sam3-onnx / README.md
RamRom's picture
Document SAM3.1 Multiplex release
68ce274 verified
|
Raw
History Blame Contribute Delete
2.3 kB
---
license: other
library_name: onnxruntime
tags:
- onnx
- onnxruntime
- segment-anything
- image-segmentation
---
# SAM3 and SAM3.1 ONNX Runtime bundles
This repository publishes manifest-driven CUDA ONNX Runtime release bundles for
SAM3 base and SAM3.1. All plans require the CUDA Execution Provider and
IOBinding; there is no CPU fallback.
| Bundle | Public plan | Fixed profile | Capability |
|---|---|---|---|
| `sam3-image-pcs-ortcuda-v2` | `sam3_base_image_pcs_text_ortcuda_v1` | B1 / 1008 / L32 / Q200 / FP16 | Text-only image PCS |
| `sam3-interactive-image-pvs-ortcuda-v2` | `sam3_base_interactive_image_pvs_ortcuda_v1` | B1 / 1008 / P16 / box1 / mask288 / FP16 | Interactive image point, box, and prior-logit masks |
| `sam3-base-video-tracking-ortcuda-v2` | `sam3_base_video_tracking_ortcuda_v1` | B4 / 1008 / P16 / box1 / mask288 / M10 / ptr16 / FP16 | Per-object base-video preview, commit, and tracking |
| `sam3-multiplex-video-tracking-ortcuda-v2` | `sam3_1_multiplex_video_tracking_ortcuda_v1` | bucket16 / 1–2 buckets / at most 32 objects / 1008 / P16 / mask288 / M10 / ptr16 / FP16 | SAM3.1 native Multiplex video tracking and point-box-mask correction |
Each bundle includes its release manifest, graph/external-data files, capture
signatures, reference fixtures, provenance, and validation report. The
manifest is the machine-readable source of truth for artifact names, hashes,
profiles, Host Runtime boundaries, and public plan dispatch.
The [M2–M4 ZeroGPU Space](https://huggingface.co/spaces/RamRom/sam3-onnx-zerogpu)
demonstrates the SAM3 base plans. It does not currently run the SAM3.1
Multiplex plan.
## Scope
M2 covers text-only image PCS. M3 adds interactive image PVS. M4 adds
in-memory SAM3 base video tracking with point/box/prior-mask-logit correction
and per-object batching. M5 adds SAM3.1 native bucket-space Multiplex video
tracking for up to 32 public object IDs.
The SAM3.1 bundle excludes SAM3 base, text/geometry image PCS, streaming video,
CPU fallback, unbounded dynamic buckets, and M6 backend bundles. Consult each
bundle manifest for its exact scope and exclusions.
The original SAM models and these derived ONNX artifacts are made available
under the [SAM License](https://huggingface.co/facebook/sam3.1/blob/main/LICENSE).