Fizgig
A LoRA and fine-tuning studio for image and video generation models such as Flux 2 Klein, Krea 2, MiniMax H3 and Qwen Image, covering training, repair and remixing on consumer GPUs down to 16 GB.
Fizgig is a workbench for fine-tuning diffusion and image-generation base models on photos, video, sound and voice, from quick LoRAs up to full base-model fine-tunes. It can fix broken LoRAs without retraining them and remix an existing LoRA into new variations in seconds, and can also run on rented cloud hardware for users without a suitable GPU.
Recent releases add LoRA and LoKR training for Qwen Image down to 10 GB cards, reference-model files that turn a folder of photos into a reusable likeness file for MiniMax H3 with a companion tool for testing them, and repair-studio presets that boost detail and likeness in Krea 2 LoRAs. Explorer tools let a user compare a LoRA or reference file side by side against the base model before using it elsewhere.