IRIS: Latent Neural Field Enables Pose-Free Novel View Synthesis, Accepted at ACM MM 2026
zhenjun_zhao · x · 2026-09-17
Researchers propose IRIS, a fully self-supervised framework for pose-free novel view synthesis, accepted at ACM Multimedia 2026.
- Problem: Without pose supervision, models must jointly learn scene representation and camera parameters. Existing approaches split into two extremes: implicit latent-space rendering is flexible but weakly grounded, while explicit 3D representations offer geometric grounding but are heavy and fragile to optimize.
- Method: IRIS takes a middle ground by representing the scene as a latent neural field. Projected features from reference views are aggregated at sampled 3D points into point-wise latent features, then composed along target rays for rendering under self-predicted cameras.
- Results: Strong rendering quality with competitive pose accuracy under fully self-supervised learning.
Pipeline: reference views → projected features → sampled 3D points → point-wise latent features → rendering.
More from Research
- Sakana AI's Smart Cellular Bricks featured in Scientific American: 197 bricks self-identify and self-repair — SakanaAILabs · 2026-09-17
- "AI found new elliptic curves in days" claim disputed: it was brute-force compute on known methods — tak3sh8 · 2026-09-17
- Fixing one simple grader doubled the score—and the rest were errors in the tasks — xeophon · 2026-09-17
- A proposal for the future of scientific communication in the age of agents — tensorqt · 2026-09-17
- GPT-Policy: In-Context Robot Learning with VLM Agents, No Gradient Updates — Dongzhou Cheng · 2026-09-17
- Fathom Speeds Million-Token KV Scans 1.67x with Per-Query Read Depth — Vivek Kalyanarangan · 2026-09-17