Starts a generation process and charges team credits. V1 types: image-generation, video-generation, audio-generation, image-upscale, video-upscale, video-lipsync. The model must match the type (see GET /models). For seedance-2.5-edit, the first referenceVideos item is the required base video; editOperation defaults to edit, while extend, sequel, and prequel allow an omitted prompt. video-lipsync takes sourceVideoAssetId and audioAssetId from the team asset library; the server-measured audio duration determines its quote and charge. Sync-Max is train-once: omit trainingId on the first run, then reuse the trainingId returned on the process for faster later runs against the same trained source. Sync-Pro and Sync-v3 are single-shot and require no training. Always send an Idempotency-Key: retries with the same key and body return the original process instead of double-charging; the same key with a different body returns idempotency_conflict. input is validated per process type AND against the model's published capabilities (aspectRatios, resolutions, videoLengths from GET /models) — unsupported values are rejected with invalid_request, never silently coerced. Internal fields (batchId, filteringId, agent context) are ignored. metadata (flat string/number/boolean/null values, max 16 keys) is echoed back on reads. webhookSecret is optional for legacy unsigned webhook delivery; when supplied with webhookUrl, notifications are HMAC-signed and may be attempted twice with a stable delivery id.