Family Studio

Open-source image and video models are now good enough to make a child's drawing dance or turn a bedtime story into an illustrated book. But the tool that runs them, ComfyUI, expects you to wire up graphs of nodes, pick model files and tune samplers. Nobody in my house except me was going to do that.

And a raw model will draw anything it's asked to. If four kids were going to use it, safety had to be part of the design, not a filter bolted on afterwards. On top of that, the one GPU in the house already runs Jake, the family's AI.

• Recipes, not graphs. Each card hides a tested ComfyUI workflow behind a short form with presets: restyle a photo, upscale photos and video, cut out a subject, text or photo to video, animate a photo, copy a style, a card maker and more. You see live progress and your place in the queue, and get a phone alert when a long job finishes.
• A kids' corner. My drawing comes alive turns a crayon drawing into a 3D or clay character and a short clip with sound. Make a picture book writes and paints a book (or uses the kid's own words) and lays it out as a PDF, with an optional read-along in Jake's voice. Learn to draw builds a step-by-step sheet and gives tips on a photo of the kid's attempt. There's also stop-motion, and a "Coach me" button that suggests a more vivid idea and explains why.
• Movies overnight. An AI director plans a film in short scenes, paints each opening frame and shoots it as clips with sound. The kids can even ask Jake for a movie: a parent approves it, it renders overnight, and in the morning it's on the kitchen screen.
• Family stars. Parents can train small add-on models (LoRAs) from photos, so a family member or pet can star in a scene. Style packs from Hugging Face go through a trusted-maker filter, test renders, a safety check and a parent's approval before anyone can use them.
• Profiles by age. "Who's making something?" Kids tap their name; the parent profile needs a password. Little ones get a short list of recipes, older kids get more. Every word a kid types passes a guard, and every result is checked by a vision model before they see it. Anything flagged waits for a parent. Kids never reach ComfyUI itself.
• Power users aren't boxed in. I can build a workflow in ComfyUI and click "Save to Studio": it finds the prompt, photo, size and length inputs on its own and turns the graph into a new recipe card. A local LLM and coding agents can run the same recipes as tools.

• One way to run a job. The app, the LLM and the coding agents all go through the same recipe runner, and every job is logged with who started it. That made auditing and testing simple, because there's only one path to check.
• Fail closed. The kids' rules are an allow-list. An unknown profile gets nothing, and an unchecked result stays hidden rather than showing by default.
• Sharing one 24 GB GPU. A small gate sits in front of the language model and a hook inside ComfyUI. Big video jobs unload the LLM first, small image jobs run next to it, and a chat waits for any render to finish. Studio runs on its own VM, separate from Jake's family stack.
• Family photos stay home. Backups are local only, with no cloud copy, because the inputs are pictures of my kids.
• Test the real thing. The drawing lessons were first built by erasing a finished picture backwards, and the steps didn't line up. Building them forward from a blank page fixed it. I only found that by running every kids' tool end to end before handing it over.

Built in October 2026 and in use by my family. It's private: it runs on my home lab and is only reachable from our own devices.

Next: star models for every kid, so each of them can wake up to their own "hero" clip on the kitchen screen, and more kids' recipes as they ask for them.

All projects  Park Family OS  Talk shop