Migrating from Novita after the August 2026 AUP change

Source. Novita AI Acceptable Use Policy, "Effective Date: August 5, 2026", section 5 Content Standards, item (7): "Is sexually explicit." (novita.ai/legal/acceptable-use-policy, verified 2026-08-19.) Whether or not enforcement has reached your account yet, sexually explicit generation on Novita is a terms violation as of that date.

Model equivalence

Full list with every mode and price: Models · Pricing · API reference.

What changes in your integration

1. Auth — one API key: Authorization: Bearer sk-…. Send an Idempotency-Key header on every paid request so retries never double-charge.
2. Submit → poll — POST /ent/v2/generate returns HTTP 202 with a task_id; poll GET /ent/v2/tasks/{task_id} until status is success or failed (states: pending → processing → success | failed). Back off from 2 s to 5 s between polls. Read outputs from result_urls.
3. Request shape — a flat JSON body: model, prompt (up to 5,000 characters), optional negative_prompt (up to 500 on Wan video), media as top-level arrays images / videos / audios / files of public HTTPS URLs, and model parameters under params — resolution (480p/720p/1080p by model), ratio (16:9, 9:16, 1:1, 4:3, 3:4; HappyHorse adds 4:5, 5:4, 21:9, 9:21), duration (Wan 2–15 s, HappyHorse 3–15 s), seed, watermark, prompt_extend. Map your Novita duration / size fields onto these; unknown fields return 400 rather than being silently dropped.
4. Media inputs — public HTTP(S) URLs only (no base64, no short-lived signed URLs). Upload to your own storage or our upload endpoint first.
5. Billing — prepaid USD wallet, priced per second of video or per image. Cost is reserved at submit and settled on completion; failed tasks are refunded in full. Top up with crypto/stablecoins (USDT, USDC and others) — no card network sits between you and your generation capacity.

POST /ent/v2/generate
{
  "model": "wan2.7-t2v-uncensored",
  "prompt": "…",
  "params": { "resolution": "720p", "ratio": "16:9", "duration": 5 }
}

Five-step checklist

1. Create an account and claim $5 test credit (via contact or the community — roughly 100 seconds of HappyHorse 480p, enough to validate quality on your real prompts).
2. Re-point your generation calls per the table above; run your prompt regression set.
3. Compare outputs and unit costs side by side (pricing page lists every SKU).
4. Move production traffic; keep Novita as a fallback for SFW workloads if you like.
5. Read the Content Policy — minors and real-person likenesses are banned absolutely, and we screen for both. If your app respects those lines, nothing else changes.

FAQ

Is this legal advice about my Novita account?

No. Read their AUP yourself — we quote it with the effective date and section so you can. Their terms can change again; ours are published at /content-policy and material changes are announced 7 days ahead.

Do you support LoRA or custom checkpoints?

Not today. Multi-subject reference-to-video (wan2.7-r2v-multiple-uncensored) covers many character-consistency needs: pass reference images, a reference video and a voice sample and the model keeps the subject consistent across shots.

Does 18models really allow NSFW output?

Yes, in writing: fictional adult content is permitted; content involving minors or real people is prohibited absolutely and screened on both input and output. See the Content Policy.

What if 18models' policy changes too?

Policy changes are announced 7 days ahead via the changelog. The whole point of writing the line down is that you can plan against it.