The LoRAs were extracted from fine-tuned checkpoints (Schnell), with the official Flux.1-Dev also used as a basis for comparison. As a result, these differential LoRAs retain the four-step Schnell optimizations and can therefore be used with Dev and Krea checkpoints. I did this primarily for myself: LoRAs are simply easier to work with — you can mix them in real time and set different weights and strengths. They also save storage space on your local machine. Choose a rank to suit your preferences and hardware capabilities (the lower the rank, the less memory is required). After extraction, the weights were normalized. There is a direct correlation between rank and the acceptable LoRA strength range: for r32, keep the strength up to 120% at most; for r384, up to roughly 250%.

A lot of time, electricity, and compute went into this on my small and not particularly powerful Mac Mini M4. But it was an interesting research project.

Download them in the Files & versions tab.