HiVG: 3B Parameter Open Model Beats GPT-5 & 6 Proprietary Rivals in Image-to-SVG
tom_doerr · x · 2026-08-04
The open-source GitHub project HiVG (Hierarchical SVG Tokenization) converts images into clean, editable SVGs while faithfully preserving structure, layout, and details.
Key Highlights:
- Small Model, Frontier Results: With only 3B parameters, it outperforms 7 top proprietary models, including GPT-5 and Gemini 2.5, on image-to-SVG tasks.
- Efficient Compression: Introduces hierarchical tokenization (Raw SVG → Atomic tokens → Segment tokens), achieving 2.76x sequence compression.
More from Multimodal
- Runway Teases Seedance 2.5, Offers 7 Days Unlimited Access to Max Subscribers — runwayml · 2026-08-04
- No Photo Studio Needed: Generating High-Quality LoRA Datasets Iteratively from a Single Photo — Sudden-Complaint7037 · 2026-08-04
- MiniMax H3 Video Generation Tested: Shows Excellent Handling of IP Content — animovirtus · 2026-08-04
- ComfyUI Troubleshooting: How to Fix Frozen Initial Frames in MiniMax I2V Workflow — giuzootto · 2026-08-04
- Testing Seedance 2.5: A Seamless 30-Second One-Take Musical Instrument Transformation — LinusEkenstam · 2026-08-04
- Idea Exploration: Can PiD Upscaler Be Used as a Refiner Stage for Video Generation — bonesoftheancients · 2026-08-04