Research map / Image generation / Autoregressive and latent generation
Vector Graphics Generation via Mutually Impulsed Dual-domain Diffusion
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes a dual-domain (vector and pixel) diffusion framework where each domain's denoiser is conditioned on the other via cross-attention, with a shape prior regularizer enforcing smoothness and closure, for vector graphics generation on font and icon datasets.
Problem
Vector graphics generation lacks pixel appearance supervision, while two-stage generate-then-vectorize methods suffer from inaccurate gradients through differentiable rasterization.
Approach
Jointly diffuse vector instruction sequences and their rasterized pixel images in synchronized parallel denoisers, with cross-attention injecting each domain's representation into the other's sampling step, plus independent supervision per domain and a differentiable shape prior regularizer (smoothness and closure) instead of differentiable rasterization.
Key results
- On the Icon dataset, the method reaches FID 6.09 and VD 0.0119, versus 15.90 / 0.0195 for the two-stage baseline and 27.42 / 0.0482 for DeepSVG.
- On the Font dataset, FID 6.66 and VD 0.0066, versus 8.66 / 0.0190 for DeepSVG.
- Ablation on Icon: removing cross-domain injection (pure vector diffusion) raises FID from 6.09 to 7.70 and VD from 0.0119 to 0.0214.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use dual-domain diffusion with cross-domain conditional information injection (ablation) | Vector graphics generation · Icon | FID: 7.70 -> 6.09 | Helped |
| Use dual-domain diffusion with cross-domain conditional information injection (ablation) | Vector graphics generation · Icon | VD: 0.0214 -> 0.0119 | Helped |
| Add Shape Prior Regularizer (ablation) | Dual-domain diffusion · Icon | FID: 6.37 -> 6.09 | Helped |
| Add Shape Prior Regularizer (ablation) | Dual-domain diffusion · Icon | VD: 0.0133 -> 0.0119 | Helped |
| Use the proposed dual-domain diffusion method instead of Two-stage (Image Diffusion + LIVE) | Vector graphics generation · Icon | FID: 15.90 -> 6.09 | Helped |
| Use the proposed dual-domain diffusion method instead of Two-stage (Image Diffusion + LIVE) | Vector graphics generation · Icon | VD: 0.0195 -> 0.0119 | Helped |
| Use the proposed dual-domain diffusion method instead of Deepsvg | Vector graphics generation · Font | FID: 8.66 -> 6.66 | Helped |
| Use the proposed dual-domain diffusion method instead of Two-stage (Image Diffusion + LIVE) | Vector graphics generation · Font | VD: 0.0120 -> 0.0066 | Helped |
Limitations
- Evaluation relies on CLIP-feature-based FID and VD metrics rather than direct shape or human measures.
- Icon dataset is a custom subset that caps sequence length at 96 and filters categories, limiting comparability.
Experimental setup
Datasets: Font dataset (DeepSVG) Icon dataset (derived from FIGR-8-SVG)
Benchmarks: Font dataset Icon dataset
Metrics: FID Visualization Distance (VD)
Baselines: DeepSVG Image Diffusion + LIVE Im2Vec Pure Vector Diffusion
Builds on: DDPM DeepSVG LIVE
Closest papers by meaning
- SuperSVG: Superpixel-based Scalable Vector Graphics Synthesis (CVPR 2024 · similarity 0.81)
- VecFusion: Vector Font Generation with Diffusion (CVPR 2024 · similarity 0.81)
- Flash Diffusion: Accelerating Any Conditional Diffusion Model for Few Steps Image Generation (AAAI 2025 · similarity 0.81)
- StochSync: Stochastic Diffusion Synchronization for Image Generation in Arbitrary Spaces (ICLR 2025 · similarity 0.81)
- MVDream: Multi-view Diffusion for 3D Generation (ICLR 2024 · similarity 0.81)
- Spectral Progressive Diffusion for Efficient Image and Video Generation (NeurIPS 2026 · similarity 0.80)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.