TIPS styles
Take the top styles from danbooru 2026 and the style dataset, make them compatible with the Cosmos model.
These embeddings are compatible with both the T5 and the BERT text encoders.
This is because the embeddings are injected directly into the diffusion model's blocks.
Before:
@tag, ...
After:
...
The text encoder no longer requires the style tag within the text input.
The target image size was set to between 512 and 1024px; the training was done by random patch cropping.
Two things have been uploaded:
- The style embeddings, these are the new inputs of the diffusion model. Thanks to them, TIPS can be discarded at inference time.
- The patch file for the diffusion model, this is why the model understands the new format.
Coincidentally, TIPSv2 was released around the same time as Cozyberry. TIPS has a wider spatial understanding, BERT has a larger Japanese vocabulary.
Although the bidirectional BERT model was confirmed to be capable of mixing style tags, the base model of the TIPS style saw thousands of times more images than the Cozyberry model.
References
- 2604.12012
- Cozyberry
- ruwww's waifu_diffusion_2026
- What is Yann LeCun cooking