Vidu is_rec and Wan prompt_extend: prompt rewrite vs Sume

Both vendors can change your prompt for you. Vidu's is_rec uses a recommended prompt when true, and Alibaba lists prompt_extend with a default of true. Sume's /v1/videos request table has no field for either, so your prompt is the prompt. Remove the two keys when you port and write the extra detail into the text yourself.

What do the two switches do?

Both default to on, so a port silently turns them off.

Why does that change your results?

A short prompt that looked good on Vidu may have been expanded for you. The same short prompt sent to wan-3.0 on Sume goes through as written, so the clip can look plainer. Compare like for like: write a longer prompt with subject, camera move, light and sound, and run it once on each side.

How do you strip the vendor keys in a port?

The function also drops seed, because Sume's models report seed: false and reject the field (see the Sume differences table).

VENDOR_ONLY = ("is_rec", "prompt_extend", "payload", "seed", "watermark")

def to_sume(body: dict) -> tuple[dict, list]:
    clean = {k: v for k, v in body.items() if k not in VENDOR_ONLY}
    dropped = [k for k in body if k in VENDOR_ONLY]
    return clean, dropped

body = {"model": "wan-3.0", "prompt": "A barista pours latte art",
        "prompt_extend": True, "seed": 0}
print(to_sume(body))

What should you test first?

Run a small, cheap test before any batch.

• One prompt, short and long, on wan-3.0 at 480p so the test costs 5 x $0.0625 = $0.3125 per clip.
• Whether the long prompt recovers what the vendor's expansion added.
• Keep the prompt text next to the job id so you can repeat the winner.