The system will allow users to edit the SVG UI elements by describing their desired changes in a text format, which will then be interpreted by a chatGPT-like system and applied to the SVG element.
| Technology readiness | Risks | Complexity |
|---|---|---|
| 🟢 Ready for implementation | 🟡 Moderate risk |
🟡 Moderately complex |
Following LLM for SVG editing [Paper], the input can be an optimized SVG with text instructions to edit it. The language Model can be tuned on the created target dataset using raster-to-vector tools.
| Demo #1 | Demo #2 | Demo #3 |
|---|---|---|
 |
 |
 |
- ML model: LLM
- SVG optimizer: SVGO [Github]
- Raster-to-vector tool:
- Vectorisation [Tutorial][Sample code]
- Vtracer [Github]
- Input: SVG component + text prompt
- Output: edited SVG component
Dataset: Open-source SVG components with text description for finetuning LLM
-
Iconify: >150,000 open source SVG icons [Website] [Description] [Figma Plug-in] [Figma Plug-in Github]
-
FIGR-8: containing 17,375 classes of 1,548,256 images representing pictograms, ideograms, icons, emoticons or object or conception depictions (with both png and svg format) [Github]
-
SVG Repo: with 500,000+ open-licensed SVG vector and icons [Website]
[Research]
-
LLM for SVG editing [Paper]
-
LLM for image editing [Github]: GroundingDINO [Github] + GLIGEN [Github]
-
Text prompt for image editing: InstructPix2Pix [Github]；Prompt-to-prompt [Github]
🟢 Pros
- It could leverage foundation models to understand image content in SVG format
- Provides obvious AI performance on a cross-domain task
- Has sufficient dataset
🔴 Cons
- Generation quality has a risk of not meeting the designer's requirements since there is limited research on SVG generation
- It might be limited on generated detailed SVG components