GeoFace uses geometry-constrained diffusion to keep multi-view face generation consistent
rsasaki0109 · x · 2026-07-29
KAIST researchers introduce GeoFace, a geometry-constrained multi-view diffusion framework for generating consistent face images from a single input.
- GeoFace jointly generates multi-view RGB images and a shared canonical UV position map.
- The architecture uses a dual-stream design with geometry-guided attention alignment so that appearance and 3D structure stay coherent across viewpoints.
- The paper says this improves consistency, especially under large pose changes.
- The authors position it for 3D reconstruction, digital avatars, and immersive content creation.
More from Multimodal
- ComfyUI demo builds an anime 2D motion sequence in one hour, end to end — petewoodbridge · 2026-07-29
- Kimi K3 via MCP models a 190-room hotel in 9 days, cutting costs to $10.5K — anselm · 2026-07-29
- Dreamina teases Seedance 2.5 with 50 multimodal refs and 30-second video output — eyishazyer · 2026-07-29
- Netflix Under Fire for Using Deepfake Instead of Blurring Faces in Documentary — burkov · 2026-07-29
- KRAFTON AI and SK Telecom release a 21B SpeechLM with tool use and emotion recognition — Kangwook_Lee · 2026-07-29
- Google releases GNM, a parametric 3D head model with identity and expression control — rsasaki0109 · 2026-07-29