·
9 commits
to main
since this release
What's New
Apologies for the Delay — Fine-Tuned Models Are Here!
We know it's been a while, but good things take time! We're happy to share that the fine-tuned EfficientSAM3 models are finally ready.
This release brings lightweight image encoders — EV-M, RV-M, and TV-M — trained end-to-end on 5% SA1B data with SACap labels for full Promptable Concept Segmentation (PCS) capabilities.
| Model | Vision Encoder | Text Encoder | Decoder | Other | Total Params | vs ImageSAM3 |
|---|---|---|---|---|---|---|
| EV-M | EfficientViT-B1 (22.2M) | MobileCLIP-S0 (42.5M) | 21.0M | 3.5M | 89.2M | 90% smaller |
| RV-M | RepViT-M1.1 (25.6M) | MobileCLIP-S0 (42.5M) | 21.0M | 3.5M | 92.7M | 89% smaller |
| TV-M | TinyViT-11M (28.3M) | MobileCLIP-S0 (42.5M) | 21.0M | 3.5M | 95.3M | 89% smaller |
Note: ImageSAM3 comparison: Vision (463M) + Text (354M) + Decoder (30.3M) + Other (14.2M) = 861.5M
Grab the checkpoints on HuggingFace.
We've also given the README a clean-up — easier to navigate, clearer model descriptions, and better organized for contributors.
Three-Stage Progressive Distillation
EfficientSAM3 delivers a complete distillation pipeline:
- Stage 1: Compact encoder distillation on SA-1B (image) + Recap-DataComp-1B (text)
- Stage 3: End-to-end fine-tuning on SAM3 data for full PCS quality