| kind | Yes | what to make: image | video | speech | music. Each argument below says which kinds take it | |
| lines | No | speech: the script, in order — each line is {speaker, text, style} | |
| model | Yes | a model id from list_models for this kind | |
| style | No | music: the style — a genre, a mood, the instruments (custom songs) | |
| title | No | music: the song's title (custom songs) | |
| images | No | image, video: hosted image URLs — references for an image model; for a video model the first frame, or with images_are_references the subject/style references (at most the model's max_images). Import a link with import_media first | |
| lyrics | No | music: the words to be sung. With lyrics, title and style are required | |
| prompt | No | image, video: what to make. music: the description of the song (simple mode; for a song with its own lyrics use lyrics) | |
| funding | No | who pays: personal (the person's own coins) or team (their team's pool — only when the person allowed this connection to spend team coins and the team approved this app). Omit to use the payer the person chose when connecting | |
| duration | No | video: seconds, one of the model's durations. music: the song's length in seconds (custom songs only) | |
| speakers | No | speech: the voices, in order (list_models has the voices, accents, styles and paces). Each is {name, voice, accent, style, pace, profile}, all optional except that a script needs at least one speaker. A model with line_style lists no accents, styles or paces: direct each line with its own style instead | |
| last_frame | No | video: a hosted image URL for the LAST frame — only for models with supports_transition; then images holds exactly the first frame | |
| resolution | No | image, video: one of the model's resolutions — a higher tier costs more | |
| video_task | No | video: what reference_video is for — edit (change that one clip; images are references, and with no duration the answer runs the clip's length) or extend (add new footage to it; duration is the new part's length, the answer is that part alone, and images are refused). Omitted, it is a reference; on a model with clip_is_source it is required, and images are refused either way (there a clip is never a reference, an edit runs the clip's length, and an extension answers the clip plus the new part and takes up to 10 seconds). Only for models with video_tasks; needs a prompt and exactly one clip | |
| aspect_ratio | No | image, video: the frame, one of the model's aspect_ratios (e.g. 16:9) | |
| filler_words | No | speech: add filler words and natural interjections to a TWO-speaker dialogue — only for models with filler_words | |
| instrumental | No | music: no vocals | |
| expected_cost | Yes | the coins estimate_generation answered for these very arguments — omit it only when the host confirms prices with the person itself | |
| generate_audio | No | video (and any image model that has it): make the sound too — only for models with supports_audio, and it costs audio_cost more | |
| idempotency_key | No | optional: repeat it on a retry so the retry answers the first generation instead of charging again | |
| reference_audio | No | video: a hosted audio URL to take sound from — only for models with max_reference_audios above 0 | |
| reference_video | No | video: a hosted video URL to take motion or style from, or on a model with clip_is_source the clip to edit or extend (video_task is then required) — only for models with max_reference_videos above 0 | |
| images_are_references | No | video: read images as references to draw from, NOT as the first frame — the video does not start on them. Needs at least one image, cannot be combined with last_frame, and only models with supports_reference take it | |
