ComfyUI Studio
A private video and image studio running on my own RTX 5090, reachable from this site, and the first AI ad it rendered overnight.
I got tired of paying hosted AI video services per clip while a 5090 sat under my desk doing nothing at night. So I built a small studio around ComfyUI that a handful of people I trust can use from their phones.
How it works
- The page lives at xs.dev/studio. You enter a personal key, pick text to video, image to video or text to image, choose an ad format (Reels, Feed, Square, Wide, up to 4K for short clips) and hit generate. A shared gallery shows what everyone has made.
- A gateway on the PC (Node, no dependencies, SQLite) checks the key, queues one job at a time, fills a fixed ComfyUI workflow and streams the result back. ComfyUI itself is never exposed.
- A Cloudflare Tunnel connects the PC to the internet without opening a single router port. Keys can be revoked in one command.
- Models: LTX 2.5 (22B, distilled) for video with sound, Qwen-Image with the Lightning LoRA for stills. On the 5090 a 5 second 1080p clip takes about a minute, a 2K image about 12 seconds.
The first real job: an ad, rendered overnight
To test it I gave Claude a shot list for a 40 second commercial for Alfred, the AI marketing service I sell by day, and went to bed. The workflow borrows from how the big AI ad shops do it: generate a handful of reference stills first (the hero, his van, the kitchen, the van cab, a front door, a home kitchen), lock them, then animate each shot from its still so the light and the character stay the same across cuts.
By morning: six stills, eight clips at 1080p, cut together with crossfades and Swedish captions, with only the sound the model generated. One clip needed a second take and the GPU needed a memory reset halfway through, both of which the agent handled on its own. Not broadcast ready, but a real first cut from a single prompt list, for the cost of one night of electricity.
What I learned
- Reference stills are the whole trick. Without them every clip drifts.
- 1080p is the sweet spot on 32 GB of VRAM. 4K works for 3 second hero shots and nothing more.
- A shared queue needs guard rails: per size duration caps, one job at a time, a few waiting per person.
- Presets beat empty text boxes. The page now ships with ready prompts for the kinds of ads a local business actually runs.
The studio is private and stays that way. If you know me, ask for a key.