- Rating
- 1
- Confidence
- 3
- Informativeness
- 1
- Presentation
- 2
- Soundness
- 1
Read review
Paper summary
The paper presents DesignX, which has two parts.
(i) A template-based design agent. The agent tags an internal template library along seven facets: Type, Scenario, Purpose, Domain, Subject, Visual Style and Color Palette. It retrieves five candidates coarse-to-fine and adapts the selected template to the user’s request, producing raster designs.
(ii) A framework that converts raster designs into editable elements. A VLM (Qwen3.6-35B-A3B) parses a design into typed elements. Training uses 2.2M real layered designs and 2.1M synthetic collages.
Evaluation. The agent is evaluated by human preference against GPT-Image-2 (82.96% preferred). The decomposition is evaluated on a new 150-image benchmark against Qwen-Image-Layer, MagicLayer and GPT6-Astra Ultra. The metrics are LPIPS and valid-layer counts judged by GPT-6 Astra.
Critical strengths
- The problem matters in practice. The method recovers editable text and simple shapes, not only image layers.
- The font prediction is the strongest part. A dedicated classifier picks among 6,279 font families instead of asking the VLM. A normalization (Eq. 1–2) keeps size and spacing predictions valid when crops are resized.
- The training data is large. The ablation suggests that real and synthetic data complement each other (Tab. 2), and the comparisons include commercial systems.
Critical weaknesses
1. The core design choices are not justified, and the agent does not serve its stated goal.
- The paper generates designs as images and then converts them into editable code. It never explains why this beats producing editable designs directly, which is an established alternative in research (e.g., COLE, ART, StarVector, OmniSVG) and in commercial tools (e.g., Canva's design model, Recraft V4, Claude Design). Even the image-first workflow already exists (Accordion, ICCV 2025).
- The agent, framed as "skills", is retrieval-augmented generation over a fixed internal library. Its stated goal is to help users who "may struggle to articulate their preferred visual style" (l. 78–79). Yet it narrows the library by keyword-matching the user's request against seven broad tags (l. 229–238). Those tags leave out layout, density, text capacity, typography and imagery.
- Beyond picking one of five templates, nothing captures or learns the user's preferences. Work that targets this problem directly is not discussed (e.g., Creo, DesignScape, Sequential Gallery, AesthetiQ, DesignPref).
the pipeline is the paper's main contribution. It is not justified against direct editable generation, and it has no mechanism that models what users want, so it is unclear what the system does better than existing approaches.
2. The evaluation does not support the claims.
- The agent is evaluated with a single A/B test. It is compared only against plain GPT-Image-2 (Tab. 1).
- The test reports no criteria, rater details, agreement or confidence intervals.
- No baseline isolates the effect of retrieval, such as a random template, embedding-based retrieval or filling the template directly.
- There are no side-by-side outputs from the compared methods for the same prompt. There are also no template/output pairs that would show whether a template is adapted or simply copied, and the paper has no failure cases or appendix..
- The full system is never tested as a whole. The preference study judges images only, and the decomposition is tested on a separate set of 150 images. Nothing evaluates the final editable file.
- The decomposition numbers are weak. They are counts of "valid" layers from an LLM judge that is not checked against humans, and several reported numbers contradict the text.
the main claims rest entirely on this evidence: the 82.96% preference, outperforming commercial systems, and bridging generation and editing. As presented, the evidence cannot separate the method from simpler alternatives, and readers cannot inspect the outputs.
Questions for authors
- Why generate raster designs and then decompose them, instead of generating editable designs directly? How does DesignX differ from Accordion (ICCV 2025), which also converts AI-generated designs into editable layers?
- How many templates does the library contain, and how were they selected? Beyond the tags, what does the “careful inspection” step (l. 237–238) consider, and does the system use the user’s choices or feedback in any way?
- How many raters and judgments were collected for Tab. 1, with what expertise and agreement, and how does the agent compare with a random template of the same type? Could you show side-by-side outputs (prompt, retrieved template, DesignX, GPT-Image-2) for randomly sampled requests, including failure cases?
- Could you define in the paper the model variants (DesignX V1–V3, Text Predictor V1–V8) and reconcile the numbers that contradict the text? These include the vector-layer claim at l. 376–377 (GPT6-Astra Ultra has 2,345 vs 2,233), the V1 total in Tab. 2 (4,294 vs 4,744), the LPIPS of the same model in Tabs. 2 and 3 (0.1400 vs 0.1382), and the claim at l. 454–455, which Tab. 4 contradicts.