FiCA paper: instant Gaussian Codec avatars from a single portrait image
rsasaki0109 · x · 2026-09-16
The FiCA paper introduces a feed-forward pipeline that generates lifelike, drivable Gaussian Codec avatars from a single portrait image.
Key points:
- Single-image input gives limited visual cues for inferring 3D head appearance and geometry
- The system combines human-centric vision foundation models with a diffusion model to exploit partial observations
- A diffusion model learns a generative mapping from partial observations to complete, authentic 3D mesh reconstruction
- A feed-forward mesh refinement network further improves quality
More from Multimodal
- Zombie Survival POV Video Made in 30 Seconds with Text Prompts in Neta Studio — SimplyAnnisa · 2026-09-16
- GPT-6 Astra Rebuilds a Full Blender Studio from Five Photos in 11 Minutes — CodeByPoonam · 2026-09-16
- GPT-6 Astra Generates Houdini Procedural Modeling Animation from a Single Prompt — CodeByPoonam · 2026-09-16
- LLaDA-Image: 6B fully-diffusion DiT trained on 90% image-only data, no caption bottleneck — jiqizhixin · 2026-09-16
- MiniMax H3 motion graphics demo impresses, $50K challenge with Picsart opens — egeberkina · 2026-09-16
- After Suno dropped his chords, this user built a MIDI-to-.abc converter so YuE2 preserves them — Saren-WTAKO · 2026-09-16