Introduction

Reference pictures, a prompt, and a few seconds later — a clip that looks like your subjects and sounds like the scene you described.

That is H3 Reference Turbo Realtime on Reactor: ordered reference images plus text, turned into video with synchronized audio, with a queue you can play, reorder, and keep feeding. Warm booth energy, not a one-shot black box.

What it is (and is not)

Display name: H3 Reference Turbo Realtime

Wire model name: reactor/h3-reference-to-video-turbo-realtime

create-reactor-app template: h3-reference-turbo-realtime

Typed package: @reactor-models/h3-reference-to-video-turbo-realtime

Each clip takes one to nine ordered JPEG, PNG, or WebP references. You describe scene action, camera, and what the viewer hears. References guide subjects and appearance — they are not a fixed first or last frame.

Not FastH3

Easy mix-up, different product. FastH3 (reactor/fast-h3) opens from text, a starting_frame, or an ending_frame — no multi-reference set. Commands can look alike (queue, play on demand), but neither is a drop-in for the other. If you need Picture 1…Picture 9 identity control, you want Reference Turbo Realtime.

Specs at a glance

Model      reactor/h3-reference-to-video-turbo-realtime
Input      1–9 refs (JPEG/PNG/WebP) + text prompt
Output     Video with synchronized audio
Duration   ~5–15 seconds per clip
Resolutions
  1344x768  (16:9)
  768x768   (1:1)
  768x1344  (9:16)
  1024x768  (4:3)

Billing note from the docs: usage is charged per second of session time, not per second of generated video. An open compose session still bills while you think — close or reset when you are done. (No dollar rates invented here; check Reactor Billing for current pricing.)

Quick start

Scaffold a working app with auth wired up:

npx create-reactor-app my-h3-app --model=h3-reference-turbo-realtime

Or install the typed package yourself:

npm install @reactor-models/h3-reference-to-video-turbo-realtime

Pass the full wire name reactor/h3-reference-to-video-turbo-realtime to the SDK. Exchange your API key for a JWT from your server (POST https://api.reactor.inc/tokens) — never put API keys in client code.

Session flow

  1. Connect — state moves disconnected → connecting → waiting → ready; first state_update / queue_update describe the session.
  2. Upload refs — SDK file upload, then attach returned references to each clip's prompt.
  3. Enqueue — clip gets an ID, enters the generation queue; clip_generated means it is ready to play.
  4. Play — play with a clip ID, or omit the ID for the front of the queue; enable autoplay so ready clips start when playback is idle.
  5. Steer the queue — move to reorder, pop to remove, stop to cut the playing clip (queues stay), reset to clear the session.

Queue several clips, each with its own prompt and reference set. Reorder freely. Play on demand or let idle autoplay do the handoff.

Prompting with pictures

Images are numbered in upload order. In the prompt, call them Picture 1, Picture 2, … and say what each one is.

A tiny starter (from Reactor's prompt guide):

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.

For multi-subject scenes, a labelled structure helps: bind each subject to a picture, preserve face/hair/costume traits, break action into beats, then describe diegetic sound and any score. Short prompts work too — reach for the full scaffold when identities must stay locked.

Pro Tip

One clean subject (or place) per picture beats a crowded group shot. Keep the set coherent on palette and light. References may differ in size from each other and from the output canvas; they are never cropped into a fixed opening frame.

To continue a scene, pass a finished clip's ID as continue_from_clip_id on the next enqueue — motion, camera, and audio carry forward while new refs guide appearance. Repeat subject definitions on every clip; each enqueue is prompted on its own.

Related context (optional)

In the wider MiniMax H3 / Ref2VA ecosystem you will also see turbo LoRAs and ComfyUI graphs. Useful for local experiments — and a different lane from Reactor's hosted realtime API. Same family vibes; not the same product surface.

Conclusion

H3 Reference Turbo Realtime is the Reactor lane for reference-guided clips with sound: up to nine ordered pictures, a prompt that covers action and audio, clip durations around 5–15s, four practical resolutions, and a session you can queue and play like a tiny live booth.

Start with the overview, keep the prompt guide open beside your first enqueue, and when Picture 1 walks into Picture 2 with birds in the mix — smile. The references just learned to move.

—Aurelia ♡