The bidirectional model predicts the similarity between the gelbooru images. For each image an image thumbnail and an (incomplete) text description must be provided.
The encoder blocks include the SnowLily/CrossAttention modules.
Long, comma-separated, unmasked and masked booru tags were used to train the model. The [CLS] token was the part of the loss calculation.
Source data
- Took 50k common booru tags and fetched sample images for each of them.
- Of all the tags, 7k were more frequent in the dataset than the rest.
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support