This is a LoRA that was initialized by extracting the difference of the Ideogram 4 conditional and unconditional model weights. It was further tuned using student teacher training on real data and a loss was performed on a per layer basis to more closely match the unconditional model. This can be used on the conditional Ideogram 4 model during the unconditional pass as a replacement to the full 9B paramiter unconditional model.

Using the full unconditional model will likely yield better results, but this will work as a light weight alternative. It was originally trained to be used in Ostris AI Toolkit so samples would be more in line with what the full pipeline would produce without needing to load the entire unconditional model.