Photographer turned open-source AI engineer. I spent fifteen years shooting and directing for artists like U2, Queen and Justin Timberlake — now I build the training tools I always wished existed, with the same obsession for the person in front of the screen.
Fizgig is an open-source AI training studio, and as of v5 it does something no other public tool does: full fine-tuning of 13B–33B image and video models on a single consumer GPU, down to 16 GB — measured, not projected (the trainer prints your run's peak VRAM every epoch so you can audit the claim on your own card). The only other public route to fine-tuning the same 33B model is DeepSpeed across eight 80 GB datacenter GPUs. The saved checkpoint is completely lossless.
It also trains, profiles, repairs, and extracts LoRAs for Flux 2 Klein, Krea 2 and MiniMax H3 — photos, video clips, sound and voice — behind a GUI built for creatives, not ML engineers.
| Tool | What it does |
|---|---|
| comfyUI-Realtime-Lora | Train, block-edit and save LoRAs live inside ComfyUI |
| ComfyUI-H3Studio | A video editor for MiniMax H3 in a single node — keyframes, timeline, continuity, export |
| ComfyUI-Angelo | Click-to-refine editing canvas |
| comfyui-mesh | Distributed inference across your LAN |
| torch-nvenc-compress | C++/NVENC tensor compression for moving latents over networks |
| comfyUI-LongLook | Long-video generation tooling |
…and twenty more across the generative-AI ecosystem.
Sony Imaging Ambassador and educator — from one-on-one masterclasses to university teaching and speaking at Bett, the world's largest edtech show. I also run UltrawideWallpapers.net (1M+ downloads), whose AI workflow — 40+ LoRAs trained on my own photographic archive — was featured by AMD.
I consult on local AI training pipelines, custom workflow tooling, and private model work — especially where client IP can't touch a cloud API. peter@shootthesound.com