Skip to main content
FastH3-Omni generates video and sound together. Start by naming who or what appears, what happens during the clip, how the camera sees it, and what the viewer hears. You can write this as sentences or separate labeled lines. The labels in the examples below are prompt text, not required API fields.

​
Give each clip one clear beat

For a short clip, choose an action that can play out in the time available. Describe the setting and camera, then spell out dialogue, ambient sound, and music, including when you want no music.
Set seconds on enqueue to 5.875, 8, 10.125, 12.25, or 14.375. The model refuses other values. Give the action enough time to fit the length you choose.

​
Name image references

Upload one reference_image or up to three images in reference_images for real-time generation. The model accepts up to nine images when real-time speed is not required. Give each image a role in the prompt: a person, object, setting, or style. Use <Picture 1>, <Picture 2>, and so on in upload order, and name the details that matter for the clip.
The <Subject 1> label helps tie later actions to the person in <Picture 1>. These angle-bracket labels are prompt text, not separate command parameters. A reference image guides appearance; by itself, it does not pin the clip’s first or last frame. The model may fit an image to a supported shape with black bars. Use set_canvas to choose the output aspect ratio.

​
Pin the first or last frame

Use first_frame_image to set the opening frame and last_frame_image to set the closing frame. These are separate keyframe inputs, not <Picture N> reference images. Include at least one ordinary reference_image or reference_images entry in the same enqueue request. Use the prompt to describe what happens between the keyframes. For a continuation, continue_mode: "frame" takes the previous clip’s final frame as the new opening frame. You may add last_frame_image to choose its ending, but you cannot add first_frame_image because the continuation already supplies it. The default latent continuation carries motion, camera, and audio instead and accepts no keyframes. See the schema for enqueue examples.

​
Give audio references a role

Upload one reference_audio or up to three clips in reference_audios. Each clip must be 2 to 15 seconds long, in mono or stereo. A request with audio also needs an image reference or a retained continue_from_clip_id. A continued clip’s soundtrack takes one of the three audio slots, so you can upload at most two more. Refer to uploads as <Audio 1>, <Audio 2>, and so on. Give each uploaded clip its own sentence saying whether it supplies a voice, background sound, or music, and connect a voice to the person who should use it. Without a stated role, the model may ignore the audio reference. The Sandbox offers Voice, Sound, and Soundtrack roles and adds the selected role to the prompt:
For generated speech or singing, put the intended words in the prompt rather than assuming the voice reference supplies them. Treat references as guidance rather than a request for sample-exact audio. Listen to the result before relying on an exact sound. The Sandbox requires a picture before it lets you add audio. The raw model API also permits audio with a retained continue_from_clip_id and no new picture.

​
Continue a clip

Set continue_from_clip_id to an already generated clip from the same session and canvas. Give the next clip a new action, then name what should carry over: the person’s appearance, lighting, camera motion, and sound.
With the default continuation, the previous clip’s ending guides the next clip’s motion, camera, and audio; it is not a frame-accurate edit. The model refuses an ID from another canvas or one it no longer retains. It keeps the eight most recent generated clips as continuation sources, so make a dependent clip before its source expires.

​
Keep prompts within the budget

The model checks the prompt and image references against the deployment’s text and presentation budgets. If a request exceeds either budget, enqueue returns command_error. The model does not trim the prompt for you; shorten it or remove references before retrying.

​
See also