acceptodds
Under review as a conference paper at ICLR 2027

InvariantCanvas: Evidence-Grounded Revision with Regression-Aware Acceptance

Abstract

Dense visual artifacts must satisfy exact textual and numerical requirements while preserving layout, visual relations, and readability. However, iterative generation-and-critique systems may introduce new violations when repairing local defects. We introduce InvariantCanvas, an evidence-grounded framework for reliable dense visual generation. The framework decomposes requests into verifiable requirements, generates editable HTML/SVG artifacts, and combines browser-based checks, OCR, source inspection, and visual audits with explicit provenance. A central adjudicator determines whether the available evidence justifies a revision. A revised artifact is accepted only when the targeted requirement changes from failure to success and all previously satisfied mandatory requirements remain satisfied; otherwise, the previous version is retained. A fresh delivery review then checks the exact image selected for release. InvariantCanvas achieves 95.86% overall accuracy and 94.83% accuracy on the Hard split of BizGenEval. On IGenBench, it achieves 0.94 question-level accuracy and 0.62 infographic-level accuracy. These results show the value of evidence-based adjudication and regression-aware revision for reliable dense visual generation.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
text rendering, information graphics generation, text to image generation workflow, dense visual generation

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_zpQoOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes InvariantCanvas, an agentic framework for reliable dense visual generation using editable HTML/SVG artifacts. It decomposes user requests into verifiable requirements, combines browser-based checks, OCR, source inspection, and visual audits, and employs an evidence-grounded adjudicator to guide revisions. Experiments on BizGenEval and IGenBench report improvements over existing image generation models.

Critical strengths
  1. The regression-aware acceptance mechanism provides a reasonable approach to preventing iterative revisions from introducing new errors.
  2. The paper reports strong results on two benchmarks, particularly on challenging cases requiring multiple constraints to be satisfied simultaneously.
Critical weaknesses
  1. The paper proposes a GPT-5.6-based agentic workflow but primarily compares against standalone image generation models. Even basic agentic baselines using GPT-5.6 or Claude Opus to generate HTML/SVG are missing.
  2. The paper provides no ablations isolating its core contributions, particularly evidence-grounded adjudication and regression-aware acceptance. Consequently, it remains unclear whether these mechanisms meaningfully improve performance beyond a standard iterative generation-and-refinement workflow.
  3. The framework involves multiple expensive LLM calls, rendering, verification, and iterative revisions, yet the paper reports neither inference costs nor end-to-end latency. Without such analysis, it is difficult to assess the practical value of the reported performance gains, especially relative to simpler approaches.
Questions for authors

Please refer to the weaknesses.

Reviewer_oQayOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents InvariantCanvas, an agent framework for generating and revising visual artifacts in HTML/SVG. The paper presentation quality is bad. For example, stranges arrows in the Fig. 1 seems to be directly generated by LLMs without basic human audits. The technical framework is mostly repeating exsiting works. The paper provides no ablation studies for the claimed new mechanisms, such as the evidence adjudication and the regression-aware acceptance. Therefore, I recommend a clear reject for this paper.

Critical strengths

No.

Critical weaknesses
  1. The paper figures are hard to read. Stranges arrows in the Fig. 1 seems to be directly generated by LLMs without basic human audits.
  2. The main technical framework presents limited novelty. It's hard to distinguish the proposed framework from previous works.
  3. The paper provides no ablation studies.
Questions for authors

See weaknesses.

Reviewer_fiH4OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes InvariantCanvas, which improves the correctness of infographics by decomposing user requests, generating editable html/svg, and combining broswer, ocr, source code, and visual checks. The core design is to apply comprehensive review after revision and accept the new version only when it fixes the targeted issue and doesn't break existing checks. The proposed method achieves leading results on both benchmarks against current image models.

Critical strengths
  • The framework explicitly separates evidence judgment from revision acceptance, addressing a core reliability problem in iterative visual generation. It first determines whether feedback identifies a targeted defect, then re-evaluates the complete revised artifact before allowing it to replace the previous version.

  • I like the idea of using an agentic system to solve the information-dense visual generation/editing tasks. Compared to pure image models, the proposed system is more verifiable and controllable, which can also benefit from recent trend of scaling post-training to improve the coding capabilities of LLMs.

Critical weaknesses
  • The experiments lack ablations and controlled comparisons that establish the contribution of the core method. The complete system is primarily compared against direct image models, without direct comparisons against LLM-based HTML/SVG generation, conventional iterative refinement, or other variants. Therefore, the gains may arise from the structured representation, stronger underlying models, or additional inference rather than the proposed decision mechanisms.
Questions for authors
  • Please refer to the weakness section.

  • Must a targeted requirement transition from FAIL to PASS, or is transitioning to UNKNOWN sufficient? How are previously passed mandatory requirements that become UNKNOWN, and disagreements between verifiers, handled?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.