SemSketch: ALiBi DINOv3-S/16

The image embedder trained for SemSketch, an interactive segmentation tool for microscopy images. It is DINOv3 ViT-S/16 (timm/vit_small_patch16_dinov3.lvd1689m) with rotary position embeddings removed and replaced by a shared ALiBi distance bias, then distilled on OpenImages at 512 px for 20 epochs. Without absolute positions, its dense features stay consistent across arbitrary image sizes and crops.

SemSketch's default dino+ssd1b+sam3+vulture embedding uses this checkpoint with bilinear upsampling. SSD-1B and Vulture come from their upstream repositories. A copy of Meta's SAM 3 checkpoint is also hosted here so SemSketch users can install it without a gated download; see SAM 3 (by Meta) below.

File Bytes SHA-256
alibi_dv3_vits16_best.pth 359784265 8a1ae66de6174ca3634642413bd0bd5bb712e3b04a82233473fc1e380e96c01f
sam3.pt 3450062241 9999e2341ceef5e136daa386eecb55cb414446a00ac2b55eb2dfd2f7c3cf8c9e

alibi_dv3_vits16_best.pth is the unmodified training checkpoint: the weights under model, plus training metadata and optimizer state.

Usage

just setup (or just download-models) in SemSketch downloads it to outputs/features/openimages_full_20260914/training/alibi-high/alibi_dv3_vits16_best.pth. To load it directly:

import timm, torch
from huggingface_hub import hf_hub_download

path = hf_hub_download("radical-ai/semsketch", "alibi_dv3_vits16_best.pth")
model = timm.create_model("vit_small_patch16_dinov3.lvd1689m", pretrained=False, num_classes=0, dynamic_img_size=True)
state = torch.load(path, map_location="cpu", weights_only=True)["model"]
model.load_state_dict({key.removeprefix("model."): value for key, value in state.items()})
model.rope = None  # position comes only from the ALiBi bias that SemSketch adds to attention

The ALiBi attention bias lives in SemSketch (semsketch.model.feats.upsampler.SharedAlibiBias); use SemSketch's loader to get features that match training.

SAM 3 (by Meta)

sam3.pt is SAM 3, created and released by Meta (Meta AI / FAIR). All credit for the model, its training and its weights belongs to Meta. Radical AI did not create, train or modify it: the file is a byte-for-byte copy of sam3.pt from Meta's facebook/sam3 repository at revision 3c879f3, and its SHA-256 above matches Meta's original. The code is at facebookresearch/sam3 and the paper is SAM 3: Segment Anything with Concepts. Please cite Meta's work, not this repository, if you use it.

We redistribute it under the SAM License, which permits redistribution: section 1.a grants the right to "use, reproduce, distribute, copy, create derivative works of, and make modifications to the SAM Materials", and section 1.b.i allows making them available to third parties "under the terms of this Agreement" provided "a copy of this Agreement" is included. LICENSE-SAM3.md is that copy, taken verbatim from Meta's repository. By downloading or using sam3.pt you agree to the SAM License, including its acceptable-use and trade-control terms. This copy is not endorsed by Meta; Meta's own repository remains the authoritative source.

Licence

The ALiBi DINOv3 weights are a derivative of Meta's DINOv3 and are distributed under the DINOv3 License. sam3.pt is Meta's SAM 3, distributed under the SAM License. Neither is covered by SemSketch's GPL licence.

Downloads last month
-
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for radical-ai/semsketch

Finetuned
(2)
this model