, imagined-detail justification, reverse rationalization, and reasoning-stage boundaries.The resulting accept/reject statistics use a strict gate over these five axis-specific outputs.
  • E.3.1 PROMPT–IMAGE ALIGNMENT: The alignment filter distinguishes aligned, partial, and mismatch cases by checking named subjects, actions, settings, entities, and explicitly required elements.It does not treat unspecified style, lighting, framing, extra contextual elements, or image-quality defects as alignment failures.
  • E.3.2 INTRODUCTION QUALITY: The introduction and imagination filters accept added visual detail when it is framed as a choice, grounded in knowledge or context, or embedded in a coherent build narrative.They flag bald observational assertions and unsupported specifics because students lack access to the reference image and could learn unjustified hallucination patterns.

G.1 PROMPTER COMPARISON: PROMPTER SCALE × TRAINING · G.2 GENERAL-PURPOSE LLMS AS PROMPTERS

The section compares prompter scale and training under fixed image-generation conditions, then evaluates general-purpose LLMs against the trained prompter using a shared schema and Qwen-Image backbone. The comparisons isolate training, scale, and backend effects across common user prompts.

  • G.1 PROMPTER COMPARISON: PROMPTER SCALE × TRAINING: Four prompter variants compare two Qwen3.5 scales before and after the full SFT + Cold-start + RFT training pipeline.The variants use Qwen3.5-35B-A3B and Qwen3.5-397B-A17B, each shown in base and trained forms.
  • G.1 PROMPTER COMPARISON: PROMPTER SCALE × TRAINING: The Qwen-Image backbone and decoding settings remain fixed, so changing the prompter isolates training and scale effects.Reading within a scale isolates training, while reading across scales isolates prompter size.
  • G.1 PROMPTER COMPARISON: PROMPTER SCALE × TRAINING: Figures 20–22 present the scale-by-training comparison across rows corresponding to shared user prompts.The figures use a four-column layout, and aspect ratio is predicted by the prompter.
  • G.2 GENERAL-PURPOSE LLMS AS PROMPTERS: The general-purpose LLM comparison includes GPT-5.5, Gemini 3 Pro, GLM-5.2, Claude Opus 4.8, and the trained Qwen3.5-397B-A17B prompter.The four general-purpose models operate in single-turn schema-filling mode.
  • G.2 GENERAL-PURPOSE LLMS AS PROMPTERS: All evaluated LLMs fill the same schema and render with the same Qwen-Image backbone on shared user prompts.This setup keeps the schema, rendering backbone, and prompt rows common across prompters.
  • G.2 GENERAL-PURPOSE LLMS AS PROMPTERS: Figures 23–24 visualize the five-column LLM comparison, while Table 3 reports quantitative per-backend scores.The figures continue the same layout and shared-backbone comparison across the displayed prompt rows.
Loading 2607.29679v1…