DistillPath-KS16-Virchow2

A 22M ViT-S/16 pathology tile encoder distilled from Virchow2 (632M ViT-H/14) into the kaiko ViT-S/16 student using backbone-token distillation on 6,000 public TCGA slides.

For the ImageNet-initialized variant, see DistillPath-IS16-Virchow2.

Paper: DistillPath: An Efficient 22M Distilled Pathology Encoder Approaching Large Foundation Model Performance Ramon Kaspar, Andrey Ignatov, Valentina Boeva. ETH Zurich. Published at the ECCV 2026 Workshop on Medical Foundation Models and Benchmarks (MedFM-Bench).

Model details

Property	Value
Architecture	ViT-S/16 (vit_small_patch16_224 in timm)
Parameters	21.7M
Feature dimension	384
Input size	224 x 224
Normalization	mean=(0.5, 0.5, 0.5), std=(0.5, 0.5, 0.5)
Student initialization	kaiko ViT-S/16 (pathology-pretrained)
Teacher	Virchow2 (632M, ViT-H/14, CC BY-NC-ND 4.0)
Training data	6,000 TCGA H&E whole-slide images, 32 cohorts
Training steps	50,000 (batch size 256, approx. 24-29 GPU-hours on 1x RTX 4090)

Benchmark results

Benchmark	DistillPath-KS16-Virchow2	kaiko baseline	Virchow2 teacher
EVA mean (7 tasks)	0.795	0.764	0.810
HEST mean (9 tasks)	0.371	0.349	0.398
PLISM score	0.447	0.307	0.447

See the paper for per-task results and comparisons with all four DistillPath variants.

Usage

Load directly from the Hub with timm:

import timm

model = timm.create_model(
    "hf_hub:RamonK/DistillPath-KS16-Virchow2",
    pretrained=True,
    num_classes=0,
)
model.eval()

Or load manually:

import timm
from huggingface_hub import hf_hub_download
from safetensors.torch import load_file

model = timm.create_model("vit_small_patch16_224", pretrained=False, num_classes=0)
path = hf_hub_download("RamonK/DistillPath-KS16-Virchow2", "model.safetensors")
state_dict = load_file(path)
model.load_state_dict(state_dict, strict=True)
model.eval()

This model uses normalization mean=(0.5, 0.5, 0.5) and std=(0.5, 0.5, 0.5), inherited from the kaiko student. Apply this normalization to input tiles before inference:

from torchvision import transforms

transform = transforms.Compose([
    transforms.Resize(224),
    transforms.CenterCrop(224),
    transforms.ToTensor(),
    transforms.Normalize(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5]),
])

Distillation recipe

The recipe reads only the teacher's final class and patch tokens (no teacher pretraining heads required):

• Class-token loss: cosine distance + RKD (relational knowledge distillation)
• Patch-token loss: cosine distance after bicubic grid resizing (teacher 16x16 to student 14x14)
• Optimizer: AdamW, lr=1e-4, weight decay 0.05, cosine decay, 500 warmup steps
• Projector: DINO-style MLP (384 to 2048 to 2048 to 256 to d_t), discarded after training

Full details in the paper and the DistillPath repository.

License

This model is released under the Kaiko Non-Commercial Public License, inherited from the kaiko ViT-S/16 student weights. The Virchow2 teacher is released under CC BY-NC-ND 4.0. The distillation process used Virchow2 only to generate supervisory outputs; the released student contains no Virchow2 weights. This model is intended solely for non-commercial academic research.

Citation

@inproceedings{kaspar2026distillpath,
  title     = {DistillPath: An Efficient 22M Distilled Pathology Encoder Approaching Large Foundation Model Performance},
  author    = {Kaspar, Ramon and Ignatov, Andrey and Boeva, Valentina},
  booktitle = {Medical Foundation Models and Benchmarks (MedFM-Bench), ECCV 2026},
  year      = {2026}
}