Product shot → video, in four steps

A person who is not real, wearing a product that is not photographed, in a room that does not exist — then moving. Every step is a generator you can also run on its own; this page is the wiring between them, and the run buttons above are the actual thing, not a description of it.

Every step below shows a real run of its own settings before it asks you for anything, and the clip at the top is what following all four produces.

Why this order

• Compose before you animate. A bad still costs seconds to redo. A bad clip costs minutes. Get the frame right first and the video step becomes cheap.
• Plain backgrounds in, complicated backgrounds out. The compositing model relights what you give it. Give it a clean silhouette and it will put the subject anywhere; give it a busy one and it spends its effort deciding what is even the subject.
• Small motion. Image-to-video models are best at a breath and worst at a plot.

Starting halfway

Every input on every step takes an upload, so you are never made to generate something you already have. Bring a real product photo and a real model release, drop them into step three, and the first two steps simply do not apply to you.

What it costs

Each step is an ordinary job at the Low tier and is priced per GPU second like any other. The run buttons show the estimate before you press them, a failed job refunds its hold in full, and nothing you make here is public until you publish it yourself.

Running it somewhere else

All three graphs are published in full, with their model lists and setup notes:

• Promo image — FLUX.1-dev fp8
• Subject + product → one photo — Qwen-Image-Edit 2509
• Image + prompt → video — Wan 2.2 I2V

Take them and run the chain on your own hardware. That is what they are published for.