Hi3D v1.0 open-sources production-ready image-to-3D model trained on 2M assets
rohanpaul_ai · x · 2026-10-28
Hi3D (formerly Sparc3D) has open-sourced Hi3D v1.0, with code and weights freely available on GitHub and Hugging Face under a commercial-friendly license, allowing anyone to self-host and build on top of it.
Key points
- Pain point: most image-to-3D models are judged by their front view — backs, bottoms and thin parts like wings, horns and handles often come out blobby or missing
- Approach: takes an image plus a coarse mesh, then refines geometry in three passes at 512³, 1024³ and 1536³ resolution, built on the Sparc3D representation with a self-pruning sparse VAE
- Scale: trained on 2 million curated 3D assets, with an effective sparse extraction-grid resolution of 1536³
- Production-tested: validated in Hi3D's live product before release, evaluated on multi-view quality, local detail and geometric usability
- Also announced: closed-source Hi3D v3.0 reaches 2048³ voxel-level geometry with up to 8K textures, at lower generation cost
The repo ships inference code only; training code and the evaluation harness are not included.
Related event: Hi3D V1.0 Open-Sourced with 2M Training Assets for High-Res 3D Generation(3 posts)→
More from Models
- Cohere reports Qwen3.8 tried to exfiltrate a diagnostic file to a Beijing container — JeremyNguyenPhD · 2026-10-28
- YacineMTB rails: even the best LLMs are 'complete dogshit' and he's tired of it — yacineMTB · 2026-10-28
- Omni-InteractScope benchmarks proactive interaction in continuous omnimodal streams — Qwen · 2026-10-28
- Mistral releases Timestral time-series foundation models, tops GIFT-Eval — mistralai · 2026-10-28
- Gemini 4 Argon Hits 1.73% WER on OCR Benchmark, Surpassing Human Experts — gabriberton · 2026-10-28
- Meta's Omnilingual MT covers 1,600 languages, 1B–8B models rival 70B LLMs — kalyan_kpl · 2026-10-28