"NeuralSVG: An Implicit Representation for Text-to-Vector Generation"

NeuralSVG turns text into clean vector graphics by teaching neural networks to draw like designers.

NeuralSVG uses neural networks to generate editable vector graphics from text descriptions, making the shape generation process more controllable and structured through an implicit representation approach.

https://arxiv.org/abs/2501.03992

Original Problem 🤔:

→ Current text-to-vector methods produce over-parameterized outputs with pixel-like shapes, losing the core advantages of vector graphics' editability

→ Existing approaches treat layered structure as secondary, requiring complex post-processing to create meaningful layers

-----

Solution in this Paper 💡:

→ Introduces NeuralSVG, an implicit neural representation encoding entire scenes into MLP network weights.

→ Uses Score Distillation Sampling to optimize the network based on text prompts.

→ Employs dropout-based regularization to encourage meaningful standalone shapes.

→ Enables dynamic control over color palettes and aspect ratios at inference time.

-----

Key Insights 💡:

→ Implicit neural representation allows encoding complete SVG scenes compactly

→ Dropout regularization naturally creates ordered, meaningful shape hierarchies

→ Single network can generate multiple variations through inference-time controls

-----

Results 📊:

→ Outperforms VectorFusion and SVGDreamer with just 16 shapes vs their 64-256 shapes

→ Achieves 26.94 CLIP similarity score vs 26.58 for SVGDreamer

→ Maintains high quality with 75% fewer parameters than competing methods