Reactor H3 Reference Turbo Realtime Segment

Most reference-image nodes in ComfyUI take one picture and hope. This one takes up to nine, and - this is the part people miss - it expects you to talk about them by number. Picture 1, Picture 2, and so on, in socket order.

What it's for

H3 Reference Turbo Realtime generates clips with sound, conditioned on a set of pictures: a person, a place, an object, a look. It's the node for "here's my character sheet and my location, now shoot a clip." A person from one photo, a garden from another, camera following from the side - written in the prompt like a shot note. The pictures guide the look rather than pinning a frame; this isn't image-to-video with the first frame held, it's reference conditioning.

If you've been through the local character-consistency slog - IP-Adapter, PuLID, half a dozen LoRAs, each with its own architecture limits - this is the hosted-model version of the same problem. Less control, no LoRA training, and one API key.

How it works

Sockets appear as you connect pictures, up to nine. The prompt is the only place the model learns which picture is which, so the numbering is the interface. Reactor's own example, straight from the node's tooltip: "The person in Picture 1 walks through the garden in Picture 2. The camera follows from the side."

Like FastH3, this is a clip-per-prompt model at heart: each segment enqueues one clip with sound at 24fps, and kind decides whether a segment blends (shot) or starts fresh (cut). And because the pictures are named by socket order, reordering sockets reorders meaning - socket 1 is always Picture 1, no matter which file you had there last render.

The inputs that matter

• prompt - where you refer to pictures by number. Get the numbering right or the model confidently casts the wrong object.
• pictures - the autogrowing image input, first socket to ninth. Connect them in the order you intend to name them.
• frames - length, default 120.
• kind - shot or cut.
• aspect - 16:9, 1:1, 9:16 or 4:3, read on the first segment, so the whole chain shares a shape.
• sequence - the earlier segments; empty on the first.
• REACTOR_SEQUENCE (output) - onward to the next segment, then Reactor Render.

This model is also live in Reactor Realtime. There the picture set is held for every later clip, so you pick your references at the start and then just keep retyping prompts. That's the nicest way to find a clip worth rendering.

Install

Manager search comfyui-reactor-api ("Reactor API: Realtime Video"), or:

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

One Python dependency (reactor-sdk>=1.6), zero model downloads, and you'll need a key: REACTOR_API_KEY in ComfyUI's environment or config.ini under [API]. Restart, then copy the repo's example_inputs/ into ComfyUI's input folder for the templates.

Where people get burned

Pictures aren't frames. People wire nine images expecting the first one to be held as frame zero, then wonder why the clip drifts away from it. It's reference conditioning - the look carries, the exact frame doesn't. If you need a fixed opening frame, that's a different kind of node.

Socket order is semantics. Insert a new picture at position one and every "Picture 1" in the prompt now means something else. Renumber your prompt when you renumber the sockets, or your character walks into the wrong garden.

Mixing references across models. A picture socket on a Vidu S2-Avatar node and a pictures socket here are not interchangeable, and Vidu's reference node explicitly rejects other models. Keep each model's references on its own chain.

Nine is generous, but not free. More references is more conditioning for the model to reconcile, and the clip gets vaguer rather than more specific. Two or three strong pictures usually beat nine scrappy ones.