See our collection for all versions of DINO.

Run DINO with Keras 3: JAX, PyTorch, or TensorFlow

zeromodels/dino-vitb16

Paper: Emerging Properties in Self-Supervised Vision Transformers (arXiv:2104.14294) · HF Papers

DINO is self-supervised: a student and teacher match across crops of the same image with no labels. The resulting features are semantic for free. These checkpoints are backbones that return tokens / feature maps.

For more details on the model, please go to the upstream model card.

Pure-Keras 3 conversion of facebook/dino-vitb16 for zeromodels. One implementation runs unmodified on TensorFlow / Torch / JAX.

This is a self-supervised backbone (DinoViTModel), not a task head.

✨ Quick start

import os
os.environ["KERAS_BACKEND"] = "torch"  # or "jax" / "tensorflow"

from zeromodels.models.dino import DinoViTModel, DinoImageProcessor

# The processor resizes + ImageNet-normalizes, so build the model with
# include_normalization=False (it would otherwise normalize a second time).
model = DinoViTModel.from_weights(
    "zeromodels/dino-vitb16", include_normalization=False
)
processor = DinoImageProcessor.from_weights("zeromodels/dino-vitb16")

pixel_values = processor("your_image.jpg")["pixel_values"]
features = model(pixel_values, training=False)
print(pixel_values.shape, features.shape)

Load any DINO variant the same way with from_weights("zeromodels/<variant>"):

Tips

• Set KERAS_BACKEND before importing Keras / zeromodels.
• The processor normalizes; pair it with include_normalization=False. To skip it, feed raw [0, 255] pixels and keep the default include_normalization=True.
• dino-resnet50 was converted from torch.hub facebookresearch/dino.
• See DINO docs and Loading Weights.
• Community / upstream weights: DinoViTModel.from_weights("hf:facebook/dino-vitb16").

Special Thanks

A huge thank you to the Facebook AI Research DINO authors for creating and releasing these models.

License: Apache 2.0.