FoundYou: a single 52M-param SAM 2-based model unifies personalized segmentation and retrieval
rsasaki0109 · x · 2026-11-09
An ECCV 2026 paper from Politecnico di Torino and collaborators, FoundYou unifies personalized segmentation and instance retrieval in one compact model: given a reference image, it segments the exact physical instance in a new image or ranks gallery images by instance-level similarity.
- Unified tasks: segmentation and retrieval are framed as the same instance-matching problem, sharing a SAM 2 representation that yields either a dense mask or an image-level retrieval score.
- Promptable: frozen SAM 2 decoder preserves mask/box/point prompting, cutting annotation cost.
- Few-shot: multiple references stored in a memory bank jointly improve matching without retraining.
- Efficient: built on SAM 2-small with only 5.9M trained parameters (52M total), running at 90.2 images/s on an RTX 4090.
Paper, code, Gradio demo, and Colab are available.
More from Multimodal
- Solo creator releases AVIDUS ODYSSEY, a 96-minute AI-assisted space opera animated film — Gunkavoider · 2026-11-09
- Trama: a minimalist open-source orchestrator for multi-AI automations on ComfyUI — Cold-Voice2737 · 2026-11-09
- GLITCH Studio seeks beta testers for all-in-one local AI filmmaking app built on ComfyUI — Upstairs-Giraffe-189 · 2026-11-09
- Local image app ZPix adds MOSS-Image SFT, SDXL and FLUX support — SamuelTallet · 2026-11-09
- Running MiniMax H3 locally: 0.2MP preview looks good, 1MP render takes 8 hours and drifts — ok_n_na · 2026-11-09
- On-device CPU-friendly English TTS model Underdog-Woof trends on Hugging Face — ConwayResearch · 2026-11-09