Reactor FastH3 Segment

FastH3 is the pack's "clip with sound from a text prompt" model, and this node is one clip of it. Write a prompt, say how long, optionally drop in a reference image, chain another one after it if you want a second beat. That's the whole idea.

What it's for

You're reaching for FastH3 when you want a generated shot that already has audio in it - footsteps, a door, a voice saying something - without going near a foley pass or a lip-sync pipeline. Locally that means Wan plus MMAudio or HunyuanVideo-Foley and a second model swap. Here it's one node and a prompt, which is the value proposition of the whole pack: the audio layer comes free with the generation.

A segment is one clip. FastH3 enqueues a clip per prompt, so a sequence of three segments is three clips playing back to back, each with its own sound. That's a very different feel from the "one continuous scene" models in this pack, and it's the honest reason to pick FastH3: you want distinct beats with their own audio, not a single drifting shot.

How it works

Each segment you queue becomes a clip request against the model. Because it's a clip model rather than a chunked one, your frame count isn't snapped to a grid - the frames field steps by ones - and the model decides the clip's real length. Measured on Reactor's own sessions, a "6 second" clip comes back at 158 frames at 24fps, not 144. Don't budget your edit around round numbers.

kind decides how a segment enters from the one before it: shot blends softly, cut starts fresh. The first segment just opens the video.

The inputs that matter

• prompt - one clip's worth of description plus whatever it should sound like. The model has a prompt guide on Reactor's docs, and it's worth reading before you write a paragraph of scene-setting.
• frames - how long this segment plays, default 120.
• kind - shot or cut.
• image (optional) - a reference image, and FastH3 takes one on any segment, not just the first. FastH3 fits images to its own canvas by the short side, so an odd aspect won't be stretched, just cropped.
• aspect - 16:9, 1:1, 9:16 or 4:3. Read on the first segment, so every clip in a chain shares one shape.
• sequence - the segments before this one. Empty on the first.
• REACTOR_SEQUENCE (output) - chain into the next segment, or feed the final one to Reactor Render.

FastH3 also runs live: the same model is available in Reactor Realtime, which is the fun way to audition prompts before you commit to a sequence.

Install

Manager, search comfyui-reactor-api, or:

cd ComfyUI/custom_nodes
git clone https://github.com/reactor-team/comfyui-reactor-api
pip install -r comfyui-reactor-api/requirements.txt

One dependency (reactor-sdk>=1.6), no weights. Add your key - REACTOR_API_KEY in ComfyUI's environment or under [API] in config.ini - and restart. Then copy the repo's example_inputs/ into ComfyUI's input folder so the bundled templates load, and open them from the template browser under comfyui-reactor-api.

Where people get burned

Aspect is set once. Change it on segment four and nothing happens; the first segment decides. Get the shape right before you build out the chain.

Only the frames you asked for, roughly. Clip length isn't chunked, so the visualizer won't warn you about granularity here - but plan on the model returning its own length, several percent off your number.

Reference images on later segments are fine here - which is unusual in this pack, where most models read an image only at the start. If you came from a LingBot or Helios article expecting that restriction, it doesn't apply to FastH3.

"Reactor" is two projects. Searching for this node will surface gourieff's ReActor face-swap pack, which is unrelated. You want reactor-team/comfyui-reactor-api.