arkennemasis fal · Sync Lipsync 2 · $3/min

If you have a face and a voice and you need them to agree, this is the node you'd actually reach for. Sync Labs' lipsync-2 model, $3 per minute of video, no length cap in the node's checks, four widgets. It doesn't do emotion and it doesn't do 4K detail; it makes a mouth move correctly, and $3 a minute is the sane middle of this pack's lip-sync ladder - a 20-second talking head is a dollar.

The one setting that matters

There's a model combo with two options, lipsync-2 (default) and lipsync-2-pro, and the field's own text tells you the trade: pro costs roughly 1.67× as much for the same duration. The node's price badge multiplies accordingly, so you don't have to hold the arithmetic in your head. Leave it on lipsync-2 here - if you want the pro pass, there's a dedicated node in this pack and the decision is worth making deliberately rather than by leaving a combo in the wrong place.

Then sync_mode, which is what happens when the audio and the video disagree about how long the shot is: cut_off (default), loop, bounce, silence, remap. The names describe the behaviour - the shorter of the two sets the pace, and cut_off simply stops. If your narration runs longer than your clip, the default is the mode most likely to leave you with a mouth that stops talking mid-sentence. Match the durations before you wire them: the pack's Narration Length node exists to make the shot long enough for its own voice-over.

Inputs and outputs

video and audio are both required; there's no upper length limit enforced by the node, which is why this one is usable for real dubbing work rather than per-15-second fragments. Outputs are video (ComfyUI VIDEO - save it, or feed the assembly chain), video_path (STRING, the file under output/fal/fal-ai_sync-lipsync_v2/), and info (JSON: request id, cost estimate, saved files, fal's full answer).

Around that, the shared fal furniture: FAL_KEY from the .env next to your launcher and never from a widget anywhere in this pack, a cost estimate compared against max_cost_usd (default $20) before any upload, max_concurrent (default 1) so paid nodes go one at a time unless you raise it, and reuse_identical_run (default on) so re-running an identical job doesn't bill twice - even after a ComfyUI restart.

Also check the cheap option first. If the video is your own render and you already know the exact dialogue timing, a local audio-driven model may get you there for electricity instead of $3 a minute. Where this node wins is a real face from real footage, in one call, with no local model wrangling.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/Hishamahmer/comfyui-arkennemasis

Restart, or install through ComfyUI-Manager by searching arkennemasis (pack title comfyui-arkennemasis). The only dependency in requirements.txt is httpx, and that serves the pack's Codex nodes - the fal provider is standard library only, so this is a clean install into an existing environment.

FAL_KEY=your-key-from-fal.ai/dashboard/keys

goes in the .env beside run_nvidia_gpu.bat.

Where people get burned

• A mouth that stops mid-sentence. sync_mode was cut_off and your audio was longer. Fix the durations, or pick a mode that pads or repeats.
• You accidentally paid 1.67× for the pro model. That model combo is the difference between $3 and $5 a minute.
• A slow, expensive run on the wrong face. Lip sync can't fix a face that's turned too far away, half off-screen, or inside a hard shadow. Frame the subject toward camera, in decent light, mouth visible.
• You forgot the audio is a ComfyUI AUDIO value. Wire it from a speech node or a loader; it's not a file path.
• ComfyUI died while a long job was running. The fal request didn't - and you were charged. Take the id from output/fal/_requests.jsonl and collect the result with the pack's free fal Recover Result node.