Guides
Running models
Generate images and video with clockingcloud run: prompts, sizes, seeds, batches, inputs and results.
The basics
clockingcloud run z-image-turbo --prompt "a ceramic mug on a linen table, soft light" --out ./shotsclockingcloud run wan-2.2 --prompt "waves rolling in at dusk" --seconds 5 --out ./clipsclockingcloud run starts a short-lived replica, runs the job, writes files to --out, and releases the GPU. It is the right tool for experiments, one-off jobs and scripts. For steady traffic use clockingcloud deploy so warm replicas serve requests without a cold start.
Image options
| Flag | Notes |
|---|---|
| --width, --height | Default 1024. Must be multiples of 16. Larger sizes use more GPU memory and time. |
| --steps | Denoising steps. z-image-turbo is distilled and works best with 1 to 4. Others default to the model recommendation. |
| --guidance | How strongly to follow the prompt. Not used by distilled models. |
| --negative-prompt | Supported by Qwen-Image and Wan 2.2. Ignored, with a warning, by models that do not use it. |
| --seed | Same model, prompt, size, steps and seed give the same output on the same GPU type. |
| --n | Generate 1 to 16 images in one batch. A batch is cheaper per image than separate runs. |
| --image | Input image for qwen-image-edit editing. Local path or URL. |
Video options
| Flag | Notes |
|---|---|
| --seconds | Clip length. Longer clips take proportionally longer and are split across GPUs on multi-GPU models. |
| --fps | Frames per second. Defaults to the model recommendation. |
| --image | Required for wan-2.2-i2v: the first frame to animate. |
| --gpu | Video models default to h200 with four GPUs. Override only if you know the model fits. |
Licenses and tokens
Every catalog model is open and ungated under Apache 2.0. You do not need a Hugging Face account, a token or any proprietary key, and the license allows commercial use.
Waiting, background jobs and cancellation
By default clockingcloud run waits until the files are written. With --no-wait it prints a request ID and returns, and the generation continues on the platform.
ID=$(clockingcloud run wan-2.2 --prompt "city at night" --no-wait -q)
clockingcloud trace "$ID"Pressing Ctrl+C stops waiting but does not cancel the generation unless you passed --cancel-on-exit. --timeout bounds how long the CLI waits and never cancels work already started.
Output files
Files are named <request-id>-<index>.<ext> (png for images, mp4 for video) and a <request-id>.json sidecar records the model, parameters, seed, GPU, seconds and cost so any result can be reproduced. Use -o json for a machine-readable summary on stdout.
Something unclear or missing? Email hello@clockingcloud.com.