Midjourney's Best Trick Is Still Trapped Inside Midjourney

Midjourney just shipped consistent characters — same face, same person, across multiple generations — and it's good. Like, actually solves a real problem good. Every comic artist, every storyboarder, every person who's tried to use image generation for anything with a narrative arc has hit the wall where your protagonist has a different nose in every panel. This is the feature that makes image generation a tool instead of a toy.

And it will stay inside Midjourney's Discord server, behind their subscription, controlled by their roadmap, subject to their content policies, available right up until it isn't.

The thing is — I want this. I want it in a model I can run locally, that I can fine-tune on my own characters, that can't be revoked in a terms-of-service update. The technical work here is genuinely impressive and the application is obvious. It's just that "impressive technical work with an obvious application" is also a description of a thing that someone should open source, and nobody has.

Stable Diffusion exists. The tooling ecosystem is real. But consistent character generation at Midjourney quality isn't sitting in a Hugging Face repo right now. So we're back to applauding features we don't own, in a model we can't inspect, from a company that has been notably cagey about what's actually under the hood.

Good feature. Wrong hands.