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. Its other feature sources (SSD-1B, SAM3, Vulture) come from their upstream repositories and are not hosted here.

File	Bytes	SHA-256
alibi_dv3_vits16_best.pth	359784265	8a1ae66de6174ca3634642413bd0bd5bb712e3b04a82233473fc1e380e96c01f

The file 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.

Licence

These weights are a derivative of Meta's DINOv3 and are distributed under the DINOv3 License, not SemSketch's GPL licence.