Post
EVA-CLIP 🦖 is the CLIP scaled to the moon! 🔥
The new SotA CLIP-like model 🏆
Highlights ✨
- Performs better in linear probing
- Outperforms in Zero-Shot Image-Text Retrieval
- Higher zero-shot accuracy in IN-1K
As usual, try it with the notebook I built for you https://colab.research.google.com/drive/1K7DdCORC3x4qyhwhuB4fT4wcfJ_BQLKw?usp=sharing#scrollTo=0ZS_lJ7SK6Ys
I also built a Space for you to compare the output probabilities to CLIP, seems that EVACLIP is more "sure" of it's results 😊 merve/EVACLIP
The authors have shared 8B checkpoints open with Apache 2.0 license 💜 and it's built on top of transformers, super easy to use! BAAI/EVA-CLIP-8B
Read the paper EVA-CLIP-18B: Scaling CLIP to 18 Billion Parameters (2402.04252) 📄
The new SotA CLIP-like model 🏆
Highlights ✨
- Performs better in linear probing
- Outperforms in Zero-Shot Image-Text Retrieval
- Higher zero-shot accuracy in IN-1K
As usual, try it with the notebook I built for you https://colab.research.google.com/drive/1K7DdCORC3x4qyhwhuB4fT4wcfJ_BQLKw?usp=sharing#scrollTo=0ZS_lJ7SK6Ys
I also built a Space for you to compare the output probabilities to CLIP, seems that EVACLIP is more "sure" of it's results 😊 merve/EVACLIP
The authors have shared 8B checkpoints open with Apache 2.0 license 💜 and it's built on top of transformers, super easy to use! BAAI/EVA-CLIP-8B
Read the paper EVA-CLIP-18B: Scaling CLIP to 18 Billion Parameters (2402.04252) 📄