Fizgig is an open-source LoRA and fine-tuning studio for today’s image, video and music models, built to train on consumer GPUs and to fix broken LoRAs without retraining from scratch. It covers the whole loop: prepare and caption data, train, compare epochs visually, then profile, repair and extract adapters. The project is released under the Apache-2.0 license, has about 584 stars on GitHub, and was actively updated in October 2026.
Fizgig targets a broad set of current base models, including the Klein and Flux families, Krea 2, MiniMax H3, Qwen Image, Z-Image Turbo, SDXL and Anima for images and video, and Ace-Step for music. Training works on photos, video, sound and voices, from quick LoRAs to full base-model fine-tunes, and the project says consumer cards down to about 8 GB of VRAM are usable. Two features stand out for working artists: broken LoRAs can be repaired rather than discarded, and checkpoints can be compared by eye to find the best epoch. If you have no suitable GPU, the studio can also run on rented hardware in one click.
The project releases frequent updates as new base models appear, and the workbench also covers the chores around training, including dataset captioning and LoRA profiling, so a complete fine-tuning project can stay inside one application.
Fizgig is free and open source under the Apache-2.0 license. Running it locally costs nothing beyond your hardware; rented GPU time, if you choose it, is billed by the hardware provider.
Fizgig is for AI artists and model trainers who fine-tune diffusion and generative models locally and want a complete workbench instead of scattered training scripts.