AstolfoKarmix: AstolfoMix-XL + Karmix-XL

An experimental merge for ?????? 2 different base model.

Discord: "Good luck".

Specification

• Merge log / receipe: e2e script, log. 2 x 3 models + 1 tie breaker.

• Algorithm used: "Read from bottom". "Modded DELLA". Parameter in e2e script.

• (Measured in L2 norm) 87.5% NoobAI eps + 12.5% "everything else".

• Tech report.

How to use

• Karmix: "Another NoobAI / IL merge". You know the drill.

• AstolfoMix: Train LoRA / merge on top of it. Tested for realistic model already. Introduced forgotten general contents.

• NoobAI eps: Recovered some artist style and NSFW content. It replaced many "conflicted vectors".