W3.0 Omni and Qwen Image 3.0 Pro — video, text-to-image and instruction editing — as a REST API. Your prompt is sent as written. Nothing is expanded, softened or replaced before it reaches the model. Director's Mode compiles a multi-character scene from one JSON payload.
Most endpoints that serve these models sit behind a rewriter that changes your prompt, or a filter that rejects jobs the model itself would have run. We send the request through as you wrote it.
Both are production models we run daily on our own platform — not a wrapper over someone else's reseller.
Omni-modal video generation. A single request can carry images, video, audio and text together.
Text-to-image and instruction-driven editing, synchronous — the edited image comes back in the same call.
extra_urls when a request yields more than one image| W3.0 mode | What you send | Use it for |
|---|---|---|
| Text to video | prompt only | Pure generation, no inputs. |
| Reference (omni) | ref_image_urls, ref_video_urls, ref_audio_urls | Character consistency. References supply identity and performance; the model composes a new scene around them. |
| Frames | first_frame_url, optional last_frame_url | Animating a still, or locking both ends of a shot for an edit that has to cut cleanly. |
| Video extend | extend_video_url + ref_video_seconds | Continuing an existing clip. Input plus output must not exceed 30s total. |
| Document | file_url | PDF, DOCX, PPTX, XLSX, TXT, MD → video. Max 1 file, 50 pages, 100MB. |
| Link | link_url | A public webpage turned into a clip. Cannot be combined with file_url. |
| Director's Mode | characters[] + beats[] | Multi-character scenes with timed action and dialogue. See below. |
Reference mode and the frame / extend / file / link modes are mutually exclusive — pick one per request.
A lot of APIs run your text through a second model before generation. That pass is why a scene comes back softer than you wrote it, or why a wardrobe changes between shots. We do not run that pass.
Hard limits, checked at the gateway before a request is billed:
Blocked requests return 403, are never billed, and are written to an immutable audit log. Repeat attempts terminate the key permanently and without refund.
Prompting a video model for a multi-character scene is the hardest thing you'll do with one. Get the reference ordering wrong and two characters fuse into one. Put two speech beats inside ten seconds and the dialogue scrambles. Write "then they swap places" and a duplicate person walks into frame. Director's Mode is the layer we built to stop all of that — and it's callable as one JSON object.
Instead of one long paragraph, you describe your scene the way a shot list describes it: who is in it, what they look like, and what happens second by second. Our compiler turns that into the prompt the model actually needs.
Send characters and beats to /video — four lines of scene description instead of four hundred words of prompt engineering. Full field-level schema is in your account dashboard.
{
"model": "w3.0",
"resolution": "1080P",
"duration": 20,
"ratio": "16:9",
"characters": [
{
"id": "a",
"name": "Ciri",
"anchor": "silver-white hair, green eyes,
scar on left cheek",
"ref": "https://cdn.example.com/ciri.jpg",
"side": "left"
},
{ /* character "b" … */ }
],
"beats": [
{
"start": 0, "end": 5,
"action": "Ciri steps out of the treeline,
rain on her shoulders.",
"speaker": "Ciri",
"line": "You shouldn't have come."
}
/* … beats 2–4 omitted … */
],
"style": "cinematic",
"wardrobe": [ /* … */ ]
}
REST over HTTPS, JSON in and out, bearer-token auth. If you've integrated any generation API before, this will take you an afternoon.
image_urls, edit when you send 1–3. Synchronous: returns 200 with the finished result URL. Do not poll it.{"url":"…"} to pull it once from somewhere else. 18MB max. Use this instead of pointing us at a free image host.{"delete_all":true}. Scoped to your own folder.processing, completed with a result URL, or failed — and a failure refunds your balance on that same call.# 1 — submit curl -X POST https://tekno3d.com/wp-json/tekno3d/v1/api/v1/video \ -H "Authorization: Bearer tk_live_YOUR_KEY" \ -H "Content-Type: application/json" \ -d '{ "resolution": "720P", "duration": 8, "ratio": "16:9", "prompt": "A woman in a red coat walks through neon rain, handheld camera, shallow depth of field." }' → 202 {"job_id": 50412, "status": "processing", "charged": 0.72, "balance": 499.28} # 2 — poll every 10s until status leaves "processing" curl https://tekno3d.com/wp-json/tekno3d/v1/api/v1/job/50412 \ -H "Authorization: Bearer tk_live_YOUR_KEY" → 200 {"job_id": 50412, "status": "completed", "result_url": "https://cdn.tekno3d.com/…mp4"}
{"error":{"code":"…","message":"…"}}. Branch on error.code: bad_request (400), unauthorized (401), insufficient_balance (402), not_found (404), content_policy (422), internal_error (500), provider_error (502).curl -X POST https://tekno3d.com/wp-json/tekno3d/v1/api/v1/image/edit \ -H "Authorization: Bearer tk_live_YOUR_KEY" \ -H "Content-Type: application/json" \ -d '{ "prompt": "Change the wardrobe to a black leather jacket, keep the face and lighting identical.", "image_urls": ["https://cdn.example.com/subject.jpg"], "size": "2K" }' → 200 {"job_id": 50413, "status": "completed", "charged": 0.15, "result_url": "https://cdn.tekno3d.com/…png", "extra_urls": [], "balance": 499.13}
image_urls and the same call becomes text-to-image. Full reference — every field, every mode, and the complete Director's Mode schema — is issued with your API key.Video is quoted per 30-second clip and prorated by length — a 10-second 720P clip is $2.00, not $6.00. You top up a balance and spend it. Nothing recurring, no seat fees, no minimum commit. High-volume accounts are priced individually — if you're doing four figures of clips a day, talk to us before you sign up at list rate.
Prompts are screened at our gateway for the limits above. That check is for legality, not for rewriting your scene. A request that passes goes to the model as you wrote it.
ref_video_seconds or the request is rejected.Keys are issued manually — we want to know what you're building and what volume you expect so we can set your rates and limits properly. It usually takes under a day.
W3.0 Omni, Director's Mode and Qwen Image 3.0 Pro Edit, on a prepaid balance with no monthly commitment.
Generated footage is 8-bit SDR. VESAI masters it to true HDR10 with real BT.2020 colour and AI upscaling to 4K, 8K or 12K — desktop or cloud.