acceptodds
Under review as a conference paper at ICLR 2027

Beyond Starry Night: Shortcut-Aware Control-State Planning for Artist-Grounded Text-to-Image Generation

Abstract

Artist-grounded image generation requires more than appending an artist name to a prompt. Image models often respond to artist names through canonical shortcuts, such as recurring motifs, generic palettes, or overrepresented period signatures, rather than preserving the user's intended scene. We introduce Shortcut-Aware Artistic Control Planning (SACP), a formulation that translates underspecified artistic intent into an explicit control state separating scene anchors, preserve/transform decisions, style-regime hypotheses, role-bound artist evidence, and shortcut-avoidance constraints, and Atelier, its closed-loop runtime. Atelier grounds this state using artist-level knowledge and local patch references, compiles backend-aware generation plans, and iteratively refines candidates through global and local authenticity feedback. We further introduce ArtIntentBench, a benchmark covering Van Gogh and Qi Baishi across artwork re-rendering, period/style-controlled generation, historically unseen subjects, shortcut auditing, and human preference evaluation. Across open-weight and closed-source generators, Atelier improves artist-level style fidelity, preserves source structure more faithfully, and substantially reduces shortcut substitution compared with prompt-engineered, retrieval-augmented, and general-purpose agent baselines. These results suggest that artist-grounded generation is bottlenecked not only by image synthesis, but by the upstream inference of explicit, evidence-grounded artistic controls.

Primary area:
generative models
Keywords:
Multimodal Agent, Artist-Grounded Image Generation, Vague Intent Translation, Iterative Refinement, Style Authenticity Evaluation

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_TiHXOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper studies canonical shortcutting in artist-grounded text-to-image generation: conditioning on an artist name often causes a model to insert high-frequency motifs, palettes, or material treatments that are unsupported by the requested scene, instead of reliably performing a context-sensitive stylistic transformation. The paper introduces SACP, which represents underspecified artistic intent as an explicit control state z = (s, q, h, r, b), and Atelier, a closed-loop system combining an artist knowledge base, scene-role-bound local patches, backend-aware planning, a global critic, and a local AUTHCRITIC. ArtIntentBench evaluates Van Gogh and Qi Baishi across re-rendering, period control, historically unseen subjects, shortcut auditing, and human preference. The results show lower shortcut substitution rates and improvements in style proximity, structural preservation, and human preference. My main reservations are that the component-level methodological novelty is limited, scaling to new artists requires substantial artist-specific resources, and some evaluation axes share the same artistic ontology used to build the system.

Critical strengths
  1. The paper identifies a real failure mode. Outputs often substitute frequent artist associations for context-sensitive style. Formalizing this behavior as canonical shortcutting and quantifying it with a frozen taxonomy and majority-vote annotation is a clear and valuable contribution.
  2. Atelier integrates intent parsing, knowledge retrieval, local evidence binding, backend-aware planning, two critics, and reflective memory into a complete loop. The schema, prompts, thresholds, stopping rules, and ablations are documented, making the system auditable and the paper-level reproducibility details strong.
Critical weaknesses
  1. Limited component-level novelty. Atelier mainly combines LLM planning, retrieval, backend routing, critic-guided refinement, and memory, all common in agentic workflows. Its novelty lies more in the artist-grounded control representation and system integration; the paper should distinguish systems contribution from scientific novelty.
  2. A gap remains between an artist-agnostic schema and artist-specific instantiations. The experiments cover only Van Gogh and Qi Baishi, and each instantiation requires artist priors, curated summaries, an annotated patch bank, a dedicated AUTHCRITIC/rubric, and a frozen taxonomy. Qi Baishi demonstrates retargeting, but not low-cost or scalable generalization.
  3. Some evaluation axes are coupled to the system's artistic ontology. The retrieval resources, critic rubric, anti-shortcut constraints, taxonomy, and parts of the state audit share related artist-specific dimensions. Generic metrics and human evaluations mitigate this concern, but gains under independently constructed taxonomies, reference sets, or rubrics remain untested.
Questions for authors
  1. What is the core scientific contribution of SACP beyond an engineering schema? If the explicit factorization z = (s, q, h, r, b) is a testable inductive bias, what property or prediction can it provide that an unstructured planner with the same information and compute cannot?
  2. How can shortcuts be distinguished in principle from legitimate uses of canonical motifs? Could the authors give a taxonomy-independent counterfactual criterion? With the artist fixed and scene or period compatibility varied, a legitimate motif should track context, whereas a shortcut should remain driven mainly by the artist name. This would handle the contextual ambiguity of starry skies or inscriptions.
  3. Can the control dimensions be learned rather than manually specified? Could preserve/transform decisions, style regimes, retrieval targets, and anti-shortcut factors be learned from edit trajectories, expert critiques, or preferences while remaining interpretable? If they must be hand-specified, how can the cost of constructing s, q, h, r, b and the associated resources be controlled at scale?
3 comments

Q1: Response to concerns about the scientific contribution of SACP

Thank you for this insightful question.

1. Core scientific contribution

SACP formulates artist-grounded generation as semantically constrained style adaptation. SACP separates the content that must be preserved from the stylistic changes allowed by the request, and grounds those changes in artist evidence. The state z=(s,q,h,r,b)z=(s,q,h,r,b) provides a shared control representation for retrieval, planning, generation, evaluation, and revision.

The inductive bias is to preserve request-specific meaning while allowing evidence-supported stylistic variation. Preservation decisions and reference bindings guide what should appear, where stylistic changes should occur, and how those changes should be evaluated. Negative prompts help enforce these constraints.

2. Testable prediction and supporting evidence

Under matched information and computational resources, explicitly maintaining these decisions across stages should reduce context-unsupported substitutions relative to unstructured planning, while retaining the requested stylistic changes. This prediction concerns whether artist-associated content priors override the request when both planners have access to the same evidence.

Table 9 evaluates the structured-intent pathway as a whole. Removing this pathway increases SSR from 32.50% to 55.83%, with the generator pool held fixed. These results support the contribution of explicit intent modeling to canonical-shortcut reduction.

Q2: Response to concerns about legitimate motifs and shortcuts

Thank you for this insightful question.

Criterion. A canonical motif is legitimate when it fits the requested scene, period, and content. To examine this distinction, we can hold the artist and core subject fixed while changing whether the request supports or conflicts with that motif. Its use should respond to this change. Repeated insertion despite conflicting requirements, particularly when it replaces requested content, provides evidence of shortcutting. This criterion applies to any motif, including those outside our frozen taxonomy.

Examples. As stated in §1, a swirling sky is appropriate for a Van Gogh Saint-Rémy nocturne and is not inherently a shortcut. In an indoor close-up explicitly excluding any sky, inserting that motif would violate the request. Likewise, Qi Baishi inscriptions should be retained when requested and omitted when the user explicitly asks for a painting without text.

Supporting evidence. Our period-control experiment keeps the scene fixed across four requested periods for the same 57 requests, with W₂ measured against period-specific reference works. Atelier achieves the lowest W₂ in all four periods under both configurations. These results support responsiveness to historical context; the counterfactual criterion above specifies how to assess the appropriateness of individual motifs.

The objective is to preserve legitimate artistic expression while preventing familiar artist associations from overriding the request.

Q3: Response to concerns about learning control dimensions and scaling

Thank you for this insightful question.

What is specified, and what can be learned? The five control fields are shared across artists. The Perceiver and Planner infer and revise their values from the request, available evidence, and feedback, within predefined artist-specific rules. Individual requests therefore do not require manually written control states.

Learning from trajectories and preferences. The explicit representation provides concrete targets for supervision: expert feedback can correct a preservation decision, a style-regime hypothesis, an evidence binding, or an inappropriate constraint. Our recorded plans, failure tags, and repair targets could be combined with expert corrections or preference labels to train these decisions. Keeping the decisions and their evidence bindings explicit would allow the learned outputs to remain inspectable. This is a future training direction; the current system uses feedback during inference.

Controlling adaptation costs. New artists reuse the schema, workflow, annotation pipeline, and training recipes. VLM annotation and Grounded-SAM automate local evidence construction, while AuthCritic labels follow sample provenance. For Monet and Gauguin, annotation uses 14.03M/12.50M tokens, and synthetic-data generation plus adapter training requires 23.44/17.59 A800 GPU-hours, respectively. Artwork selection and validation of artist-specific knowledge and evaluation rules remain necessary.

The same control fields can be reused across artists and serve as targets for future supervised learning.

Thank you for the clarifications. I still view the contribution as primarily engineering-focused, with limited methodological novelty, though it may offer useful insights for future research on artist-specific style generation. I will keep my score there but would be fine with rejection. I would also welcome the other reviewers’ perspectives.

Thank you for the follow-up. Our methodological claim concerns how a request and artist references are translated into a generation plan.

1. What SACP adds to planning. An artist’s painting contains both stylistic properties and particular subjects. Using it as a reference requires deciding which properties to transfer and which parts of the requested scene to preserve. SACP represents these decisions explicitly and assigns retrieved evidence to specific scene roles.

For a library scene, for example, a reference may guide the palette and brushwork of a surface, while the furniture and spatial layout remain preservation targets. The reference therefore has a defined role in the plan: it supports a particular change to the requested scene.

2. How these decisions guide generation and revision. The generation plan and the critic use the same preservation requirements and reference bindings. Critic feedback identifies violations and repair targets; the next round updates the plan and evidence bindings while retaining the original request. SACP connects what should change, the evidence supporting that change, and the feedback used to revise it. This is the planning procedure whose contribution we aim to establish.

3. Evidence from the ablation. With the generator pool, sampling settings, and three-round limit held fixed across three seeds, removing the structured-intent pathway increases SSR from 32.50% to 55.83% and W₂ from 82.11 to 85.23. This supports the usefulness of the proposed planning decisions within Atelier.

Reviewer_h6gcOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies a problem with text-to-image models: asking for an image “in the style of <artist>” often brings in familiar motifs, such as Starry Night skies for Van Gogh. The authors call these shortcuts and measure how often they appear. They introduce a method which expands the user's prompt into a structured set of controls. These specify the main elements of the scene, what to keep or change, the likely artistic period, and patterns to avoid. Their system, Atelier, retrieves information about the artist and reference patches linked to specific visual roles. It uses a pool of image generators to produce candidates. A trained critic checks the authenticity of local patches, and a vision-language model also evaluates the results. The system can repeat generation and feedback up to three times. The authors evaluate it on a new benchmark that measures shortcut rates (SSR), similarity to the source, and style distances.

Critical strengths

S1. The paper defines a fixed list of shortcut patterns for each artist and measures how often they appear. The authors finalize these lists before evaluating their method. Table 2 reports the rate of each pattern for every method. Three annotators label the results, and the majority vote determines each label. This provides a clear basis for future comparisons. I expect this part of the paper to be useful to later work.

S2. The evaluation across artistic periods is a good idea. The scene stays the same, but the requested period changes. The style distance W2 is measured against reference works from that period only (Section 4.1, Table 3). This tests whether a method keeps using the same familiar look across different periods. General style scores or CLIP scores cannot detect this problem.

Critical weaknesses

W1. The experiments do not clearly show what causes the improvements. Atelier has several advantages that the baselines do not have: a pool of generators, a trained critic that selects among multiple candidates, and up to three rounds of generation and feedback. Its critic prompt also lists the shortcut patterns counted by SSR (Prompts 10 and 13). Each baseline uses a single generator without a critic. As a result, the experiments do not directly test the main claim that the structured controls reduce shortcuts. The ablations that keep the generator pool fixed (Tables 9 and 10) report style distance, but not SSR. The remaining improvements are small, and the paper does not report results across random seeds or their variation. I believe that the current comparisons therefore cannot separate the benefit of the controls from the benefit of selecting candidates and choosing generators. W2. The contribution is quite narrow and depends heavily on the two chosen artists. Most of what is added to earlier text-to-image agent systems is specific to each artist. Each artist needs a curated knowledge base, a list of shortcut patterns, a critic adapter trained on their paintings, critic evaluation rules, and rules for combining feedback (Section 3.4, Appendix A.3). Van Gogh also has a FLUX style LoRA. Much of this would need to be rebuilt for a new artist, which limits the method's use. The authors acknowledge that they test only two artists, but say that their method can be adapted simply by replacing the data. Yet Qi Baishi required a separately trained critic, custom evaluation rules, and a custom policy for combining feedback.

Questions for authors

Q1 [Requested experiment] Please report shortcut rates (SSR), as well as style distance (W2), for the rows in Table 9 and the prompt-expansion rows in Table 10. Q2 [Clarification] How was the FLUX2-LORA adapter trained? Which artworks and periods were used? Do any of these artworks also appear among the re-rendering source images or the W2 reference sets? Does the direct FLUX baseline use this adapter too? Please also report results after removing the adapter from Atelier's generator pool. Q3 [Clarification] The Qi Baishi paragraph says Atelier performs best on “two of three closed-source patterns” (p. 8). However, Table 2 shows it performs best on only one. Please correct the text or the table.

4 comments
AuthorsOpenReview ↗

Response to Reviewer h6gc

Q1: Response to the request for ablation results

Thank you for this insightful suggestion.

The results are shown in the tables below. We report SSR and W₂ using three seeds: 42, 123, and 2026, with the generator pool and sampling settings held fixed within each comparison. Across the component-ablation and prompt-construction experiments, W₂ standard deviations range from 0.48 to 1.04.

Component ablations

ConfigurationSSR ↓W₂ ↓
Full Atelier32.50%82.11
Without Artist KB & Patches37.50%86.48
Without Iterative Writeback44.17%84.77
Without Structured Intent55.83%85.23
Without Both Feedback53.33%85.20

Full Atelier achieves the lowest W₂ and SSR. Removing structured intent produces the largest SSR increase (23.33 percentage points), followed by removing both feedback channels (20.83 points). These results support the contribution of structured intent and feedback to shortcut reduction within the same generator pool.

Prompt-construction comparisons

In these reruns, the prompt-construction baselines share Atelier's generator pool, critic feedback, final-selection rule, and three-round limit; the original comparison used single-pass expansion.

ConfigurationSSR ↓W₂ ↓
Free-form expansion46.15%80.05
Expert-template expansion40.71%78.52
LLM-proposed references45.19%80.87
Atelier (prompted critic)33.33%77.74
Full Atelier29.85%75.97

Full Atelier achieves the lowest SSR and W₂. Compared with expert-template expansion, the strongest prompt-construction baseline, SSR decreases by 10.86 percentage points and W₂ by 2.55. Relative to the prompted-critic variant, SSR decreases by 3.48 points and W₂ by 1.77.

Q2: Response to concerns about FLUX LoRA

Thank you for this insightful question.

1. Training data and procedure

The Van Gogh adapter was trained on 101 complete artworks spanning Paris, Arles, Saint-Rémy, and Auvers. Each artwork was paired with a caption containing its title, artwork analysis, year, and artist-style description. We adapted FLUX2.Klein-9B using LoRA with rank 16, alpha 4, a learning rate of 8 × 10⁻⁵, and a maximum of 1,800 training steps.

2. Overlap with evaluation data

The 101 training artworks are disjoint from both the 100 re-rendering source artworks and the W₂ reference sets.

3. Direct FLUX baseline

The direct FLUX baseline uses the base model without the artist-specific LoRA adapter.

4. Results without the adapter

Following the Table 9 protocol (40 cases and three seeds: 42, 123, and 2026), we retain the FLUX backend and disable only its artist-specific adapter. SSR changes from 32.50% to 33.33%, and W₂ from 82.11 to 83.27. Both remain lower than the corresponding values in the control-component ablations (SSR: 37.50–55.83%; W₂: 84.77–86.48). Thus, Atelier retains its advantage over these ablations when the artist adapter is removed.

Q3: Response to concerns about the Qi Baishi results

Thank you for pointing out this error.

We have corrected the text to state that Atelier achieves the lowest shortcut rate in one of the three closed-source categories, and updated the corresponding highlighting in Table 2.

Thank you for the additional experiments and clarifications. They strengthen the evaluation, but my main concern remains: the contribution is narrow and relies heavily on artist-specific resources and adaptation. I therefore maintain my score of 2 (Weak rejection). I look forward to hearing the other reviewers’ views.

AuthorsOpenReview ↗

Additional Comments

Thank you for the follow-up. We agree that Atelier needs per-artist resources, and we do not want to understate that. Our first response did not address this point directly, so we give the evidence here.

1. What is rebuilt for a new artist, and what it costs (Appendix C.1). We adapted Atelier to two further painters, Monet and Gauguin. The Perceiver, Planner, Global critic, retriever, Executor, Memory, and FinalSelect kept their structure, prompt templates, output formats, and runtime logic. Three resources were rebuilt: the knowledge base, the AuthCritic adapter, and the FLUX style adapter. No manual labeling was involved: a multimodal LLM describes each artwork, Grounded-SAM extracts the patches, the LLM annotates them, and AuthCritic labels follow from each patch's source. The artist-specific instructions are also LLM-generated. Per artist, annotation cost USD 6.68 / 5.84, and synthetic-image generation plus adapter training took 23.44 / 17.59 A800 GPU-hours. With these resources swapped in, Atelier has the lowest W₂ on both artists (Monet: 65.25 vs. 67.06 for the best baseline; Gauguin: 76.10 vs. 81.19; Table 14).

Every per-artist step after collecting the artworks is therefore performed by an LLM or a fixed training recipe, so the adaptation can be run as an automated procedure in under one GPU-day of compute. This is the same kind of requirement as the per-concept training used by personalization methods such as DreamBooth, Textual Inversion, and StyleDrop.

2. Why the method uses artist-specific evidence at all. Our goal is fidelity to a real painter's visual language (period, brushwork, material handling), not generic stylization. Generators that rely on the artist name alone are the ones that fall back on canonical cues: 47–77% of baseline Van Gogh outputs contain an unrequested shortcut (Table 2). Evidence about what the painter did beyond a few famous works is what removes that fallback. What we claim to be general is the control schema and the closed loop, which now run on four painters across oil and ink traditions. For the same reason, our blinded preference study includes painting majors (n = 5) as well as non-art majors (n = 10), and both groups give the same method ordering (Appendix B.9).

3. Qi Baishi. You are right that this instance has its own critic adapter, its own rubric, and an artist-conditioned merge rule (Appendix A.3). The output contract and score aggregation are shared with Van Gogh (Appendix A.4.2), and the Monet and Gauguin adaptations required no change to the runtime logic.

4. What we will change. We will move the C.1 summary and cost figures into the main text, and replace "retargeted by replacing artist-specific data" in Section 5 with an explicit list of what is rebuilt per artist. [C.1 currently reports W₂ only; shortcut rates for the two new artists need a frozen, human-annotated taxonomy, which is an evaluation cost and not a requirement of the method.]

Reviewer_XK5tOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates canonical shortcutting in artist-style text-to-image generation, where models inject iconic artistic elements that conflict with the user's prompt. To address this, the authors propose SACP and its operational system, Atelier. The system achieves memory-guided iterative generation through explicit control states, an artist knowledge base, patch-level evidence binding, and a dual-level (global/local) critic mechanism. Evaluations on ArtIntentBench across Van Gogh and Qi Baishi styles cover original painting redrawing, period-sensitive control, and unseen subject generation.

Critical strengths
  • Clear Problem Definition: The paper accurately identifies the conflict between artistic style expression and the preservation of user-specified content. By highlighting the context-dependent nature of shortcut identification, the work demonstrates clear research value.
  • Comprehensive System Design: The explicit control state effectively structures preserve/transform decisions, period hypotheses, reference bindings, and anti-shortcut constraints. The clear task division between the global critic and the local AUTHCRITIC is well-conceived.
  • Rich Experimental Evaluation: The experiments encompass shortcut auditing, control-state analysis, component ablations, and human preference studies. Validations across two distinct artistic traditions (Western oil painting and Chinese ink wash) provide initial evidence of generalizability. Furthermore, Tables 9 and 10 partially rule out the possibility that performance gains stem solely from prompt expansion or iteration.
Critical weaknesses

Limited Evidence of Generalizability: The evaluation covers only two artists, which is insufficient to fully substantiate the "artist-agnostic" claim. The data and manual effort required to adapt the framework to new artists remain unclear.

  • Lack of Efficiency and Cost Analysis: The system relies on multi-model calls, retrieval, and iterative evaluations, yet lacks detailed reporting on end-to-end latency, token consumption, and compute costs. Consequently, the trade-off between performance gains and computational overhead is difficult to assess.
  • Weak Implicit Period Inference: In Table 7, the Top-1 period recovery rate for Van Gogh text redrawing is only 38–40%. This bottleneck in implicit period inference raises concerns, and its downstream impact on final generation quality is not sufficiently analyzed.
  • Need for Stronger Statistical Rigor: Main results lack confidence intervals or variance over multiple runs. The ablation study covers only 40 cases, and the human evaluation includes only 15 participants (including 5 experts), leaving the statistical stability of the conclusions open to further verification.
  • Unclear Budget Fairness: While Tables 9 and 10 control for certain generation resources, it remains vague whether the "equal generation budget" accounts for retrieval, planning, critic calls, and candidate selection overhead. Minor Comments: The prompt text in Figure 1 ("Create a painting of a quiet library corner with a wooden chair in the style of Vincent van Gogh.") should be rendered as selectable vector text rather than an image. Additionally, the layouts of Table 2 (too sparse) and Table 3 (too dense) should be polished for better readability.
Questions for authors
  • Which system components can be directly transferred to a new artist, and which require rebuilding or retraining? Could the authors provide validations on additional artists along with the associated data and annotation costs?
  • Could the authors report the end-to-end latency, token consumption, number of model calls, and total monetary cost for generating a final image, broken down by major system modules?
  • How are the ground-truth labels, candidate set, and evaluation stage defined for the period recovery rate in Table 7? How severely does incorrect period inference affect final output quality?
  • Could the authors supply confidence intervals, statistical significance tests, and multi-run variance? How does the statistical analysis account for repeated subjects and repeated ratings from the same participant?
  • What specific resources are included in the "equal generation budget"? Could the authors provide a clear breakdown of baseline resource configurations, alongside comparisons under similar overall compute/time budgets?
5 comments
AuthorsOpenReview ↗

Response to Reviewer XK5t

Q1: Response to concerns about cross-artist generalization

Thank you for this insightful question.

1. What transfers directly, and what requires adaptation?

Atelier uses the same framework across artists, updating artist-specific resources and adapters.

Reused components: The Perceiver, Planner, Global critic, Retrieval module, Executor, Memory, and FinalSelect retain their shared structure, prompt templates, output formats, and runtime logic. Artist-specific instructions are populated with model-generated text, and the Global critic requires no fine-tuning.

Adapted components: The artist knowledge base is rebuilt, and the AuthCritic and FLUX style LoRA adapters are retrained on the new artist's data using the same base models and training configurations. These changes are needed because artists differ in composition, palette, brushwork, and the visual cues that distinguish authentic work from imitation.

2. What data, annotation, and computation are required?

Data: The Monet and Gauguin datasets contain 224 and 160 artworks, respectively, with 659/603 reference patches and 3,009/3,006 AuthCritic examples.

Annotation: We use a multimodal LLM to annotate both complete artworks and entity patches, without manual labeling. The LLM first describes each artwork and identifies its entities. Grounded-SAM segments these entities, after which the LLM annotates each patch’s content, brushwork, palette, and material characteristics. AuthCritic’s source-class labels are assigned from the known origins of the patches: authentic works, synthetic images, or other artists’ works. For Monet/Gauguin, LLM annotation uses 14.03M/12.50M tokens and costs USD 6.68/5.84, respectively.

Computation: Adapter training takes 5.47/5.32 A800 GPU-hours, and synthetic-image generation takes 17.97/12.27, giving a measured subtotal of 23.44/17.59 A800 GPU-hours per artist.

3. Validation on additional artists

We extend the paper's Van Gogh and Qi Baishi study to Monet and Gauguin using the same control structure and closed-loop workflow.

MethodMonet W₂ ↓Gauguin W₂ ↓
FLUX.2 Klein base 9B67.0681.19
Qwen Image67.3396.47
LongCat68.7781.57
Hunyuan Image 3.072.1181.56
Atelier65.2576.10

Atelier achieves the lowest IntroStyle W₂ among the compared methods for both Monet and Gauguin. These results support Atelier’s ability to adapt to artists beyond Van Gogh and Qi Baishi by updating artist-specific resources and adapters.

Q2: Response to concerns about end-to-end resource costs

Thank you for this insightful question.

We report average resource use per final image across Perceive, Plan, Execute, and Evaluate, including all candidate generations and evaluation rounds. Values are shown as open-source / closed-source.

Open Execute reports cumulative local-generation time. The open-stage values sum to 1,104.9 s (18.42 minutes), and the closed-stage values sum to 1,007.32 s (16.79 minutes).

GPU allocation: Local GPU time uses an equal-share estimate for two open-source/four closed-source runs sharing one A800, assuming the GPU intervals are shared throughout. The unallocated totals of 15.71/8.55 GPU-min are divided by 2/4, giving 7.86/2.14 A800 GPU-min per final image. Request times, token consumption, model calls, and API charges are not divided by concurrency.

ModuleAverage model calls(open / closed)Tokens, K (open / closed)Reported time, s (open / closed)API cost, USD (open / closed)Estimated amortized GPU time, A800 GPU-min (open / closed)
Perceive1 / 19.4 / 8.753 / 600.0087 / 0.00310 / 0
Plan3 / 3.556.7 / 50.2431 / 4530.0414 / 0.02190 / 0
Execute9.1 / 43.0 / 8.4218.9 / 174.320.0630 / 0.12351.83 / 0
Evaluate9.3 / 4.598.7 / 22.2402 / 3200.0724 / 0.03066.03 / 2.14
Total22.4/13.0167.8 / 89.51,104.9 / 1,007.320.1855 / 0.17927.86 / 2.14

Q3: Response to concerns about period recovery and output quality

Thank you for this insightful question.

  • Ground truth: The artistic period recorded in each source artwork’s metadata.
  • Candidate set: Six periods of Van Gogh’s career: Early Dutch, Nuenen, Paris, Arles, Saint-Rémy, and Auvers. The system ranks these six periods.
  • Evaluation stage: Table 7 evaluates the Perceiver’s initial period ranking. Its input is a textual description derived from the source artwork, without access to the source image.

Impact on final quality: The re-rendering results in Table 3 include cases with incorrect period predictions. Atelier achieves the best W₂, DreamSim, DINO, and LPIPS results under both configurations. Period matching contributes to the retrieval score rather than imposing a hard filter.

Q4: Response to concerns about statistical reliability

Thank you for this insightful question.

User study and repeated ratings: We recruited 15 participants: 5 painting majors and 10 non-art majors. Each participant ranked all five methods on the same 35 questions, yielding 15 × 35 = 525 complete ranking responses. We resample whole participant records and perform paired tests at the participant level, preserving repeated ratings within each participant. Atelier achieves the highest first-choice rate, 41.52% (95% CI: 35.24%–47.81%). All 15 participants rank Atelier above Qwen and FLUX on average, and 13 of 15 rank it above Nano Banana. These comparisons remain significant after Holm correction (p = 0.000244, 0.000244, and 0.0148, respectively). Painting majors and non-art majors produce the same overall method ordering.

Main experiments: We report pointwise 95% confidence intervals and Holm-adjusted p-values within each setting. Open-source Van Gogh SSR decreases by 15.64 percentage points relative to the strongest baseline (95% CI: −23.18 to −8.10; p = 0.0004). Re-rendering DINO improves under both configurations (p = 0.0054/0.0004), and open-source LPIPS improves (p = 0.0397). W₂ differences and the closed-source SSR difference do not reach adjusted significance. Open-source re-rendering CLIP is significantly lower (p = 0.00175).

Multi-run variability: We repeated the experiments in Tables 9 and 10 using three seeds: 42, 123, and 2026. W₂ standard deviations range from 0.48 to 1.04, smaller than the mean W₂ differences of 1.77–4.90 between full Atelier and the comparison settings. These results support the stability of Atelier’s improvements under the tested seed settings.

Q5: Response to concerns about budget fairness

Thank you for this insightful question.

1. What specific resources are included in the “equal generation budget”?

Claude, GenArtist, and Atelier share a maximum of three generation rounds. In the component-ablation and prompt-construction comparisons, the compared settings additionally share the generator pool; the prompt-construction baselines also share the critic and final-selection rule. These are the matched generation constraints. Module-level time, token consumption, API charges, and local GPU use are reported in Q2.

2. Could the authors provide a clear breakdown of baseline resource configurations, alongside comparisons under similar overall compute/time budgets?

Claude and GenArtist also evaluate images and use feedback to regenerate them. After generation, each agent inspects the candidate and provides an evaluation. If the image does not pass, the agent uses this feedback to generate a revised candidate, continuing until acceptance or the three-round limit. The configurations used in Tables 2 and 3 are summarized below.

MethodGenerator(s)Maximum roundsBetween-round feedback
Direct generationOne backend per baseline1None
GenArtistGPT-Image-2, Nano Banana Pro3Image inspection, feedback, and revision
ClaudeGPT-Image-2, Nano Banana Pro3Image inspection, feedback, and revision
Atelier, closed-sourceGPT-Image-2 and Nano Banana Pro3Global critic + AuthCritic
Atelier, open-sourceFLUX LoRA, Qwen Image, LongCat, and Hunyuan Image 3.03Global critic + AuthCritic

In Table 2, Atelier's closed-source SSR is 51.96%/43.75% on Van Gogh/Qi Baishi, compared with 59.22%/55.00% for the strongest GenArtist configuration (Nano Banana Pro) and 75.70%/68.75% for the strongest Claude configuration (GPT-Image-2). Atelier achieves lower shortcut rates against agents that also inspect and revise images under the same maximum round budget.

Minor comments

Thank you for these helpful suggestions.

  • Figure 1: We have replaced the rasterized prompt with selectable vector text.
  • Tables 2 and 3: We have reduced excess whitespace in Table 2 and adjusted the headers and spacing in Table 3 to improve readability.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Thanks for the feedback

We thank all reviewers for their careful and constructive comments. This note summarizes the points raised by more than one reviewer and where the revised paper addresses them. Detailed answers are in the individual replies.

Scope of the work.

Atelier targets faithfulness to a specific painter, not generic stylization. For this goal, an unrequested Starry Night sky or a wrong period is an error, and current systems make it often: 47–77% of baseline Van Gogh outputs and 40–79% of baseline Qi Baishi outputs contain an unrequested canonical cue (Table 2). Faithfulness is the hard part of the problem and the contribution of the paper: defining and measuring shortcuts, and the control state that reduces them. Adapting the system to a further painter is the inexpensive part: data preparation with no manual labeling plus adapter training, at about USD 6–7 of annotation and under one GPU-day per artist (point 1). Two uses follow directly. For research, the frozen shortcut taxonomies and SSR give a way to audit any text-to-image model for canonical shortcutting, including on non-Western traditions, where baselines replace Qi Baishi's ink language with generic decorative conventions (Tables 1–2). For practice, art education is a natural application, for example showing one scene across a painter's periods (Section 4.3), where correctness about the painter matters more than speed.

1. Generalization to new artists (Appendix C.1).

Van Gogh and Qi Baishi were chosen as a deliberately distant pair: European oil painting and Chinese ink painting differ in medium, tradition, and control challenges (Section 4.1). The same control schema and loop lower SSR below the strongest baseline for both painters in both configurations (Table 2). To test adaptation further, we added Monet and Gauguin.

The seven runtime modules (Perceiver, Planner, Global critic, retriever, Executor, Memory, FinalSelect) transfer with unchanged structure, prompt templates, output formats, and runtime logic. Only the knowledge base, the AuthCritic adapter, and the FLUX style adapter are rebuilt, using LLM-generated artist instructions and no manual labeling, at USD 6.68 / 5.84 of annotation and 23.44 / 17.59 A800 GPU-hours per artist. Atelier has the lowest W₂ on both artists (Table 14). Because every LLM-based module is specified by a prompt template and a fixed output schema (Appendix D), the pipeline can be packaged as reusable agent skills. We will replace the "artist-agnostic" wording with this explicit split between shared and per-artist parts.

2. Cost per image (Appendix C.4, Table 16). Per final image (open-weight / closed-source), Atelier uses 22.4 / 13.0 model calls, 167.8K / 89.5K tokens, USD 0.19 / 0.18 in API charges, and 7.86 / 2.14 A800 GPU-minutes. Summed stage time is 18.4 / 16.8 minutes, about three quarters of it in planning and evaluation. Table 16 gives the breakdown by module.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.