Any Model, Any Time: Load It Into the Generator Yourself

TL;DR:

• Every checkpoint works in the generator. Any SafeTensor checkpoint on a supported base model, loaded or not.

• Not loaded? A member can pull it in, free. Your job waits in the queue until it's ready.

• Members download faster, and can Boost with Buzz to skip the line.

• Members-only to start, while we watch the load. We're opening it to everyone shortly.

Getting a checkpoint into the generator used to mean winning an auction. Now any checkpoint on a supported base model works. If it isn't loaded yet, we download it for you — and to start with, that last step is a member one, just while we see what it does to the queue.

How it works

Step 1. Open any model. Versions already in the generator have a green dot, and those work for everyone.

Step 2. Pick one that isn't loaded and hit Create — members only for now. The generator tells you it needs to download first, and about how long that'll take.

Step 3. Hit Generate. Your job waits in the queue and starts when the download lands. The queue shows your lane, your place in line, and the progress.

Download lanes

Every download waits in one of three lanes. Each lane is its own line with its own speed limit. No lane cuts in front of another.

• Standard is free. It's the slowest, and it's the one that opens up when we lift the members-only step.

• Priority is for members, with a higher speed limit.

• Express is Boost. No speed limit. You see the Buzz price and how much time it saves before you pay, and one boost covers everything the job needs.

A download only stays in line while you have a job waiting on it, and that job holds one of your generation slots until the download finishes. That keeps the free lane from filling up with downloads nobody's going to use.

Checkpoint auctions

Auctions are staying. Win one and your checkpoint is featured in the resource picker, and anyone who loads it gets Express for free. That's how it works for now, and we may change it as we see how people use this.

Worth knowing

• Members pull in the new ones at first. Anything already loaded works for everyone, and the free lane is built and waiting — this is a setting we turn up, not something we still have to write.

• Checkpoints need a SafeTensor file. Diffusers, Core ML and ONNX checkpoints won't load.

• Only base models the generator already runs. This opens up checkpoints, not new ecosystems.

• Times are estimates. They move with demand.

• Loading it doesn't guarantee it works. Most checkpoints run fine, but models vary, and we can't promise every one behaves in our generator the way it does on the system it was made on. We're monitoring this.

Pick a model. Hit Generate. Tell us what broke.