Show Hiroska a dozen pictures that share your look, whether that is your illustrations, your product shots or your colour grade. Every image and clip afterwards can come out in it, with nothing extra in the prompt.
A look, a subject, or your footage
Three kinds, and the difference matters. A look is trained from pictures that share a style. A subject is trained from one product or person, so it can appear in scenes you never shot. A video look is trained on your own clips.
A look: the same colours and feel across everything you make
A product or a person: the same subject, in any scene you can describe
A video look: trained on your own clips, for video you describe from scratch
Trained once for the workspace rather than per project
Pick pictures, name it, walk away
Choose the files from your library, give the style a name, and start. Training runs on our side, so you can close the tab and come back later. What comes back stays in your workspace until you delete it.
Search the whole library from the picker, not only the last page of it
Ready in minutes on the standard model, with the price on the button before you press it
Nothing to type when you use it: the style is applied for you, at a strength you set
Came out too faint? Train the same files again in one click, with no re-picking
The difference between AI output and your output
A prompt describes a look once. A trained style holds it across a forty-shot series, which is what makes a channel recognisable, and it does so without being described again.
Consistency across a series rather than one lucky generation
Works with the editor: generate in your style, then cut, grade and export as usual
Combine it with a cloned voice and a project template and a whole episode is on-brand
Priced from what the training costs us, in tokens, and refunded if a run fails
Questions
How many pictures do I need to train a style?
Around a dozen is plenty. What matters is that they share one look: same colours, same feel. Twelve consistent pictures teach more than forty with nothing in common, and a mixed set is the usual reason a style comes out weak.
Do I have to write anything special in my prompts afterwards?
No. The style has a tag of its own and we add it for you, every time. You write the prompt you would have written anyway.
What if the style barely shows in the result?
Turn the strength up with the slider under the style, which fixes it most of the time. If a full-strength result still looks untouched, train the same files again from the style card: one click, no re-picking.
Can I train a style on photos of a person?
Yes, if it is you or you have their explicit written permission. We ask which before training and record the answer with the wording you agreed to. Training on somebody’s likeness without permission is unlawful in many places, and the account that trained it is responsible for how it is used.
What does training a style cost?
Tokens, priced from what the training costs us and shown on the button before you start. If a run fails the tokens come back automatically, and the same applies if you delete a style while it is still training.
Is a trained style private?
It belongs to your workspace and is never offered to anyone else. The trained file is stored for you and deleted, permanently, when you delete the style.
Can I use a trained style in the video editor too?
Yes. A trained style is offered anywhere you generate, including from a clip inside the editor. Image looks apply to generated images, and a video look, trained on your own clips, applies to generated video.