shootthesound/Fizgig

Fizgig — LoRA & Fine-tune Studio for Klein 9B, Krea 2, MiniMax H3, Qwen Image 2.1, Anima & Z-Image Turbo and Ace-Step XL: train, fine-tune, profile, repair and extract LoRAs & LoKR for Image, Video & Music Models

Fizgig – A LoRA & Fine‑tune Studio for Modern Generative Models

What it is

Fizgig is a desktop‑plus‑web application that lets you train, fine‑tune, and edit generative AI models on a single consumer‑grade GPU (as low as 8 GB) or on rented cloud GPUs. It focuses on LoRA (Low‑Rank Adaptation) workflows but also supports full‑model fine‑tuning, “LoKR” training, and special “slider / edit” LoRAs that let you blend between two visual styles with a single dial.

Core capabilities

Feature What you get Why it matters
Multi‑model support Flux 2 Klein 9B, Krea 2, MiniMax H3, Qwen Image 2.1, Z‑Image Turbo, SDXL, Anima, ACE‑Step 1.5 XL Covers the most popular image, video, and music generation models released in 2023‑24.
LoRA creation & repair Train LoRAs, edit them block‑by‑block, blend donors, and export ready‑to‑use .safetensors for ComfyUI. Saves you from re‑training when a LoRA is slightly off‑target.
Slider / Edit LoRAs A single LoRA whose strength is a slider between two looks (or an edit such as relight). Gives intuitive, real‑time control over style or image edits.
Full fine‑tuning Train the entire base model on 8‑16 GB cards (experimental for most models). Lets you customise the underlying model when LoRAs aren’t enough.
Work‑bench tools Studio (live preview & sliders), LoRA Royale (epoch‑by‑epoch visualisation), Profiler (per‑block identity/style report), Extract (rank reduction). All the post‑training utilities you need in one place.
Automatic dataset curation While training, the trainer watches loss per image, throttles, re‑captures captions, or discards images that stop teaching. Improves data efficiency without manual cleaning.
Cross‑device / cloud execution Run locally, on another PC on the same network, or on a RunPod GPU with a single click. Phone can monitor runs via QR‑code. Removes the need for a powerful workstation; you can rent a GPU for a single run.
Presets & adapters Model‑specific presets, training adapters for Qwen Image 2.1 and Z‑Image Turbo, and a “Ultra mode” for Krea 2. Reduces the fiddly configuration work for each model.
CLI & headless mode All trainers are scriptable; you can launch runs without the UI. Fits into pipelines or remote servers.

How it works (high‑level)

  1. Prepare data – optional resizing, face‑crop, and a Look Consistency Filter.
  2. Caption – either simple trigger‑word captions or AI‑generated captions using Qwen‑3‑VL / Florence‑2.
  3. Pick a model & preset – the UI loads the appropriate training adapter and VRAM‑aware settings.
  4. Start training – the trainer monitors VRAM, swaps precision/block‑wise, and queues subsequent runs.
  5. Explore results – use Studio sliders, LoRA Royale epoch movies, or the Profiler to decide which epoch to keep.
  6. Export – saved LoRAs or full checkpoints drop straight into ComfyUI or other pipelines.

Who it’s for

  • Artists & hobbyists who want to customise diffusion models on a laptop or a modest desktop.
  • Musicians (ACE‑Step) who wish to train LoRAs on their own songs.
  • Researchers / developers needing a quick UI to prototype LoRA‑based style transfer, edit‑style LoRAs, or full fine‑tuning without writing training scripts.
  • Teams that want a shared, web‑accessible interface and the ability to rent a cloud GPU for occasional heavy runs.

Installation & requirements

  • GPU: NVIDIA RTX 30/40/50 series (or AMD Radeon with ROCm). Minimum VRAM 8 GB (some models need 10‑16 GB for full fine‑tune).
  • OS: Windows 10/11 or Linux.
  • Python: 3.10‑3.13.
  • System RAM: 32 GB recommended.
  • Install via the provided install_fizgig.bat (Windows) or install_fizgig.py (Linux). A one‑click RunPod deployment is also offered for those without a GPU.

License & community

  • Apache 2.0 – free for personal, academic, or commercial use.
  • The project bundles permissively‑licensed components (kohya‑ss, AI‑Toolkit, etc.) and provides links to model adapters on Hugging Face.
  • Sponsorship via Buy‑Me‑A‑Coffee or GitHub Sponsors helps sustain development; the RunPod deployment also supports the author.

Bottom line: Fizgig is a full‑featured, open‑source studio that brings LoRA creation, repair, and full model fine‑tuning to consumer hardware, with a polished web UI, automatic dataset handling, and seamless cloud‑GPU integration. It is very much in scope for AI/ML practitioners interested in generative model customization.

Related

  • Dispatch
    Using LoRA for Efficient Stable Diffusion Fine-Tuning

    Hugging Face has integrated Low-Rank Adaptation (LoRA) into the diffusers library, enabling faster Stable Diffusion fine-tuning with significantly lower VRAM requirements and model weights as small as 3 MB.

  • Dispatch
    Fine-Tuning Gemma Models in Hugging Face

    Hugging Face provides a guide on using Parameter-Efficient Fine-Tuning (PEFT) and Low-Rank Adaptation (LoRA) to customize Google's Gemma models on GPUs and Cloud TPUs.

  • Project
    shootthesound/comfyUI-Realtime-Lora

    A comprehensive ComfyUI toolkit for training, analyzing, and selectively loading LoRAs across multiple model architectures, featuring granular block-level and sub-component control.

  • Project
    perfectgf/lora-dataset-studio

    A browser-based studio for building, curating, captioning, and training image datasets for LoRAs, integrating ai-toolkit and ComfyUI.

  • Project
    huggingface/finetrainers

    A library for the accessible training and fine-tuning of diffusion models, focusing on memory-efficient text-to-video and text-to-image generation.