H3 Reference Turbo Realtime prompt guide

Refer to your images

Send up to nine JPEG, PNG, or WebP upload references with each clip. For one image, use reference_image; for an ordered list, use reference_images. Supply one field, not both. Images are optional: omit both fields, and the clip comes from your prompt alone. The list order determines the names used in your prompt: Picture 1 is the first image, Picture 2 is the second, and so on. Describe the role of each picture. For example, with a subject image followed by a setting image:

The person in Picture 1 walks through the garden in Picture 2. The camera follows from the side.
Leaves rustle and birds call in the background.

References guide subjects and appearance. They do not define a fixed first or last frame. A reference does not have to be a character. A picture can carry a person, an object, or a place, and a set that mixes them works well: one image per character, plus one for the setting they act in. Order the list the way your prompt reads, then name every picture you send.

Refer to your audio

A clip can also carry up to three audio references: one reference_audio, or an ordered reference_audios list. Supply one field, not both. Audio is optional, and a clip that has audio needs at least one image, or a continue_from_clip_id whose clip the session still holds. Each audio clip must be 2–15 seconds long, mono or stereo, in WAV, MP3, AAC/M4A, OGG/Opus, FLAC, or WebM. The list order names the clips the way it names pictures: Audio 1 is the first, Audio 2 the second. A label does not say what the audio is for, so define each one in subject_definitions. Without that sentence the model conditions the soundtrack on a waveform it was never told the role of.

To do this	Write this in `subject_definitions`
Give a character a voice	`<Audio 1> is the voice-timbre reference for <Subject 1> (S1).`
Carry a soundtrack over	`<Audio 2> is the synchronized audio track of <Video 1> and is reused in this clip.`

<Subject N> is who the viewer sees. (Sx) is who speaks, numbered in the order speech happens in detailed_description, not in the order of the reference list. Use both for a subject who speaks: <Subject 2> (S1). A voice reference whose (Sx) never appears on a spoken line binds to nobody. Reused music needs no (Sx), because a voice inside a track is not a speaker.

Describe the clip and its sound

Use prompt to describe what the clip should show and sound like. It must contain nonempty text. The model stores the prompt unchanged and never truncates it. There is no character limit, but there is a budget. The prompt is measured in text tokens, and the default budget is about 2,000 — roughly 8,000 characters of English prose. A prompt past the budget is accepted by enqueue and then fails the build, so it arrives as clip_failed rather than as a refusal. The structure below fits comfortably; a prompt long enough to hit the budget is usually one that is repeating itself. Each enqueue keeps its own prompt and ordered reference list, including while other clips build or play. Supply the references for each new clip; new uploads do not change earlier requests.

Structure a detailed prompt

Short prompts work. For a clip with several subjects, a specific camera, and its own sound, a labelled structure gives the model far more to hold onto. The following six sections are a convention rather than required syntax — the model reads the prompt as plain text — but they cover the decisions the model has to make, in the order it helps to make them.

Section	What it carries
`subject_definitions`	One line per subject, binding it to a picture and listing what to preserve.
`summary`	One sentence: the clip length, and what happens from start to end.
`retention_analysis`	Per subject, when it appears and which traits must stay fixed.
`detailed_description`	The look of the clip, what to avoid, then the action broken into beats.
`overall_soundscape`	What the viewer hears in the scene, including dialogue or its absence.
`non_diegetic_music`	The score, if any, and how it tracks the action.

Use whichever sections earn their place. A single subject on a plain background needs a sentence, not six headings. Reach for the full structure when you are holding several identities steady at once.

Name and preserve your subjects

Give each subject a stable label, bind it to its picture, and then say which traits must not drift. Repeat those labels in the action so the model knows who is doing what.

subject_definitions:
<Subject 1> is the baker in <Picture 1>, preserving her face, grey braided hair, flour-dusted navy
apron, and rolled shirtsleeves.
<Subject 2> is the morning market lane in <Picture 2>, preserving its stone paving, striped awnings,
crates of fruit, and low sun from the left.

Two habits pay off here:

Count the things you do not want duplicated. Writing “a single canvas tote” or “one silver blade” is more reliable than naming the object alone, because it tells the model how many should appear.
List traits, not adjectives. Face, hair, garment, and the props a subject carries are what hold an identity together across a clip.

Break the action into beats

Say the look of the clip once, state what you do not want, then walk the action in order. Labelled beats keep a five-second clip from trying to do everything at once.

detailed_description:
Photorealistic morning market cinema, 16:9, warm low side light, physically coherent motion, stable
identities. One continuous shot with no cuts, no duplicate people, no costume changes, and no
readable text.

[Shot 1, 00:00.000-00:02.000] A waist-height camera frames <Subject 1> at her stall in <Subject 2>,
turning a loaf in both hands as steam rises from the crate beside her.

[Shot 1, 00:02.000-00:05.166] She lifts the loaf toward the camera and smiles. The camera pushes in
slowly to a medium shot while the awning shifts in the breeze behind her.

Timecodes are optional, and they help most on short clips where every beat is tight. When you use them, let the last beat end at the clip’s real duration, which the clip_queued reply reports. Longer clips read fine with plain [Shot 1] beats in order. Stating what to avoid is as useful as stating what to show. Phrases like “no cuts”, “no duplicate people”, “no costume changes”, and “no readable text” reliably remove the artefacts they name.

Describe the sound

Video and audio are generated together, so the sound belongs in the same prompt. List what the scene itself makes, and say plainly when there is none.

overall_soundscape:
Quiet market chatter, a crate set down on stone, paper rustling, and one distant bicycle bell. No
dialogue.

non_diegetic_music:
A light acoustic figure that fades in under the push-in and holds.

For a spoken line, name the speaker, describe the voice, and wrap the words with the language tagged:

Looking at the camera, <Subject 1> (S1) says in a warm, unhurried voice, <d>[English] Still hot from
the oven.</d>

Choose reference images that hold up

The reference set does most of the work on identity, so it is worth assembling deliberately.

One subject per picture. A clean, well-lit image of a single character or a single place is easier to hold than a group shot the model has to pull apart.
Show what you want preserved. If a costume detail or a prop matters, it should be visible in the picture you bind it to.
Keep the set coherent. References that share a palette, time of day, and light direction produce a clip that looks like one scene rather than a collage.
Any aspect is fine. Each image must be a single-frame JPEG, PNG, or WebP, at most 25 MiB and 25 million pixels, with an aspect ratio between 1:4 and 4:1. Images may differ in size from each other and from the output canvas, and they are never cropped into a fixed first frame.

Continue a scene across clips

To continue a scene, pass a generated clip’s ID as continue_from_clip_id in the next enqueue. Motion, camera, and audio carry forward while the new clip’s references guide its appearance. The source must have finished generating. A blank, unknown, or dropped ID produces an independent clip. Write a continued clip as the next stretch of the same shot. Three things keep the seam quiet:

Repeat the subject definitions. Each clip is prompted on its own, so the labels and the traits to preserve have to appear again.
Restate the look. Naming the same lighting, weather, and palette as the previous clip stops the scene from drifting, and opening the action with “the shot continues from…” tells the model where it is picking up.
Update a subject that changed. If a clip ends with a mask pushed up or a coat removed, the next clip’s definition should describe the subject in that new state, even though the reference image still shows the old one. The references guide appearance; your prompt says which moment of it you are in.

A continued clip uses one of its three audio references for the previous clip’s soundtrack, so it accepts at most two of your own. That soundtrack already conditions the new clip, and overall_soundscape steers it. When you add a voice reference as well, say what carries over and what the voice applies to. Setting set_flush_on_clip_end.enabled to false holds the last frame and avoids flushing between ready autoplay clips. This setting controls playback boundaries; it does not request continuation. Use continue_from_clip_id to continue from a generated clip.

Try several seeds

The same prompt and references produce a different take under a different seed. When a scene matters, queue it a few times with explicit seed values and keep the take you like. An explicit seed leaves the session default untouched, so a sweep does not disturb the clips you queue afterwards. See the schema for parameters and the tutorial for the command sequence.

Overview

FastH3-Omni

Vidu S2-Avatar

Vidu S2-Editing