HappyHorse 1.1 takes a seed on Cloudflare; Sume video models do not

Cloudflare's HappyHorse 1.1 T2V page accepts a seed from 0 to 2,147,483,647 and an optional watermark. Sume's video catalog marks seed: false for every v1 model, so you cannot pin a take there. To repeat a look, fix the first frame instead.

What each side exposes

Why this matters for retakes

A seed lets you rerun a prompt and get the same or a closely related take, which makes it cheap to change one word and compare. Without it, each submission is a fresh sample, and a retake costs a full generation. The Sume docs describe seed as a request field but say the result is not deterministic on all providers; the catalog descriptor says no v1 model accepts one.

A workaround on Sume

• Generate a still you like, then send it as first_frame in frame_images so every retake starts from the same image.
• Draft at 480p on Wan 3.0, Seedance 2.5 or MiniMax H3 and only render the final at 1080p once the motion works.
• Keep the prompt identical between takes and change only one input at a time.

What a retake costs on Sume

Because a retake is a new generation, budget for it. At Wan 3.0's $0.125 a second at 720p, a 10-second retake is $1.25; at Kling 3 with audio it is $2.10. Three retakes of a shot cost three times the base price, so a short 480p test of the motion is usually cheaper than rerunning the final.

Cloudflare's seed field helps with comparison, not with quality. Two seeds can both fail the same prompt. What the seed buys is the ability to change one word and see whether the change, and not the roll of the dice, caused the difference.