SigLino: Vision Foundation Models distilled from SigLIP2 and DINOv3

Accepted at CVPR 2026

This work stems from the CVPR 2026 AMoE paper, which designs and applies distillation into a Mixture-of-Experts (MoE) vision architecture. We have chosen the name SigLino for better clarity (SigLIP2 + DINOv3).

SigLino is a collection of vision encoders distilled from DINOv3 and SigLIP2 teachers, supporting multi-resolution image understanding.

Model Zoo

Model	Architecture	Active Params	Total Params	Config Name	Checkpoint
SigLino-MoE-0.3-0.6B	MoE (top-6/28)	0.3B	0.6B	siglino-moe-0.3B	Download
SigLino-MoE-0.15-0.6B	MoE (top-2/28)	0.15B	0.6B	siglino-moe-0.15B	Download
SigLino-0.6B	Dense	0.6B	0.6B	siglino-0.6B	Download
SigLino-70M	Dense	0.07B	0.07B	siglino-70M	Download
SigLino-30M	Dense	0.03B	0.03B	siglino-30M	Download

Installation

pip install -r requirements.txt
pip install -e .

Quick Start

SigLino-MoE-0.15B (MoE, top-2/28, 0.15B active / 0.6B total)

from siglino import load_siglino_model
from PIL import Image
import torch

# Download checkpoint
# wget https://github.com/tiiuae/amoe/releases/download/AMoE-checkpoint/amoe-0.15B.pt

model, processor = load_siglino_model(
    checkpoint_path="amoe-0.15B.pt",
    config_name="siglino-moe-0.15B",
    device="cuda",
)
model = model.to(torch.bfloat16)

image = Image.open("image.jpg")
inputs = processor(image, return_tensors="pt")

with torch.no_grad():
    pixel_values = inputs["pixel_values"].to("cuda", dtype=torch.bfloat16)
    spatial_shapes = inputs["spatial_shape"].to("cuda")
    padding_mask = inputs["padding_mask"].to("cuda")

    outputs = model(
        pixel_values=pixel_values,
        spatial_shapes=spatial_shapes,
        padding_mask=padding_mask
    )

    # DINOv3-style patch features
    patch_features = outputs["patch_features"]["dinov3"]  # (N, L, 1024)

    # SigLIP2-style pooled features
    pooled_features = outputs["summary_features"]["siglip2"]  # (N, 1152)

    # Native model features
    siglino_features = outputs["patch_features"]["siglino"]  # (N, L, 768)

SigLino-0.6B

from siglino import load_siglino_model
from PIL import Image
import torch

# Download checkpoint
# wget https://github.com/tiiuae/amoe/releases/download/AMoE-checkpoint/dense-0.6B.pt

model, processor = load_siglino_model(
    checkpoint_path="dense-0.6B.pt",
    config_name="siglino-0.6B",
    device="cuda",
)
model = model.to(torch.bfloat16)

image = Image.open("image.jpg")
inputs = processor(image, return_tensors="pt")

with torch.no_grad():
    pixel_values = inputs["pixel_values"].to("cuda", dtype=torch.bfloat16)
    spatial_shapes = inputs["spatial_shape"].to("cuda")
    padding_mask = inputs["padding_mask"].to("cuda")

    outputs = model(
        pixel_values=pixel_values,
        spatial_shapes=spatial_shapes,
        padding_mask=padding_mask
    )

    # DINOv3-style patch features
    patch_features = outputs["patch_features"]["dinov3"]  # (N, L, 1024)

    # SigLIP2-style pooled features
    pooled_features = outputs["summary_features"]["siglip2"]  # (N, 1152)

    # Native model features
    siglino_features = outputs["patch_features"]["siglino"]  # (N, L, 1280)

Results

All results use ensemble features (combined DINOv3 + SigLIP2 heads) unless noted.

kNN Classification (512x512)

Model	Params (active)	ImageNet	CUB-200	Food-101	DTD	Aircraft	Flowers-102	Avg
SigLino-MoE-0.3B	0.3B	85.9	88.9	95.5	80.6	92.3	99.8	90.5
SigLino-MoE-0.15B	0.15B	85.0	88.2	94.6	79.9	91.5	99.8	89.8
SigLino-0.6B	0.6B	86.1	89.2	95.8	81.1	92.3	99.8	90.7
SigLino-70M	0.07B	81.7	84.8	90.2	77.3	83.3	99.6	86.2
SigLino-30M	0.03B	79.0	80.9	87.5	75.7	77.3	99.2	83.3

Zero-Shot Image-Text Classification (512x512)

Model	ImageNet	CUB-200	Food-101	DTD	Aircraft	Flowers-102	Caltech-101	Avg
SigLino-MoE-0.3B	79.9	82.6	94.6	69.9	83.5	88.6	88.4	83.9
SigLino-MoE-0.15B	78.8	80.9	93.4	69.5	81.4	89.0	88.6	83.1
SigLino-0.6B	80.5	83.0	95.0	71.1	82.4	89.0	89.9	84.4
SigLino-70M	71.2	70.7	85.2	65.7	60.1	84.0	88.3	75.0
SigLino-30M	65.1	59.9	80.3	62.9	48.3	77.4	87.2	68.7

Retrieval R@1 (512x512)

Model	Flickr30K T2I	Flickr30K I2T	MSCOCO T2I	MSCOCO I2T
SigLino-MoE-0.3B	81.6	94.6	54.7	70.8
SigLino-MoE-0.15B	81.0	92.9	54.2	71.1
SigLino-0.6B	81.9	94.2	55.6	72.9
SigLino-70M	77.5	90.5	50.4	65.4
SigLino-30M	72.9	82.2	46.6	59.7

Linear Segmentation mIoU (1024x1024)

Model	Pascal VOC	Cityscapes
SigLino-MoE-0.3B	88.9	65.4
SigLino-MoE-0.15B	88.1	63.6
SigLino-0.6B	89.8	67.3
SigLino-70M	84.8	61.6
SigLino-30M	82.1	59.2

PCA Visualization

To visualize the principal components of the features:

python pca_maps.py \
    --ckpt_path path/to/checkpoint.pt \
    --input_dir path/to/images/ \
    --output_path ./output_viz/ \
    --num_samples 10

Sample output:

HF usage

import torch
from PIL import Image
from transformers import AutoModel, AutoImageProcessor

# Load model and processor
model_id = "tiiuae/siglino-moe-0.3-0.6B"
model = AutoModel.from_pretrained(model_id, trust_remote_code=True).to("cuda", dtype=torch.bfloat16)
processor = AutoImageProcessor.from_pretrained(model_id, trust_remote_code=True)

# Preprocess image
image = Image.open("image.jpg").convert("RGB")
inputs = processor(image, return_tensors="pt").to("cuda")
inputs["pixel_values"] = inputs["pixel_values"].to(torch.bfloat16)

# Inference
with torch.no_grad():
    outputs = model(**inputs)

# Access specialized features
# Options: 'siglino' (768d), 'siglip2' (1152d), 'dinov3' (1024d)
patch_features = outputs["patch_features"]["siglino"]    # (Batch, Tokens, 768)
summary_features = outputs["summary_features"]["siglip2"] # (Batch, 1152)

Citation

If you use this work in your research, please cite:

@inproceedings{chaybouti2026siglino,
  title={SigLino: Efficient Multi-Teacher Distillation for Agglomerative Vision Foundation Models},
  author={Chaybouti, Sofian and Narayan, Sanath and Dahou, Yasser and Khắc, Ph{\'u}c H L{\^e} and Singh, Ankit and Huynh, Ngoc and Para, Wamiq Reyaz and Kuehne, Hilde and Hacid, Hakim},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={10071--10081},
  year={2026}
}