CLIP-convnext_xxlarge-laion2B-s34B-b82K-augreg-soup
| Property | Value |
|---|---|
| Model Size | 1.2B parameters |
| License | MIT |
| Training Data | LAION-2B |
| Zero-shot ImageNet Accuracy | 79.4% |
What is CLIP-convnext_xxlarge-laion2B-s34B-b82K-augreg-soup?
This is a groundbreaking CLIP model that utilizes the ConvNeXt-XXLarge architecture as its image tower, representing the largest released ConvNeXt model ever pretrained. It achieves remarkable zero-shot classification performance without requiring previous image tower pretraining, surpassing 79% accuracy on ImageNet.
Implementation Details
The model combines a ConvNeXt-XXLarge image tower (847M parameters) with a text tower equivalent in size to ViT-H-14 models. Training was conducted on the LAION-2B dataset at 256x256 resolution, utilizing a sophisticated training procedure with varying batch sizes and precision formats.
- Global batch size: 81920-95744
- Total parameters: 1.2B
- Training samples: ~34B
- Resolution: 256x256
Core Capabilities
- Zero-shot image classification
- Image and text retrieval
- Downstream task fine-tuning
- Image generation guidance
Frequently Asked Questions
Q: What makes this model unique?
It's the first non-ViT image tower CLIP model to exceed 79% ImageNet zero-shot accuracy and represents the largest ConvNeXt model ever released for pretraining tasks.
Q: What are the recommended use cases?
The model is primarily intended for research purposes, including zero-shot classification, image-text retrieval, and fine-tuning for downstream tasks. However, it's not recommended for deployed commercial applications without thorough testing.
