PrinGS anonymizes faces and text in Gaussian Splatting scenes in under 1 second
ponguru · x · 2026-10-30
- PrinGS, an ICVGIP 2026 paper from IIIT Hyderabad students, tackles privacy leakage in 3D Gaussian Splatting, where trained splats encode faces, license plates, documents, and screens.
- The representation-level framework takes a trained splat plus one reference view marking sensitive content, localizes the regions in 3D, and anonymizes them directly — no original training images needed, completing post-training anonymization in under 1 second while preserving geometry and fidelity across all novel views.
- A privacy-by-design training variant prevents sensitive content from being embedded during optimization. Metrics: Face Similarity Measure of 9.79% at 39.47% retained Face Detection Rate; OCR Character Error Rate of 1.00 renders sensitive text unreadable.
- Paper, code, and the PrinGS-Humans dataset are released.
Related event: IIIT Hyderabad's PrinGS Paper Accepted by ICVGIP 2026(2 posts)→
More from Research
- Eilenberg's visionary 'Automata, Languages and Machines' gets a fresh recommendation — abhi9u · 2026-10-30
- Learning RL from scratch: one-page notes site de-jargonizes reinforcement learning basics — East-Muffin-6472 · 2026-10-30
- "Your LLM evaluation is lying to you": temp-0 Qwen3-235B gives 80 different answers in 1000 runs — irombie · 2026-10-30
- inclusionAI open-sources ZwZ2 vision models: +4.99 points across 25 benchmarks without zoom tools — jacek2023 · 2026-10-30
- Epoch data: leading chips get ~40% more efficient yearly, but AI power draw keeps climbing — teortaxesTex · 2026-10-30
- Soft-MTP: a one-line fix removes noisy gradients from multi-token prediction — pmttyji · 2026-10-30