Recast one video ad with several faces: consent checklist, cost

You can reuse one working ad with several different people by running H3 Max Recast once per person: same source video, one photo each, one paid job per variant. On Sume a 15 second ad is $5.63 per variant at 768p and $8.44 at 1080p, so three creator versions at 768p come to $16.88. The harder part is paper, not pixels: each person whose photo you submit needs to have agreed to be in your ad.

What does a set of variants cost?

Sume bills fal's per-second list times 1.25. The source length rounds up to whole seconds, and a variant is a separate job with its own hold, so the total is simply variants times the clip price.

• The totals use exact dollars before rounding ($5.625 and $8.4375 per variant), then round to cents.
• Add retries. Sume has no seed on this model, so a variant you reject is a new paid take.
• Photos are free and the number of people in the source does not change the price.

How do I run the batch?

Give each variant its own Idempotency-Key, derived from the creator and a version so that a network retry of the same variant is safe and a new take is distinct. The loop below submits three jobs and prints the raw responses so you keep every job id. Poll each job's status or use a signed webhook instead of holding connections open. Your workspace's concurrency and queue limits apply; extra jobs wait as queued until there is room.

import os
import requests

URL = "https://api.sume.com/v1/video-router/generate"
HEADERS = {"Authorization": f"Bearer {os.environ['SUME_API_KEY']}"}
SOURCE = "https://example.com/winning-ad.mp4"
CREATORS = {
    "anna": "https://example.com/anna.png",
    "ben": "https://example.com/ben.png",
    "chloe": "https://example.com/chloe.png",
}

for name, photo in CREATORS.items():
    body = {
        "model": "h3-max-recast",
        "video_url": SOURCE,
        "reference_image_urls": [photo],
        "resolution": "768p",
        "duration": 15,
        "mode": "async",
    }
    key = {"Idempotency-Key": f"ad-variant-{name}-v1"}
    r = requests.post(URL, json=body, headers={**HEADERS, **key}, timeout=60)
    print(name, r.status_code, r.text[:200])

Which source ad works best?

Pick a clip that fits the model's envelope. Recast takes 5 to 30 seconds with no single shot over 15, replaces the main people in the video, and accepts one to four photos. An ad with one clear on-camera person and quick cuts is the simplest case. A 15 second spot with a handful of shots is also cheap to test: at 768p it is $5.63 a variant.

Remember what stays fixed. Motion, camera, cuts and the soundtrack come from the source, per fal's description. Every variant therefore carries the original voice and script. If your plan depends on each creator sounding like themselves, this route will not deliver it, and you should say so to the people involved before they agree. The video changes who is on screen, not what they say.

What should I get from each person first?

Sume's docs do not set a likeness policy for the Recast route, and its Terms of Service put the burden on you: you represent that you have all rights and permissions for content you submit, including permission to use any person's likeness or voice. That makes a written release the safest practice. Ask for the following, in writing:

• Consent to use their photo as the reference for AI-generated video, naming the tool class (AI recast or swap), not only general use.
• The scope: which ad, which platforms, how long, paid or organic.
• Whether the footage may show them saying or doing things they did not say or do, since the soundtrack is the original one.
• Compensation and a way to withdraw, which matter under laws that treat digital replicas as a likeness right. Sume's SB 1111 checklist walks through one such law.
• Whether the original performer in the source agreed to be replaced and republished.

How do I label the finished ads?

Treat every variant as altered video of a real person. Platforms and regulators ask for disclosure of exactly that, and the details differ: TikTok's face swap rule and the EU AI Act's labeling duties are covered in earlier posts. Sume's docs do not describe a label added to Recast output, so the label and any platform setting are yours to apply. Keep the release, the source video, the photo, the prompt and the Sume job id together for each variant. That file is what you will want if a platform or a person asks questions later.