Tencent Hunyuan's Tex-Zero Trains 3D Texture Generation Without Any 3D Assets
Tencent-Hunyuan · hf · 2026-10-02
- Key insight: Native 3D texture generation is widely believed to need large-scale real 3D assets, but Tencent Hunyuan's Tex-Zero shows only fine-grained color information is essential — geometry can be manually constructed.
- Data recipe: High-quality 2D images are placed as planes in 3D space, with patch-wise random rotations and aggregation building complex geometry, turning abundant 2D images into effective 3D training samples.
- Models: A Tex-Zero VAE reconstructs real 3D assets despite never seeing them in training; the Tex-Zero DiT encodes conditioning multi-view images through the same VAE to shrink the representation gap.
- Result: High-fidelity 3D textures with fine details trained purely on images, suggesting a new scaling paradigm for 3D texture generation data.
More from Multimodal
- Redditor shares fully AI-generated music video for NCT 127 remix — leesysysysy · 2026-10-02
- Two white ducks having a 'very serious little conversation' in AI video — misovalko · 2026-10-02
- HuST Lab's Multimodal Flow: Fully Continuous Unified Language-Vision Generation — hustvl · 2026-10-02
- One prompt gets Claude Fable 5.5 to render a AAA-grade 3D Rube Goldberg machine — imjustnewatai · 2026-10-02
- Seedance 2.5 Video Goes Viral for Ultra-Realistic Early-2000s DV Camcorder Look — SimplyAnnisa · 2026-10-02
- ComfyUI Clips + sync.so Lip-Sync: Edit Before or After Processing? — Positive_Society1876 · 2026-10-02