nebulette/c-side

C/B-SIDE

A diffusion model with BERT. It's backward compatible with the T5 tokenizer.

Spatial encoding loss was calculated as it was explained elsewhere.

Cozyberry was chosen as the only text encoder. There are no adapters.

As in the waifu diffusion, the image output alignment requires 10-100x less VRAM, due to the use of random patch cropping during training.

History

• After evaluating different text encoders, the final lightweight BERT model was born
• Later on, thousands of classes from the danbooru 2025-26 were extracted, and the model learned from both the textual and visual clues
• In this release, horizontal scenes were further reinforced exclusively for the BERT model

Source data

• synthetic booru character fashion
• horizontal scenes