Models

The set

The app groups its tools into Image, Video and Audio, and every tool starts on a default model with the others a pick away. The models page has what each one is good at, its GPU class, timing and license; this page is about running them.

Turning a model on

Z-Image Turbo, Qwen Image 2.1, MiniMax H3 are on in a fresh install; the setup wizard creates their endpoints. Every other model starts off. To use one, open Settings → Models and flip its switch. The app then:

1. Creates a serverless endpoint named cozygpu-<model> in your RunPod account: at most one worker, a ten second idle timeout, FlashBoot on, the GPU pools the catalog lists for that model.
2. Waits while RunPod pulls the model's worker image to a worker. This is the one slow step: 10 to 30 minutes depending on the image (about 7 minutes for OmniVoice, about 27 for MiniMax H3), and it happens once per endpoint, not once per job.
3. Runs a small warm-up job so you can watch progress; video models cancel it as soon as the worker is up.

The tools that use the model unlock as soon as the switch is on; a tool whose model is still off shows Turn on in Settings → Models in place of its generate button. Turning a model off stops nothing that is running and leaves the endpoint in your account, where it costs nothing while idle; delete it in the RunPod console if you want it gone, and the app recreates it (and pays the pull again) the next time the model is turned on.

A model whose worker image is not published yet shows Coming soon instead of a switch. The app checks the signed catalog on its own and the switch appears when the image lands; no reinstall.

Worker slots

RunPod limits how many serverless workers an account may have at once, and a new account starts with a small limit. Every model that is on holds one endpoint with a maximum of one worker, so each model takes one slot whether or not it is busy. If RunPod answers with a quota error when you turn a model on (the app shows RunPod says this account has used up its worker quota), turn off a model you are not using, or ask RunPod to raise the limit from the console; limits grow with account history and spend.

Models that need the S3 key

SCAIL-2 and WAN 2.2 Animate take a motion video, and videos travel to the worker through a network volume in your account, so the Motion Control tool needs the RunPod S3 API key and at least one volume (the app offers to create it). The same is true of custom LoRAs and of any input over 7 MB on any model. See Keys.

Billing

RunPod bills per second of GPU time in your own account; nothing runs while idle. An endpoint that is on but idle has no worker and no bill; the first pull is a download, not GPU time. Set a spend limit under Billing in the RunPod console if you want a ceiling.

Settings per model

Every tool shows a few common settings up front and the rest under Advanced. Seed is on every model (random, or fixed to reproduce a result). The per-model pages describe each setting; this is the list.

Custom LoRAs are a setting on Z-Image Turbo, Qwen Image 2.1, Krea 2 Turbo, Ideogram 4 and MiniMax H3; see Custom LoRAs.