request_asset_upload registers a cast; it does not generate a scene
Reusable characters and products are one named setup job. Prompting a face every scene, or stuffing a logo into memory, spends generate credits on identity you already had.
request_asset_upload registers a cast. It does not generate a scene. It does not store hex codes. It does not "remember" a face as a sentence. Calling it is how the upload card appears — if the agent only writes "please upload a photo," nothing renders and you sit there. Then, if you skip the card and describe the spokesperson in every prompt, you buy a cousin in every shot.
Set up reusable characters and products is the named agent job. Upload once per character or product. Reuse across every scene. The tool opens an Asset Builder card for reference images of a character, product, prop, or location. Batch opens several cards for a full cast. Required: key, suggested name, type, description, plus exactly one of workflow_id or slideshow_schedule_id. Assets belong to that build, not a global pool. Uploading does not use generation credits. Generation uses the keys later.
Character consistency is the reason. Words do not pin a face. Pixels do.
What the job is
Onboarding. One asset at a time, wait for the saved confirmation, then the next — or batch when you know the whole cast. Hosted URLs can skip the card via create_workflow_asset. Both paths converge on a named record the scene director can reference. The tool itself writes nothing to the database; the frontend save does.
This is setup you do before the first paid scene, not a recovery you invent after scene four drifted.
What it is not
- A generate. Text-to-image or video rows mint pictures of "a woman in a red jacket." That woman will not survive shot two. Register the approved stills, then generate against the key.
- A brand kit. Set my brand kit is colors, fonts, logo, tone, default aspect. A spokesperson's turnaround is not a hex. A SKU pack shot that must match across scenes is an asset. Mixing those stores is how the logo is "remembered" as prose and never attached.
- Long-term memory. Teach a lasting preference is
remember: "always 9:16," "keep captions minimal." A face is not a preference.rememberwill not bind reference images. - The movie itself. Create a multi-scene AI movie and save a workflow consume the cast. They do not replace onboarding. Start those without assets and you are back to adjectives.
The waste is rerolling identity. Asset upload is free of generation credits; the scenes you regenerate to chase a face are not.
The test
Is there a person, product, or place that must be the same pixels across more than one scene?
Yes: this job, first. Open the card. Save the key. Then generate.
If it is a one-off extra who will never return, skip the registry. If it is a brand color, use the kit. If it is a standing rule with no file, use memory.
FAQ
Why can't I paste image URLs in chat instead of the card?
Because nothing in the pipeline treats a chat paste as a workflow asset. The card is the collection step. Skip it and the key does not exist.
Does registering the cast generate the first scene?
No. Zero generation credits for the upload. The first scene is a later confirm on a video or slideshow model.
Can one asset serve every workflow in the account?
No. An asset is bound to a workflow id or a slideshow schedule id, never both, never neither. Clone or re-onboard when you start a new build. That scoping is what stops a mascot leaking into an unrelated job.
I already have a brand kit logo. Isn't that the product shot?
A logo in the kit is a brand default. A product hero that must match labeling across twelve shots is a reference asset. Use both. Do not ask the kit to be a turnaround sheet.