TriHuman: Real-Time and Controllable Digital Human Synthesis at SIGGRAPH Asia 2024
chrisgrayson · x · 2026-08-09
Researchers from Max Planck Institute for Informatics and collaborating institutions introduce TriHuman, a novel human-tailored, deformable, and efficient tri-plane representation for digital humans.
- Core Breakthrough: Overcomes the slow runtime of previous NeRF-based methods that rely on Multi-Layer Perceptrons (MLPs), achieving real-time performance.
- Technical Details: Non-rigidly warps global ray samples into an undeformed tri-plane texture space to address mapping conflicts. The representation is conditioned on skeletal motion to account for dynamic appearances.
- Results: Supervised solely on multi-view imagery, it generates high-fidelity, topology-consistent clothed human geometry and photorealistic renders in real-time based on skeletal motion and camera views.
The project is open-source with available code and data, and will be presented at SIGGRAPH Asia 2024 on December 5th.
More from Multimodal
- Seedance 2.5 Test: Seamlessly Integrates 20 Makeup Products into a 30-Second Ad — Scobleizer · 2026-08-09
- ComfyUI Spectrum Node Boosts Ideogram Inference Speed by 1.39x — Every-Walrus · 2026-08-09
- Wan 2.2 vs Minimax i2v: Significant Differences in Character Animation — Crafty-Percentage-29 · 2026-08-09
- Generate 3D-Printable Models with Claude Code Skill via CadQuery Scripts — tom_doerr · 2026-08-09
- Seedance 2.5 Demo: Scarily Realistic Fake Paparazzi Footage — LudovicCreator · 2026-08-09
- MiniMax H3 Test: Characters Struggle to Sync Dance Moves with Music Beats — PhilMcGraw · 2026-08-09