LoRAs are extracted from fine-tuned checkpoints (Schnell), and the official Flux.1 Schnell is also used as a basis for comparison. As a result, the resulting differential LoRAs do not contain four-step Schnell optimizations and can therefore be used with both Schnell checkpoints (at 4–6 steps) and DEV checkpoints (with full accuracy and the full number of steps). I did this primarily for myself because LoRAs are easier to work with, including mixing them in real time and setting different weights/strengths. This also saves space on your local computer. Choose a rank according to your taste and the capabilities of your hardware (the lower the rank, the less memory you need). After extraction, the weights been normalized. There is a direct correlation between rank and a LoRA's acceptable strength range. So for r32, keep the LoRA strength at up to 120%, and for r384, at about 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.