nvidia/C-RADIOv3-L

Model Overview

[**Github**] [**CVPR 2025**] [**CVPR 2024**]

Description

This model performs visual feature extraction. For instance, RADIO generates image embeddings that can be used by a downstream model to classify images.

C-RADIOv3 models are available in multiple sizes:

• —Base (90M parameters).
• —Large (320M parameters).
• —Huge (653M parameters). (In training)
• —Gigantic (1.1B parameters).

C-RADIOv3 was trained for 1M steps (400k more steps than v1), using inverse frequency sampling for data balancing, and PHI Standardization for teacher distribution balancing. As well as new techniques for summary distribution matching, and domain generalization.

This model is ready for commercial/non-commercial use.

License/Terms of Use

GOVERNING TERMS: Use of this model is governed by the NVIDIA Open Model License Agreement.

Deployment Geography

Global.

Use Case

The embeddings generated by this model are expected to be used by a downstream application. For example:

• —Image-level understanding (image classification, curation, etc.).
• —Dense processing (semantic segmentation, depth estimation, etc.).
• —Integration into a Vision-Language Model.

Release Date

Huggingface: 03/26/2025 via RADIO Collection of Models.

References

• —\[CVPR 2025\] **RADIOv2.5: Improved Baselines for Agglomerative Vision Foundation Models**
• —\[CVPR 2024\] **AM-RADIO: Agglomerative Vision Foundation Model - Reduce All Domains Into One**

Model Architecture

Architecture Type: Neural Network <br> Network Architecture: Vision Transformer <br>

Input

Input Type(s): Image <br> Input Format(s): Red, Green, Blue (RGB) <br> Input Parameters: Two Dimensional (2D) <br> Other Properties Related to Input: Image resolutions up to 2048x2028 in increments of 16 pixels <br>

Output

Output Type(s): Embeddings <br> Output Format: Tensor <br> Output Parameters: 2D <br> Other Properties Related to Output: Downstream model required to leverage image features <br>

Usage:

RADIO will return a tuple with two tensors. The summary is similar to the cls_token in ViT and is meant to represent the general concept of the entire image. It has shape (B,C) with B being the batch dimension, and C being some number of channels. The spatial_features represent more localized content which should be suitable for dense tasks such as semantic segmentation, or for integration into an LLM.

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

hf_repo = "nvidia/C-RADIOv3-L"

image_processor = CLIPImageProcessor.from_pretrained(hf_repo)
model = AutoModel.from_pretrained(hf_repo, trust_remote_code=True)
model.eval().cuda()

image = Image.open('./assets/radio.png').convert('RGB')
pixel_values = image_processor(images=image, return_tensors='pt', do_resize=True).pixel_values
pixel_values = pixel_values.cuda()

summary, features = model(pixel_values)

Spatial features have shape (B,T,D) with T being the flattened spatial tokens, and D being the channels for spatial features. Note that C!=D in general. Converting to a spatial tensor format can be done using the downsampling size of the model, combined with the input tensor shape. For RADIO, the patch size is 16.

from einops import rearrange
spatial_features = rearrange(spatial_features, 'b (h w) d -> b d h w', h=x.shape[-2] // patch_size, w=x.shape[-1] // patch_size)

The resulting tensor will have shape (B,D,H,W), as is typically seen with computer vision models.

Software Integration

Runtime Engine(s):

• —TAO- 24.10 <br>

Supported Hardware Microarchitecture Compatibility: <br>

• —NVIDIA Ampere <br>
• —NVIDIA Blackwell <br>
• —NVIDIA Jetson <br>
• —NVIDIA Hopper <br>
• —NVIDIA Lovelace <br>
• —NVIDIA Pascal <br>
• —NVIDIA Turing <br>
• —NVIDIA Volta <br>

[Preferred/Supported] Operating System(s): <br>

• —Linux
• —Linux 4 Tegra
• —QNX
• —Windows

Model Version(s)

• —C-RADIOv3-B (90M parameters).
• —C-RADIOv3-L (320M parameters).
• —C-RADIOv3-H (653M parameters).
• —C-RADIOv3-g (1.2B parameters).

Links:

• —https://huggingface.co/nvidia/C-RADIOv3-B
• —https://huggingface.co/nvidia/C-RADIOv3-L
• —https://huggingface.co/nvidia/C-RADIOv3-H
• —https://huggingface.co/nvidia/C-RADIOv3-g

Training and Evaluation Datasets

Training Dataset

NV-CC-Img-Text-Dataset <br>

Data Collection Method by dataset

• —Automated <br>

Labeling Method by dataset

• —Not Applicable (no labels are needed)

Properties

• —700 Million Images <br>

Evaluation Dataset

Link: ImageNet <br>

Data Collection Method by dataset

• —Automated <br>

Labeling Method by dataset

• —Human <br>

Properties: This dataset spans 1000 object classes and contains 1,281,167 training images, 50,000 validation images and 100,000 test images.<br>

Inference

Engine: PyTorch <br> Test Hardware: A100 <br>

Ethical Considerations

NVIDIA believes Trustworthy AI is a shared responsibility and we have established policies and practices to enable development for a wide array of AI applications. When downloaded or used in accordance with our terms of service, developers should work with their internal model team to ensure this model meets requirements for the relevant industry and use case and addresses unforeseen product misuse.

For more detailed information on ethical considerations for this model, please see the Model Card++ Explainability, Bias, Safety & Security, and Privacy Subcards below.

Please report security vulnerabilities or NVIDIA AI Concerns here.

Bias

Explainability

Privacy

Safety