Media generation

Async media generation: images, video, voiceover (TTS), speech-to-text, music and 3D. You create a task and poll the result. The full model list is in the media-models catalog; webhooks, errors and billing are in the webhook section.

Creating returns 202 with { ok: true, jobId, status: "processing" }. Poll GET /api/media/jobs/{jobId} until status: "success", then take the resultUrls.

# 1. create a generation job
curl https://getvibeapi.com/api/media/generate \
  -H "Authorization: Bearer $GETVIBE_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "veo-3.1",
    "prompt": "a cat on a skateboard, slow motion"
  }'
# → {"ok":true,"jobId":"...","status":"processing"}

Besides model and prompt, you can pass extra options in opts. The set depends on the model. The full machine-readable contract is via GET /api/model-capabilities (or GET /api/model-capabilities/{model} for a specific model).

URL options (image_url, input_urls, mask_url and the rest) are links the provider fetches, so they accept only public https:// addresses. An inline data:image/...;base64,… string in one of those fields is rejected with 400.

To send your own file, use the matching *_base64 field instead: the service stores the bytes, publishes them and passes the resulting link to the provider. Up to 16 files per request in input_base64.

POST /api/media/generate and POST /api/media/price accept the same body, so you can quote it and then submit it unchanged. There are no other top-level fields: an unknown one is refused on both endpoints with 400 and code unknown_field naming it (field, fields), with or without strict. Generation parameters go inside opts.

Any source file can be sent three ways: as a link (image_url, audio_url, video_url…), as inline bytes (image_base64, audio_base64, video_base64…) or as the id of your own file from POST /v1/files. The id fields are listed in every operation’s requestJsonSchema and are accepted in strict mode.

For models whose output length follows an uploaded track — kling-ai-avatar, omnihuman-v2, omnihuman-2-0, omnihuman-v1.5, omnihuman-1-5, volcengine-lipsync (audio), kling-motion-control, wan-animate, topaz-video-upscale, p-video-edit (video), lipsync-2-pro, sync-3, react-1 (both inputs) — /api/media/price and the max_price_rub pre-flight on /api/media/generate measure that file server-side (link, data: URI, base64 or *_file_id) and price the measured seconds. The answer carries sourceSeconds (outputSeconds for lip-sync), and estimateReason says the length was measured. If the file cannot be read, the price keeps its previous basis (one second, the catalog estimate or the route maximum) and estimateReason says so. The charge is the per-second rate × the seconds actually rendered, never above max_price_rub. opts.duration on these models is accepted and ignored with a param_ignored warning.

Create: { name, consent: true, engine, audio_url | audio_base64 } or several recordings samples: [{ audio_url } | { audio_base64 }, …] (engine: "elevenlabs" only). An ElevenLabs voice is used by voice-clone-tts and by the Voice Changer (elevenlabs-voice-changer). Appending recordings to an ElevenLabs voice is free. Every recording is downloaded and measured before the supplier is called: a refusal changes and charges nothing.

Voice Changer and Forced Alignment

When the supplier’s start confirmation does not arrive, the job stays in status: "submitting" with stage: "submission_unknown", errorCode: "media_submission_uncertain" and failure.advice: "reconcile_first" — keep polling the same jobId and do not create a duplicate. To test this path at no cost, send POST /api/media/generate with a sandbox key and the header x-pv-sandbox-outcome: submission_uncertain. The answer is a normal 202 with a jobId; polling it answers as an unknown submit for 60 seconds, then as a job that never started (failure.code: "submission_not_started", charged: false). Nothing is stored, sent to a supplier or charged. A regular key with this header gets 400 sandbox_outcome_requires_sandbox_key.

Same endpoint as the rest of generation: put the STT model name in model, leave prompt as an empty string, and pass the recording in opts.audio_url. The transcript arrives in resultText on GET /api/media/jobs/{id}; resultUrls stays empty — transcription produces no file.

# speech-to-text: create the job, then read the transcript
curl https://getvibeapi.com/api/media/generate \
  -H "Authorization: Bearer $GETVIBE_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "whisper-large-v3-turbo",
    "prompt": "",
    "opts": {
      "audio_url": "https://example.com/meeting.mp3",
      "language_code": "en"
    }
  }'
# → {"ok":true,"jobId":"JOB_ID","status":"processing","priceRub":0.03}

curl https://getvibeapi.com/api/media/jobs/JOB_ID \
  -H "Authorization: Bearer $GETVIBE_API_KEY"
# → {"status":"success","resultText":"the transcribed text…","resultUrls":[],"priceRub":0.03}

Models

• whisper-large-v3-turbo — $0.00025 per minute of audio. Cheapest and fastest, 99 languages.
• nemotron-asr-multilingual — $0.00025 per minute. NVIDIA multilingual ASR, punctuates on its own.
• whisper-large-v3 — $0.000563 per minute. Full Whisper: slower than turbo, more accurate on noisy audio.
• universal-2 — $0.003125 per minute.
• speech-to-text — $0.01 per minute (ElevenLabs Scribe). Supports speaker diarization and audio-event tags.

Parameters

Formats and limits

mp3, wav, m4a/mp4, flac, ogg/opus and webm are accepted. File size is 25 MB maximum — roughly 26 minutes of 128 kbps mp3, or several hours of speech in opus. Anything larger is refused immediately with an explicit message and is not billed; split the recording instead.

To avoid polling, register one address mediaWebhookUrl in your dashboard. When the task completes (success or error), the service makes up to five delivery attempts of a POST with a JSON body:

// POST to your mediaWebhookUrl when a job finishes
{
  "jobId": "...",
  "status": "success",            // or "fail"
  "model": "veo-3.1",
  "kind": "video",                // image | video | audio | music | 3d
  "resultUrls": ["https://getvibeapi.com/api/media/file/..."],
  "failMsg": null,
  "createdAt": "2026-06-17T10:00:00.000Z"
}

The generation result URL (image, video, audio or 3D model) from resultUrls. Use it as returned by the service: it opens without an API key and works in <img>, <video> or a browser. The provider is hidden: the upstream address is resolved server-side, the client never sees it.

# The signed URL is returned in resultUrls of the /api/media/jobs/{id} response
# or in the webhook body. Opens directly in the browser / <img> — no key needed.
https://getvibeapi.com/api/media/file/JOB_ID/0?sig=HMAC_SIGNATURE