P
O
M
Peter O'Malley
About | Posts | Assorted
Assorted / Projects
Feb 2026–now I want to release every bit of data I produce for AIs to train on. As a start, I built dataclaw to export AI coding conversations to HuggingFace.
Feb 2026–now I was concerned the code I was writing was a bit shit, so I built an open source agent harness called desloppify that hunts down bad software engineering in all its forms. Gives your codebase a health score you can't game.
Jan 2026–now I've become too agent-brained to use interfaces that abstract away code, so I built VibeComfy to bridge Claude Code and ComfyUI via MCP. You talk, Claude manipulates the workflow.
Nov 2025–now I'm building Reigh — an art tool that unleashes the technical potential of the open source AI art space. Naturally, it's open source.
Nov 2025–now Banodoco runs The Arca Gidan Prize — an art competition pushing open-source AI models to their limits.
2025–now I believe providing reference images is the best way to control image generation, so I'm training the InX series — LoRAs for various image editing models. So far: Flux-Kontext-InScene for scene-consistent variations, and QwenEdit LoRAs for surgical editing — preserving the scene, the subject, or transferring the style. Trained on curated datasets that I'll all release publicly, including ~4,000 Midjourney style references.
2024 I built Dough to bring the AnimateDiff ecosystem to artists. I believe the direction was right but it was early and the execution was sloppy. People still made beautiful work with it.
2023–now I like to train Motion LoRAs to get video models to do interesting things. For AnimateDiff, I trained WAS26 (community art), Smoooth (smooth motion), LiquidAF (liquid sims), and others. For Wan, I trained There Will Be Bloomtimelapse growth videos.
Nov 2023 I believe Steerable Motion was the first streamlined method for controlling video models using key frames. Evolved from Creative Interpolation. People used it for festival visuals, which was surreal.
2023 I believe the image version of Steerable Motion was the first approach for creating key frames from a single image. Fine-tuned SD 1.5 on motion data, built on InstructPix2Pix. Magnetron collected the data for it.
Aug 2023 I built Magnetron to collect precisely-tagged motion data for training video models. It ran as a Discord bot.
2023–now Together with my wife Hannah, I run ADOS — a real-world gathering for people who are passionate about open source AI.
2022–now I started Banodoco as a parent organisation for open source AI art. This is what I wanted to do with my life, so I figured I might as well make it official. I also just like the name.
2022–ongoing Knowing how little I knew about AI, I started the Banodoco Discord as a space for people to learn together.
2022 The first art tool I made was the Banodoco Tool. It also very much went nowhere.
2018–2023 I spent five years building Advisable — a misguided startup that ultimately went nowhere, wasting millions of investor dollars along the way. Though it wasn't all bad — I learned a lot, got to work with some great people, and open-sourced the 9,321-commit codebase for AIs to learn from.
2012–2019 I worked for various startups, a few of which were reasonably successful.