nvidia/vilaPREVIEW

Multi-modal vision-language model that understands text/img/video and creates informative responses

Vision language model

image caption

image to text

API Reference

Commercial VILA Model Card

Description

Vision-language models (VILA) provides multi-image reasoning, in-context learning, visual chain-of-thought, and better world knowledge. VILA is deployable on the edge, including Jetson Orin and laptop by AWQ 4bit quantization through TinyChat framework. We find: (1) image-text pairs are not enough, interleaved image-text is essential; (2) unfreezing LLM during interleaved image-text pre-training enables in-context learning; (3)re-blending text-only instruction data is crucial to boost both VLM and text-only performance.

This model is ready for commercial use. It was trained on commercial images and videos for all three stages of training and supports single image and video inference. This version does not support interleaved and in-context learning capabilities.

References

License

The license to use this model is covered by the Model EULA. By downloading the unpruned or pruned version of the model, you accept the terms and conditions of these licenses

Model Architecture

Architecture Type: Transformer-based Network Architecture

Network Architecture

Vision Encoder: SigLIP-400M
Language Encoder: Yi-34B

Input

Input Type(s): Image, Video, Text
Input Format(s): Image (Red, Green, Blue (RGB)), Video (.mp4), and Text (String)
Input Parameters: Image (2D), Video (3D), Text (1D)

Output

Output Type(s): Text
Output Formats: String
Output Parameters: 1D
Other Properties Related to Output: N/A

Software Integration

Runtime Engine(s): TensorRT-LLM
Supported Hardware Architecture(s): NVIDIA Hopper
Supported Operating System(s): Linux

Model Versions

VILA-SigLIP-Yi-34B

Training Dataset

NV-Pretraining and NV-VILA-SFT data were used.

Additionally, the commercial subset of following datasets were used:

Data Collection Method by dataset:

Hybrid: Human, Automatic/Sensors

Labeling Method by dataset:

Hybrid: Human, Automatic/Sensors

Properties:

NV-Pretraining data was collected from 5M subsampled NV-CLIP dataset. Stage 3 NV-SFT data has 2.8M images and 3.58M annotations on images that only have commercial license. Additionally, 355K videos with commercial license and 400K annotations on videos were used.

Evaluation Data

Data Collection Method by dataset:

Hybrid: Human, Automatic/Sensors

Labeling Method by dataset:

Hybrid: Human, Automatic/Sensors

Properties:

A collection of different benchmarks, including academic VQA benchmarks and recent benchmarks specifically proposed for instruction-following LMMs.

Methodology and KPI

Benchmark	VQAv2	GQA	SQA Image	Text VQA	POPE (Popular)	MME	SEED	SEED Image	MMMU val (beam 5)	SEED Video	VideoMME w/o Sub @32f	VideoMME w/ Sub @32f	Egoschema (val)	Perception Test
Accuracy	81.70	62.13	79.62	71.14	85.61	1649.62	70.36	74.12	47.33	58.21	57.85	60.67	63.8	61.76

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 supporting 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++ Promise and the Explainability, Bias, Safety & Security, and Privacy Subcards.

Please report security vulnerabilities or NVIDIA AI Concerns here.