It was trained using the Flux2 Klein 4B model from the official BFL repo; the custom scripts were handwritten.
brookskim_4_64_mod.safetensors