I've been obsessed with MilkDrop for a long time. Several years ago I curated through all 52,000 presets, selected the very best presets, and then organized them into categories and subcategories. So I've been curious if an LLM could be fine-tuned on this dataset and then generate new presets. When the Qwen-Coder-7b-Instuct model was released, it seemed like the open-source models have matured to the point that it's worth trying out. So I teamed up a InferenceIllusionist and started doing some fine-tuning experiments of the model. The results have been exciting to play with and so we've released the model for anyone to experiment with.

What is MilkDrop? It’s a music visualization engine which reacts in real time to the music. MilkDrop has a long 23 year history, has been ported to numerous systems, and a robust community of users that have made tons of original presets. Today, there is a new generation of artists using the NestDrop app within their live concert performances. I'm one of the developers of NestDrop.

Feel free to ask me any questions.