transformers/docs/source/en/model_doc/dinov3.md

This model was published in HF papers on 2025-08-13 and contributed to Hugging Face Transformers on 2025-08-14.

DINOv3

DINOv3 is a family of versatile vision foundation models that outperforms the specialized state of the art across a broad range of settings, without fine-tuning. DINOv3 produces high-quality dense features that achieve outstanding performance on various vision tasks, significantly surpassing previous self- and weakly-supervised foundation models.

You can find all the original DINOv3 checkpoints under the DINOv3 collection.

Tip

Click on the DINOv3 models in the right sidebar for more examples of how to apply DINOv3 to different vision tasks.

The example below demonstrates how to obtain an image embedding with [Pipeline] or the [AutoModel] class.

from transformers import pipeline


pipe = pipeline(
    task="image-feature-extraction",
    model="facebook/dinov3-vits16-pretrain-lvd1689m",
)

pipe("https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/pipeline-cat-chonk.jpeg")

import torch

from transformers import AutoImageProcessor, AutoModel
from transformers.image_utils import load_image


url = "http://images.cocodataset.org/val2017/000000039769.jpg"
image = load_image(url)

processor = AutoImageProcessor.from_pretrained("facebook/dinov3-vits16-pretrain-lvd1689m")
model = AutoModel.from_pretrained(
    "facebook/dinov3-vits16-pretrain-lvd1689m",
    device_map="auto",
    attn_implementation="sdpa"
)

inputs = processor(images=image, return_tensors="pt").to(model.device)
with torch.inference_mode():
    outputs = model(**inputs)

pooled_output = outputs.pooler_output
print("Pooled output shape:", pooled_output.shape)

Quantization reduces the memory burden of large models by representing the weights in a lower precision. Refer to the Quantization overview for more available quantization backends.

The example below uses torchao to only quantize the weights to int4.

# pip install torchao
import torch
from torchao.quantization import Int4WeightOnlyConfig

from transformers import AutoImageProcessor, AutoModel, TorchAoConfig
from transformers.image_utils import load_image


url = "http://images.cocodataset.org/val2017/000000039769.jpg"
image = load_image(url)

processor = AutoImageProcessor.from_pretrained("facebook/dinov3-vitsplus-pretrain-lvd1689m")

quant_type = Int4WeightOnlyConfig(group_size=128)
quantization_config = TorchAoConfig(quant_type=quant_type)

model = AutoModel.from_pretrained(
    "facebook/dinov3-vit7b16-pretrain-lvd1689m",
    device_map="auto",
    quantization_config=quantization_config
)

inputs = processor(images=image, return_tensors="pt").to(model.device)
with torch.inference_mode():
    outputs = model(**inputs)

pooled_output = outputs.pooler_output
print("Pooled output shape:", pooled_output.shape)

Notes

• The example below shows how to split the output tensor into:

  • one embedding for the whole image, commonly referred to as a CLS token, useful for classification and retrieval
  • register tokens - learnable embeddings that act as dedicated “memory slots” for global information, they reduce high-norm artifacts in patch tokens, yielding cleaner attention maps and better performance on dense prediction tasks.
  • a set of local embeddings, one for each 16x16 patch of the input image, useful for dense tasks, such as semantic segmentation

  import torch
  from transformers import AutoImageProcessor, AutoModel
  from transformers.image_utils import load_image
  
  url = "http://images.cocodataset.org/val2017/000000039769.jpg"
  image = load_image(url)
  print("Image size:", image.height, image.width)  # [480, 640]
  
  processor = AutoImageProcessor.from_pretrained("facebook/dinov3-vits16-pretrain-lvd1689m")
  model = AutoModel.from_pretrained("facebook/dinov3-vits16-pretrain-lvd1689m", device_map="auto")
  patch_size = model.config.patch_size
  print("Patch size:", patch_size) # 16
  print("Num register tokens:", model.config.num_register_tokens) # 4
  
  inputs = processor(images=image, return_tensors="pt").to(model.device)
  print("Preprocessed image size:", inputs.pixel_values.shape)  # [1, 3, 224, 224]
  
  batch_size, _, img_height, img_width = inputs.pixel_values.shape
  num_patches_height, num_patches_width = img_height // patch_size, img_width // patch_size
  num_patches_flat = num_patches_height * num_patches_width
  
  with torch.inference_mode():
    outputs = model(**inputs)
  
  last_hidden_states = outputs.last_hidden_state
  print(last_hidden_states.shape)  # [1, 1 + 4 + 256, 384]
  assert last_hidden_states.shape == (batch_size, 1 + model.config.num_register_tokens + num_patches_flat, model.config.hidden_size)
  
  cls_token = last_hidden_states[:, 0, :]
  patch_features_flat = last_hidden_states[:, 1 + model.config.num_register_tokens:, :]
  patch_features = patch_features_flat.unflatten(1, (num_patches_height, num_patches_width))
  

DINOv3ViTConfig

autodoc DINOv3ViTConfig

DINOv3ConvNextConfig

autodoc DINOv3ConvNextConfig

DINOv3ViTModel

autodoc DINOv3ViTModel - forward

DINOv3ViTBackbone

autodoc DINOv3ViTBackbone

DINOv3ConvNextModel

autodoc DINOv3ConvNextModel - forward

DINOv3ViTImageProcessor

autodoc DINOv3ViTImageProcessor - preprocess

DINOv3ConvNextBackbone

autodoc DINOv3ConvNextBackbone - forward