Text-to-SVG: Berkeley has revealed a new approach involving an abstract pixel-based diffusion model.
April 25, 2025
In the field of text-to-image synthesis, researchers at UC Berkeley have successfully demonstrated the creation of vector graphics that are conditioned by text inputs. by
Diffusion models have shown impressive results in text-to-image synthesis, as they learn to create raster images from extensive collections of labeled photographs, capturing a multitude of diverse subjects and scenarios. However, when it comes to digital icons, illustrations, and stickers, designers generally favor vector images, such as those defined by Scalable Vector Graphics (SVGs), due to their compactness and scalability without losing quality. Alisa Davidson UC Berkeley has showcased a method to generate vector graphics that can be saved as SVGs, leveraging a text-conditioned diffusion model trained on pixel representations of images. Remarkably, this process does not rely on large datasets of labeled SVGs. Rather, the researchers at Berkeley focus on vectorizing a sample and refining it using a Score Distillation Sampling loss technique, inspired by recent advancements in text-to-3D synthesis.
Take a look at the newly created gallery of SVGs. April 25, 2025 Vector graphics remain compact while preserving their clarity at any size. Berkeley researchers have enhanced their optimization for vector graphics by refining an image-text loss that builds upon Score Distillation Sampling. The DiffVG differentiable SVG renderer, employed by VectorFusion, enables the generation of inverted visuals.
Know More
Moreover, VectorFusion introduces a multi-phase setup that improves both efficiency and the quality of outputs. This innovative approach begins with generating raster samples using a method dubbed Stable Diffusion. Subsequently, these samples are automatically traced using VectorFusion technology. However, the initial samples may often lack intricate details or be uninspiring, making it challenging to convert them into vector graphics. The method aims to enhance their vibrancy and textual coherence via Score Distillation Sampling. Know More .
With a clever limitation of SVG paths to grid-aligned squares, VectorFusion can even recreate pixel art reminiscent of classic video games.
This methodology is readily adaptable for generating text-to-sketch illustrations. The process starts by drawing 16 randomly selected strokes, which serve as the foundation for creating an abstract line drawing that accurately reflects the provided textual input. Following this, the latent Score Distillation Sampling loss is fine-tuned for optimization. Read More Check out the top 20 AI-generated music samples created with prompts from Mubert.
PlayStation Stars introduces a loyalty program featuring digital collectibles, distinguishing them from traditional NFTs.
Is the Metaverse merely an enhanced version of Second Life?
Read more
Alisa Davidson
April 25, 2025 Press Releases Addressing DeFi fragmentation: Discover how Omniston is enhancing liquidity on the TON blockchain.
Business
Vanilla has introduced super perpetual contracts with a staggering leverage of 10,000x on the BNB Chain.
Markets
