acceptodds
Under review as a conference paper at ICLR 2027

LayerCanvas: Infographic Generation in AI Learning Assistants via Hybrid Layer-Set Diffusion

Abstract

Online learning platforms are increasingly deploying AI assistants to provide personalized, on-demand tutoring. These can leverage infographics to communicate complex information and improve the learning experience. However, automating effective infographic generation remains a challenging task, requiring a combination of visual richness with precise typography, structured layouts, and continued editability. Existing representation methods satisfy these goals only in part: Scalable Vector Graphics (SVG) representations preserve geometry, hierarchy, and text but provide limited aesthetic flexibility, while raster diffusion models produce richer appearances by flattening content and layout into a single image. Conditioning a diffusion model on a rasterized SVG does not resolve this conflict because the raster preserves approximate placement but discards element ownership, occlusion relationships, and the structure needed to modify and recompose the result. We introduce LayerCanvas, a hybrid approach that retains SVG as the semantic, structural, and typographic program while using diffusion to aestheticize non-text content. Rather than generating one flattened raster, LayerCanvas partitions SVG elements into z-ordered sets of mutually non-overlapping elements and processes them through communicating parallel RGB generation streams. The resulting layers are independently editable and deterministically recomposed, while text is rendered from preserved vector metadata. Our approach combines rich visual synthesis with accurate typography and localized editing. LayerCanvas substantially improves text fidelity over state-of-the-art raster image generators while achieving the highest overall aesthetic quality. We further demonstrate non-destructive content editing, localization, asset reuse, and animation.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Diffusion models, controllable generation, editable infographics, parallel generation

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_esP1OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

LayerCanvas addresses the generation of visually rich, text-faithful, and editable infographics for AI learning assistants. It retains an SVG scene—elements, typography, hierarchy, paint order, boxes, and renderer-derived amodal/visible masks—as the structural source of truth, partitions non-text elements into z-ordered groups of mutually non-overlapping masks, and generates the corresponding raster layers in parallel using a frozen Ideogram-4 MMDiT backbone. A trainable adapter provides layer identity, graph-biased cross-layer communication, shared/private noise, and style/boundary residuals; layers are alpha-composited in SVG order, and exact vector text is rendered afterward.

Critical strengths
  1. A coherent representation for editable, text-faithful generation.
  2. Careful renderer-aligned supervision and split construction.
Critical weaknesses
  1. The main comparison does not isolate the contribution of layer-set diffusion. In Table 1, LayerCanvas receives the structural layer set and deterministic exact-SVG text compositing, whereas Qwen-Image-2512 and Ideogram-4 are prompt-only; much of the OCR advantage is therefore expected by construction rather than evidence for the learned layer model. The more comparable single-raster T2I-Adapter ablation in Table 2 actually has higher Char and Exact scores (0.851/0.568 versus 0.823/0.530), while LayerCanvas gains only 0.012 Recall and 0.22 Avg. Section 4.3 does not ablate shared noise, graph-biased communication, or the composition/hidden/boundary losses, and no recent layered-generation baseline is evaluated. Consequently, the central claim that the proposed learned machinery improves generation and editing over equally conditioned alternatives remains insufficiently established.
  2. The practical quality and editing claims are validated only indirectly. All quantitative results are on the authors' synthetic InfoLayers distribution, with no external human-authored SVG/infographic test set or blinded human preference study; the primary Claude judge is also used as the SVG design agent, although the additional judges in Table 5 mitigate judge-specific calibration. Editing evidence in Figures 4–6 and 13–16 is qualitative, without target-edit success, unrelated-region preservation, or occlusion-reveal measurements against editing baselines. This is consequential because reliable localized editing and superior real-world infographic quality are the paper's main practical value propositions, not merely static OCR on held-out synthetic layouts.
Questions for authors
  1. In Table 2, does Single-raster Conditioning use exactly the same training records, exact-SVG text compositing, prompts, backbone, and optimization budget as LayerCanvas? If so, how should the decreases in Char and Exact be interpreted, and what evidence isolates the benefit of layer-set generation from deterministic text/vector compositing?
  2. Can you add a like-for-like rasterized-SVG ControlNet/T2I-Adapter baseline with identical vector-text compositing to the test-set comparison and, where applicable, a recent layered-generation/editing baseline? The current prompt-only baselines in Table 1 do not receive LayerCanvas's structural input.
  3. Can you report component ablations for shared noise, graph-biased cross-layer communication, and the composition/hidden/boundary losses, with sample-level variability or confidence intervals? Section 4.3 currently changes the overall representation without showing which proposed learned components are necessary.
  4. Can you provide quantitative held-out editing results—target-edit success, unrelated-region preservation, and occlusion-reveal quality—against the single-raster model, together with either blinded human preferences or evaluation on external human-authored SVG layouts? These would directly test the editing and generalization claims that currently rest on qualitative examples and synthetic-data VLM scores.
Reviewer_s2JgOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

LayerCanvas aims to enable Diffusion models to generate infographics that are both aesthetically pleasing and text-accurate, and can be freely edited like PowerPoint.

Critical strengths
  1. LayerCanvas retains the SVG structure to handle Typography and Visual Generation separately, avoiding having the Diffusion Model undertake two different tasks of precise text rendering and visual generation simultaneously.
  2. Ordered Layer Set is a reasonable representation design. The author does not simply adopt the approach of one element corresponding to one layer, but utilizes the occlusion relationship of SVG to merge non-overlapping elements into the same generation stream. This processing method utilizes the structural information that already exists in the input data.
Critical weaknesses

The main concern is that the experimental evaluation does not sufficiently support the paper's central technical contributions. First, the comparisons with Qwen-Image and Ideogram are not entirely fair, as LayerCanvas benefits from explicit SVG structure, precise layout information, and deterministic text rendering, while the competing methods primarily rely on prompt-based conditioning. The improvements in text fidelity therefore cannot be directly attributed to the proposed layer-set diffusion architecture. More importantly, the paper lacks a comparison with a straightforward pipeline combining LLM-generated SVGs, Qwen-Image-Edit, and alpha compositing, which is essentially the pipeline used to construct the training dataset. Without this baseline, it remains unclear whether training a dedicated LayerCanvas model provides meaningful advantages over existing tools. Second, the paper introduces several components, including correlated noise, cross-layer attention, graph-based attention bias, and separate style and boundary residual paths, but provides no component-level ablation studies to demonstrate their individual contributions. In particular, there is insufficient evidence that the proposed cross-layer communication mechanism actually improves visual consistency compared with independently generated layers.

Questions for authors

Refer to Weaknesses

Reviewer_maN8OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

LayerCanvas generates editable infographics by combining SVG layouts and text with visual layers generated by diffusion. The model generates layers in parallel with cross-layer communication and composites them in SVG order. The paper also introduces the InfoLayers dataset.

Critical strengths
  1. Preserving SVG text provides exact typography and allows text edits without regenerating illustrations.

  2. The examples demonstrate useful applications.

Critical weaknesses
  1. The data pipeline already combines Claude Opus 4.8, Qwen-Image-Edit, and SVG composition to produce editable infographics. However, this complete pipeline is not evaluated as a baseline. A direct comparison with the full pipeline is needed to show what training the layer-set diffusion model adds in illustration quality, style consistency, and generation cost.

  2. Text accuracy comes from SVG rendering, and many editing operations follow directly from retaining separate layers. In Table 1, Illustration Appeal improves only from 2.92 for Ideogram-4 to 2.99 for LayerCanvas. Most of the aggregate gain comes from typography and semantic coverage. Table 2 compares layered and single-raster conditioning, but does not isolate cross-layer communication or its style and boundary paths.

  3. The dataset examples mainly show complete infographics. Figure 14 shows three simple groups, without detailed inspection of complex scenes.

Questions for authors
  1. How does LayerCanvas compare with the complete data-generation pipeline on the same held-out layouts in illustration quality, style consistency, and runtime?

  2. What changes when cross-layer communication is removed while keeping the layer representation and SVG text rendering fixed?

  3. Can the authors provide full RGBA decompositions for heavily overlapping examples, and evaluate edits that reveal hidden content or change layer order?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.