Generate video
generate_videoRender a raw video clip from your own prompt and return its served MP4 URL, or edit an existing clip via a reference video. Use for experimental clips or precise manual control instead of finished ads.
Instructions
Render a RAW video clip from your own prompt and return its served mp4 URL. For finished brand ADS prefer render_ad (it runs the Studio quality pipeline — composited text, clean speech, end card, music); use this for raw/experimental clips or precise manual control. ONE generation = one continuous clip up to the model’s longest listed duration — the longest-clip model in the catalog today renders a full multi-beat spot of up to 30 SECONDS in ONE unbroken take with native synchronized audio, so never assume a generic 8–10s cap and never stitch something that fits one clip; durationSeconds must be one of the model’s durations from hermoso_capabilities, which is the live list. TO GET A SPECIFIC MODEL, NAME IT in model: an unnamed render is routed by the server’s own auto-pool, which is narrower than the catalog, so the longest-clip and highest-resolution models are reached by naming them and not by omitting the field. Renders take 1–3 min. refImage anchors the opening frame; ttsScript adds a voiceover. AUDIO IS NOT FREE AND NOT OPTIONAL BY DEFAULT: a clip delivered with no audio of its own gets a music bed composed and CHARGED on top of the render (see musicMood and audio) — on a cheap short draft the bed can cost as much as the clip. Pass refVideo (a clip URL) to EDIT an existing video instead of generating from scratch — the omni engine transforms that clip per your prompt, inheriting the source clip’s canvas + length (aspectRatio/durationSeconds are ignored for an edit). RAW MODEL ACCESS: your prompt is NOT dispatched verbatim by default — a few small guards are appended (packaging/label safety when no reference image rides, a negative prompt on the models that take one, reference-binding lines when references ride) and hex colour codes are rewritten to colour names. Pass raw:true for none of that. raw:true dispatches your prompt to the model BYTE-IDENTICAL — no rewriting, no appended guidance, no negative prompt, no brand references attached on your behalf. Credits, the durable delivery of the finished asset and the per-model validation are unchanged. Spends credits (Starter plan is video-blocked server-side).
Input Schema
| Name | Required | Description | Default |
|---|---|---|---|
| raw | No | RAW MODEL ACCESS: dispatch this prompt to the model BYTE-IDENTICAL — no appended packaging/label guidance, no negative prompt, no reference-binding lines, no hex-to-colour-name rewrite. Use it when you want the model itself rather than Hermoso's render craft. Two vendor-required fixes still apply: extra @ImageN tokens are dropped and an over-long prompt is trimmed at a sentence. Billing, durable delivery and per-model validation are unchanged. | |
| loop | No | true = a seamless loop whose last frame flows back into its first. Only models with loop true in hermoso_capabilities; needs refImage and cannot be combined with endImage. | |
| audio | No | default true. false = render SILENT: no native model audio, no music bed, and no bed charge held or billed. This is the ONLY way to decline the automatic bed (see musicMood) — leave it alone for anything that should have sound, and do not combine it with ttsScript. | |
| model | No | video model id from hermoso_capabilities; a named model is never swapped without asking. Omit to let the router pick | |
| shots | No | MULTI-SHOT: one clip cut into these shots, in order. The seconds must add up to a length the model renders; replaces `prompt` (send either). Only models with multiShot true in hermoso_capabilities. | |
| extend | No | true = EXTEND refVideo: the same clip continues per your prompt for durationSeconds more (each model’s extend.minSeconds..maxSeconds in hermoso_capabilities), delivered as ONE clip, source then continuation. Needs model named (a model with extend in hermoso_capabilities). | |
| prompt | Yes | the video prompt / shot description (for a refVideo edit, this is the transformation instruction) | |
| endImage | No | local path or URL of the LAST frame: the clip travels from refImage (required with it) to this image. Only models with endFrame true in hermoso_capabilities take it; any other named model is refused by name, nothing charged. | |
| refImage | No | local path or URL to anchor the first frame | |
| refVideo | No | URL of an existing video to EDIT rather than generate from scratch — the clip is transformed per your prompt, inheriting the SOURCE clip’s canvas (aspect ratio) and, on every model hermoso_capabilities marks `sourceLength`, its LENGTH too: those endpoints have no duration parameter, their listed `durations` are the per-second price ladder, and a durationSeconds you send is reported back as unused rather than silently dropped. Trim the source to change the length. Omit `model` for the default editor, or name a model whose videoEdit is true in hermoso_capabilities (a named model that cannot edit is refused, nothing charged); refImage rides along as the look of what the edit adds. With extend:true this is instead the clip to EXTEND. Omit to generate a fresh clip. | |
| ttsVoice | No | voice name, e.g. Rachel / George | |
| musicMood | No | WHICH mood the music bed is composed in (upbeat / calm / warm / epic / tense / playful / elegant / hype / chill / dramatic). It does NOT decide WHETHER there is one: a clip that comes back with no audio track — every model hermoso_capabilities lists as "silent", plus any audio model that returned mute — gets a bed composed and CHARGED automatically, at the flat per-track fee hermoso_capabilities reports as explainerMusicCredits, and omitting this field only means the mood defaults to "warm". Pass audio:false for a genuinely silent clip with no bed and no bed charge. | |
| refImages | No | SEVERAL reference images (local paths or URLs) — a person, products, a place — that must all appear in the clip. Only models whose `refs.max` in hermoso_capabilities is above 1 use more than one, and each uses at most that many; with `refs.promptAddressed` true, name them in your prompt as Image 1, Image 2… in this order. minimax-h3-max-ref takes up to 9 and keeps each one as a reference rather than a first frame. On a model that takes one image, only the first is used. | |
| ttsScript | No | voiceover script to speak | |
| cameraMove | No | A named camera move around the still in refImage, spelled exactly as the enum gives it: an orbit (a quarter turn, the default), orbiting left, a half turn or a full turntable, a rise, a crane up, a push in, a pull back, or a reveal. Only the camera-controls model renders one (minimax-h3-max-camera, listed with its moves in hermoso_capabilities): pass it with model omitted and that model is picked, or with that model named; any other named model is refused by name with nothing charged. Needs refImage. | |
| resolution | No | '1080p' default (what we ship and bill for); '480p'/'720p' = cheaper draft passes, '4k' = premium final delivery (more credits). NOT EVERY MODEL OFFERS EVERY TIER — this enum is what the tool accepts, and each model's OWN `resolutions` list in hermoso_capabilities is what it can actually render. Ask for a tier the chosen model does not list and it is rendered at that model's best available tier instead, with nothing in the reply saying so — so check `resolutions` before promising anyone 1080p or 4k. | |
| aspectRatio | No | default '9:16' | |
| interactionId | No | with extend:true on an Omni model: the interactionId returned by an earlier render on that model — continues it from its own stored context instead of re-uploading refVideo. | |
| durationSeconds | No | length of THIS ONE clip in seconds — pick one of the CHOSEN model’s listed durations from hermoso_capabilities (never a generic guess: the lists differ per model, from 4–8s on the short models up to 30s on the longest-clip one). This is a single continuous generation, so it CANNOT exceed that model’s longest clip: a longer ask is REFUSED with nothing rendered and nothing charged (it is never quietly truncated). Past ~15s only the long-clip models qualify, and an unnamed render is routed by the narrower auto-pool — so NAME the model in `model` when you are asking for a long single take. For a spot longer than any one clip, use plan_ad with durationSeconds then render_ad, which stitches acts of at most one model clip each (on a 15s-clip model, 40s = 15+15+10). | |
| cameraTrajectory | No | Your own ordered camera path, 2 to 12 keyframes, for the camera-controls model only (same rule as cameraMove; overrides it). The first pose is held until its time and the last pose is held to the end. A value outside these bounds is refused by name, nothing charged. |