The docs' 3+4+5 second UGC avatar ad: $2.94 at Plus, $6.60 at Max

Sume's own multi-scene example is a UGC-style ad with three scenes: a 3-second hook, a 4-second silence beat and a 5-second call to action. That plans 12 seconds, so at the tier rates it is $2.94 at Plus and $6.60 at Max. Standard is 12 x $0.184 = $2.208, shown before any cent rounding.

The scene plan

The request uses ordered video_inputs. Scene hook has a text voice with duration: 3, scene demo is voice.type: silence with duration: 4, and scene cta has a text voice with duration: 5. All three use the same background prompt, casual bedroom framing with native UGC lighting.

Cost

Rates: Standard $0.184, Plus $0.245, Max $0.55 per second without a product image; $0.194, $0.258, $0.58 with one (Sume pricing page, as of 2026-10-08). The docs say the total planned duration must stay in the 4 to 60 second window. They do not say a silence beat is free, so budget the full 12 seconds.

Making it yours

Silence beats require a duration and forbid script and input_text. Spoken scenes use type: text with one of script or input_text, not both. Because the current execution supports one resolved avatar and one shared scene per final video, keep the background prompt identical across scenes, as the example does. Swap the hook line to test openings; the other two scenes can stay the same.

Practical notes

Output format. Avatar video output is 720p at this time, and the aspect ratio defaults to 9:16 with 1:1, 3:4, 4:3 and 16:9 also accepted. Choose the ratio in the request rather than cropping afterward, since a crop throws away part of a frame you paid for by the second.

Job lifecycle. Every Sume avatar request is job-backed. You submit, store the job id, and poll GET /v1/jobs/{id}/status until it completes, then read GET /v1/jobs/{id}/result. The shared lifecycle also lets you wait with sync or subscribe mode for a limited time instead of polling, and the result carries Sume-hosted artifacts rather than provider URLs, so you never handle a provider queue id. Plan your client around the job, not around a single blocking response.

Media inputs. Any media field, whether product_image, scene.image_url or an avatar image_url, must be a fetchable public HTTPS URL. Sume rejects localhost, private-network and non-HTTPS addresses before it submits anything, so a broken link fails fast instead of consuming a render. Host reference images where Sume can fetch them and keep them available until the job finishes.

Idempotency. The docs examples send an Idempotency-Key header on every create call. Use one key per intended job and reuse it only when you retry the same request. A network timeout followed by a fresh submit with a new key is the one path that can create, and bill, a second job. Keep keys human-readable, such as the campaign and scene, so a support lookup is easy.