shootthesound/Fizgig

Krea 2, MiniMax & Klein 9B LoRA - LoKR Studio — train, profile, repair, and extract Krea 2, Flux 2 Klein 9B & MiniMax LoRAs & LoKRs

What it solves

Fizgig is a comprehensive workbench for training and refining Low-Rank Adaptation (LoRA) models for specific AI image and video generators. It addresses the common problem of "overbaked" or imperfect LoRAs by providing tools to fix, mutate, and compare them without needing to complete a full retraining process.

How it works

The software provides a fast, VRAM-efficient trainer that supports the Flux 2 Klein 9B, Krea 2, and MiniMax H3 models. It uses techniques like block swapping, quantization (INT8, NF4, 4-bit HQQ), and weight averaging (EMA) to allow training on consumer GPUs as low as 8GB to 16GB. Beyond training, it includes a "workbench" of post-processing tools:

  • Repair Studio: Allows block-by-block adjustment of a LoRA using sliders to fix identity or style issues.
  • LoRA the Explorer: Uses evolutionary mutation to discover new LoRA variations.
  • LoRA Royale: Renders multiple epochs on a single seed to find the best version via crossfading.
  • Profiler: Generates HTML reports to identify which blocks carry style or identity.
  • Extract: Distills LoRAs to a lower rank for efficiency.

Who it’s for

AI artists and developers who want to fine-tune models on photos, video, sound, and voices, and who need precise control over the final output without the overhead of professional-grade hardware.

Highlights

  • Low VRAM requirements: Train large models (up to 33B parameters) on consumer GPUs (8GB-16GB).
  • Post-training surgery: Fix broken LoRAs block-by-block without retraining.
  • Multimodal support: Supports training on images, video clips, and audio/voice recordings for MiniMax H3.
  • Dataset curation: Experimental features for Krea 2 that detect problem images and auto-recaption stuck images mid-run.
  • Full fine-tuning: Ability to train the base models themselves and then distill the result into a shareable LoRA.

Related

  • 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.

  • Dispatch
    Fine-Tuning FLUX.1-dev with QLoRA on Consumer Hardware

    Hugging Face demonstrates how to fine-tune the FLUX.1-dev model using QLoRA and the diffusers library to reduce peak VRAM usage to under 10 GB on a single consumer GPU.

  • 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.

  • 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 20B LLMs with RLHF on a 24GB Consumer GPU

    Hugging Face released an integration of TRL and PEFT that enables reinforcement learning fine‑tuning of 20‑billion‑parameter language models on a single 24 GB GPU using 8‑bit quantization and low‑rank adapters.