Render-in-the-Loop
ECCV 2026

Render-in-the-Loop: Vector Graphics Generation via Visual Self-Feedback

A step-wise SVG generation paradigm where a multimodal LLM draws code, renders the partial canvas, sees its own visual state, and uses that feedback to generate the next primitive.

Guotao Liang, Zhangcheng Wang, Juncheng Hu, Haitao Zhou, Ziteng Xue, Jing Zhang, Dong Xu, Qian Yu†

Beihang University · The University of Hong Kong · 4Paradigm
† Corresponding author.
Paper arXiv BibTeX
Render-in-the-Loop SVG generation teaser.

Abstract

Current autoregressive SVG generators often draw blindly: they produce symbolic code without seeing the visual effect of previous strokes. Render-in-the-Loop turns SVG synthesis into an interleaved visual-code process. The model observes the prompt, its generated SVG fragments, and the rendered intermediate canvas, enabling it to reason over geometry, layering, occlusion, and completion.

Feedback

Visual Self-Feedback

Intermediate SVG states are rendered and fed back into the multimodal model as visual context for the next primitive.

Data

Dense Drawing Trajectories

Fine-grained path decomposition converts static SVGs into multi-step drawing traces with richer visual supervision.

Inference

Render-and-Verify

The decoder filters visually stagnant or repetitive primitives, reducing degenerate loops and redundant over-drawing.

Method

We model SVG generation as a drawing session: prompt, code fragment, rendered canvas, next code fragment. This gives the MLLM a synchronized drawing hand and seeing eye.

Decompose

Split long SVG paths into visually meaningful components while preserving topology and fill semantics.

Condition on Canvas

Train the model to generate each next primitive from both the target condition and the current rendered state.

Verify

Render candidate primitives before accepting them, rejecting code with negligible visual contribution or repetition.

Fine-grained path decomposition.
Fine-grained path decomposition makes intermediate visual feedback denser and more semantically useful.

Results

On MMSVGBench, Render-in-the-Loop achieves strong Text-to-SVG and Image-to-SVG performance while using substantially less data than several large autoregressive baselines.

127.64

Icon FID

Text-to-SVG on MMSVG-Icon. Lower is better.
0.293

Icon CLIP

Semantic alignment on MMSVG-Icon. Higher is better.
0.928

Illust. SSIM

Image-to-SVG reconstruction on MMSVG-Illustration.
0.178

Illust. LPIPS

Perceptual distance on MMSVG-Illustration. Lower is better.
Qualitative comparison with representative baselines.
Render-in-the-Loop produces cleaner SVG structures and better preserves semantic details compared with direct autoregressive and optimization-based methods.

Analysis

Visual feedback improves instruction following and completeness, while Render-and-Verify suppresses repetitive or visually empty drawing steps.

Instruction following examples.
Fine-grained prompts preserve attribute-object bindings such as color, count, and part-level details.
Ablation study.
VSF improves semantic completeness; RaV breaks repetitive loops and encourages meaningful progress.

Gallery

The model generates diverse vector graphics with clean topology and editable structure across icons and illustrations.

Gallery of SVGs generated by Render-in-the-Loop.

Citation

Please consider citing our work if you find it useful.

@inproceedings{liang2026render,
  title     = {Render-in-the-Loop: Vector Graphics Generation via Visual Self-Feedback},
  author    = {Liang, Guotao and Wang, Zhangcheng and Hu, Juncheng and Zhou, Haitao and Xue, Ziteng and Zhang, Jing and Xu, Dong and Yu, Qian},
  booktitle = {European Conference on Computer Vision (ECCV)},
  year      = {2026}
}