Research map / Image generation / Autoregressive and latent generation
NIVeL: Neural Implicit Vector Layers for Text-to-Vector Generation
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 NIVeL, a layered neural implicit field MLP optimized with SDS through a frozen DeepFloyd diffusion model to generate editable, resolution-independent vector graphics from text.
Problem
Text-to-vector graphics generation is hard because vector representations have variable structure, scarce training data, and SDS optimization over Bezier curves yields redundant or self-intersecting geometry.
Approach
Represent vector graphics as a small set of decomposable layered neural implicit fields (MLP over a frequency-encoded 2D unit square with per-layer colors and alpha compositing), optimize them with SDS plus an entropy regularizer, and initialize by fitting a low-frequency implicit RGB generator; the implicit layers are then converted to cubic Bezier curves via marching squares and Inkscape tracing
Key results
- At 12K parameters, NIVeL reaches CLIP Sim 32.0 and R-Prec 78.5 versus VectorFusion's 26.7 and 71.3.
- At 1K parameters, NIVeL reaches CLIP Sim 25.1 and R-Prec 68.2 versus VectorFusion's 21.4 and 59.5.
- In MTurk studies with 100 participants (1000 comparisons each), NIVeL outputs were preferred by significantly larger proportions for both plausibility and text matching.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use NIVeL instead of VectorFusion at 1K parameters | Text-to-vector generation with DeepFloyd · VectorFusion prompt dataset | R-Prec: 59.5 -> 68.2 | Helped |
| Use NIVeL instead of VectorFusion at 1K parameters | Text-to-vector generation with DeepFloyd · VectorFusion prompt dataset | Sim: 21.4 -> 25.1 | Helped |
| Use NIVeL instead of VectorFusion at 12K parameters | Text-to-vector generation with DeepFloyd · VectorFusion prompt dataset | R-Prec: 71.3 -> 78.5 | Helped |
| Use NIVeL instead of VectorFusion at 12K parameters | Text-to-vector generation with DeepFloyd · VectorFusion prompt dataset | Sim: 26.7 -> 32.0 | Helped |
| Initialize NIVeL from random parameters rather than the RGB generator output (ablation) | NIVeL · Text-to-vector generation | visual quality | Hurt |
| Initialize NIVeL by reconstructing a sampled raster image rather than the RGB generator output (ablation) | NIVeL · Text-to-vector generation | visual quality | Hurt |
| Initialize NIVeL by reconstructing the implicit RGB generator output (ablation) | NIVeL · Text-to-vector generation | visual quality | Helped |
| Fine-tune with SDS after initialization from the RGB generator (ablation) | NIVeL · Text-to-vector generation | visual quality | Helped |
| Use Dreamtime timestep scheduling instead of the vanilla DreamFusion scheduler (ablation) | NIVeL SDS optimization · Text-to-vector generation | visual quality | No clear effect |
Surprising findings
- At 1K parameters, NIVeL still outperforms VectorFusion at 12K parameters on the CLIP-based Sim metric's counterpart is not claimed; however, VectorFusion struggles to capture object geometry at low parameter counts.
Limitations
- Number of layers is bounded by a fixed upper limit L=5.
- Conversion to Bezier curves relies on an external open-source vectorizer (Inkscape) rather than a differentiable module.
- SDS optimization remains sensitive to initialization and timestep scheduling.
Experimental setup
Datasets: VectorFusion prompt dataset
Benchmarks: VectorFusion prompt dataset
Metrics: CLIP similarity (Sim) R-Precision user preference
Baselines: VectorFusion
Builds on: DreamFusion VectorFusion NeRF DeepFloyd
Closest papers by meaning
- You Only Sample Once: Taming One-Step Text-to-Image Synthesis by Self-Cooperative Diffusion GANs (ICLR 2025 · similarity 0.86)
- SwiftBrush: One-Step Text-to-Image Diffusion Model with Variational Score Distillation (CVPR 2024 · similarity 0.86)
- Not All Parameters Matter: Masking Diffusion Models for Enhancing Generation Ability (CVPR 2025 · similarity 0.85)
- Neural Diffusion Models (ICML 2024 · similarity 0.85)
- Denoising Task Difficulty-based Curriculum for Training Diffusion Models (ICLR 2025 · similarity 0.85)
- Score Distillation via Reparametrized DDIM (NeurIPS 2024 · similarity 0.85)
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.