Fizgig Adds Live Reference Tools to Its Free LoRA Studio

    
        By vramkickedin    
     | 
    
            June 15, 2026 at 2:33 pm        
    
     | 
    
        2 min read

Fizgig, a free open-source LoRA studio purpose-built for Flux 2 Klein 9B, has been updated to version 1.7 with a set of live control and reference-image tools. The release lets you train, repair, remix, and profile LoRAs on a single NVIDIA card, with new options for conditioning previews on real photos and tweaking generation settings mid-run. It turns a training session into a real-time editing workspace that requires no cloud access and keeps all data local.

Developer Shootthesound who also developed ComfyUI-Lora-FindingLora and ComfyUI-Angelo, created Fizgig to fill the gap that most trainers leave after training finishes. Where other tools stop, this workbench begins, offering block‑level repair, evolutionary variation discovery, and side‑by‑side epoch comparison. The project is now available under an Apache 2.0 license and targets users who want full command over their Klein 9B LoRAs without retraining.

Reference-guided previews and live training controls

Key additions in v1.7

Reference-image conditioning for previews and samples.
Live VRAM and system RAM usage bar.
Change sample prompts mid-run without restarting.
Toggle gradient checkpointing off for faster steps.
Distilled preview model cached in RAM between epochs.
Auto block-swapping prevents crashes on 16 GB cards.

Privacy-conscious artists and small studios running Klein 9B locally will find the new workflow especially helpful. They can repair overbaked character LoRAs, remix them into fresh variations, and visually pick the best epoch, all on a single 16 GB GPU. Seeing how a LoRA edits a reference image while training or fine‑tuning gives an immediate, intuitive check that was previously hard to achieve outside of a full ComfyUI render.

Developer notes and open-source licensing

Shootthesound highlights that Fizgig is written entirely for Klein 9B, making block‑aware repair and evolutionary exploration possible without bolting onto generic backends. The new gradient checkpointing toggle and fp8 matmul speedups give RTX 40‑ and 50‑series owners faster steps at no extra memory cost, while 4‑bit NF4 mode keeps training viable on 10–12 GB cards. Future improvements will likely focus on extending the profiling and extraction tools based on community input.