InternViT-300M-448px
| Property | Value |
|---|---|
| Parameter Count | 304M |
| License | MIT |
| Image Size | 448x448 |
| Tensor Type | BF16 |
| Paper | View Paper |
What is InternViT-300M-448px?
InternViT-300M-448px is a compact vision foundation model developed by OpenGVLab through knowledge distillation from the larger InternViT-6B-448px-V1-5. This efficient model maintains powerful capabilities while significantly reducing the parameter count to just 304M, making it more accessible for various applications.
Implementation Details
The model operates on a dynamic input resolution of 448×448 pixels and supports processing of 1 to 12 tiles during training, expanding to 1-40 tiles during inference. It's built using transformer architecture and optimized for BF16 precision, offering an excellent balance between performance and resource utilization.
- Pretrained on multiple datasets including LAION-en, LAION-zh, COYO, and specialized OCR datasets
- Supports dynamic multi-tile processing for handling various image sizes
- Implements efficient knowledge distillation techniques
- Optimized for bfloat16 precision
Core Capabilities
- High-quality image feature extraction
- Robust OCR capabilities inherited from parent model
- Efficient processing of high-resolution images
- Flexible tile-based processing system
Frequently Asked Questions
Q: What makes this model unique?
This model uniquely combines efficiency with powerful capabilities, offering the robust features of larger models while maintaining a relatively small 304M parameter count. Its ability to handle dynamic resolutions and multiple tiles makes it versatile for various vision tasks.
Q: What are the recommended use cases?
The model excels in image feature extraction tasks, OCR applications, and scenarios requiring high-resolution image processing. It's particularly suitable for applications where computational resources are limited but high-quality vision capabilities are needed.
