react-1

react-1, from the company Sync, does something you can't get from an open-weights video model on a consumer GPU: it takes an existing video of a person and re-animates their facial performance to match a different audio track. New dialogue, new emotional delivery, same face, same camera move. That's the whole node. Two required inputs - video and audio, each a UUID or URL - and a VIDEO comes back.

Let's be clear about what this is: it's performance transfer, the thing people call "dubbing that looks native" and the thing that also makes deepfakes trivially easy. It's a hosted API model, and Runware's content safety is the only thing standing between you and abuse of someone's likeness. Use it on your own footage or with permission; the tooling gives you rope.

What you set

• video / audio - required, as UUIDs or URLs (Runware's upload nodes will stage these for you). The video is the appearance and motion source; the audio is the new performance you're driving it with.
• providerSettings.sync.editRegion - what moves: face, lips, or head. face is the broad default, head adds head motion for bigger deliveries. This is your first quality lever when the output looks "pasted on."
• providerSettings.sync.emotionPrompt - happy, sad, angry, disgusted, surprised, neutral, or default. You're not just matching the audio; you can push the emotional register it's read with. "Angry" on the same line can sell a scene that flat lip-sync would kill.
• providerSettings.sync.temperature - expressiveness of the facial movement, 0–1, default 0.5. Low is stiff and clinical; high is theatrical. Where most people get burned: they leave this at the default and blame the model for dead-eyed output when cranking it to 0.7–0.8 fixes the "uncanny" feel.
• providerSettings.sync.syncMode - what to do when audio and video lengths don't match: bounce (default, ping-pongs), loop, cut_off, silence, or remap. remap is the one that stretches or squeezes the video to fit the audio; the others fill or trim.
• providerSettings.sync.activeSpeakerDetection.autoDetect - automatically find the talking person in a multi-person shot. Turn it on when there's more than one face.
• providerSettings.sync.occlusionDetectionEnabled - handles faces that are briefly covered (hand over mouth, passing object). On for talking-head footage where people gesture; it buys robustness at some latency.

numberResults (up to 4) gives you takes, outputFormat is MP4/WEBM/MOV, and safety.mode (none/fast/full) controls the content gate.

How it works

The node sends sync:react-1@1 as a videoInference task through the Runware SDK - your audio and video are referenced by their hosted URLs, the API does the performance transfer, and the finished clip comes back as a native VIDEO socket. It runs entirely server-side; nothing is computed on your machine, and the title bar prints the cost and NSFW verdict when it's done.

Install and gotchas

cd ComfyUI/custom_nodes
git clone https://github.com/Runware/ComfyUI-Runware
pip install -r ComfyUI-Runware/requirements.txt

Restart; API key from the Runware dashboard into Settings (or RUNWARE_API_KEY). It's paid, per-run, minimum top-up.

Real-world pain points: it's not a fast node - performance transfer over the API takes a while, so batch sensibly. The temperature default really is the difference between watchable and uncanny, so touch it before you touch anything else. And the input format trips people up: both inputs are strings (UUID or URL), not file paths or IMAGE/AUDIO sockets, so stage media through Runware's upload nodes first and paste the reference in.