← all repositories
shootthesound/Fizgig

The LoRA trainer that keeps working after training ends

Fizgig trains LoRAs on 8 GB cards, but its real pitch is the workbench: repair, remix, and profile LoRAs you already have.

Fizgig
Collecting fresh signals — velocity needs a few days of history.
collecting data…
star history

What it does

Fizgig is a local training studio for a wide roster of image models — Flux 2 Klein 9B, Krea 2, MiniMax H3, Qwen Image 2.1, Z-Image Turbo, SDXL, and Anima — covering LoRA, LoKR, slider/edit LoRAs, and full fine-tunes. It claims LoRA training from 8 GB VRAM and full fine-tunes from 8–16 GB depending on model. It also handles dataset prep: AI captions, face crops, and for MiniMax H3, video and voice datasets with Whisper transcription.

The interesting bit

Most trainers stop when the .safetensors file lands. Fizgig’s workbench picks up there: Repair Studio lets you fix a broken LoRA block-by-block with live previews instead of retraining, the Profiler colour-codes which blocks carry identity vs. style, and LoRA the Explorer mutates a LoRA evolutionarily while you pick favourites. The tools also work on LoRAs you downloaded from anywhere, not just your own runs.

Key highlights

  • Repair Studio: per-block sliders, donor-LoRA blending, exact baked saves — no retraining needed
  • LoRA Royale renders every epoch on one seed with crossfades, so you pick the best epoch by eye
  • Context LoRA: train on top of a frozen, active LoRA so a face and a style coexist — the README says no other trainer does this
  • The trainer curates your dataset mid-run: throttles, recaptions, or sets aside images that stop teaching
  • Loads LoRAs from kohya, PEFT, OneTrainer, AI-Toolkit, and LyCORIS; saves ComfyUI-compatible files

Caveats

  • Full fine-tuning is marked experimental for every model, as are Z-Image Turbo, SDXL, and Anima support
  • NVIDIA RTX 30/40/50 or AMD ROCm only, with 32 GB system RAM recommended — the low VRAM numbers come with fine print
  • The README’s claims like “no other trainer does this” are the author’s own, not independently verified

Verdict

Worth a look if you train character or style LoRAs and have ever wished you could fix one after the fact instead of starting over. If you just want a one-shot trainer and never touch the file again, kohya already exists.

Frequently asked

What is shootthesound/Fizgig?
Fizgig trains LoRAs on 8 GB cards, but its real pitch is the workbench: repair, remix, and profile LoRAs you already have.
Is Fizgig open source?
Yes — shootthesound/Fizgig is open source, released under the Apache-2.0 license.
What language is Fizgig written in?
shootthesound/Fizgig is primarily written in Python.
How popular is Fizgig?
shootthesound/Fizgig has 501 stars on GitHub.
Where can I find Fizgig?
shootthesound/Fizgig is on GitHub at https://github.com/shootthesound/Fizgig.

heatdrop uses Google Analytics to see which pages get read — nothing else. Your call. How we handle data.