LipSync GAP Sampler (Legacy)

PerfectSyncSampler ("LipSync GAP Sampler (Legacy)") is the original name for the pack's lip-sync sampler, preserved as an alias of GapLipSyncSampler. In the source's NODE_CLASS_MAPPINGS both names point to the one class - this is pure backward compatibility for workflows saved before the rename. If you see it in an old graph, it's the same sampler you'd get by placing the current node; nothing about the behavior changed.

What it does

Full LatentSync inference: take the models bundle, a batch of video images, and an audio track, and it runs the masked latent-diffusion pass that makes the mouth move in sync. The pipeline resamples audio to 16 kHz mono for Whisper, locates the mouth via Mediapipe, applies the feathered mask, and steps the UNet3D with a DDIM scheduler.

The inputs worth your attention, unchanged from the current sampler:

• steps - 20 default, 20–30 for solid sync.
• guidance_scale - 1.5 default; the "how hard does the mouth track the audio" dial.
• frame_rate - match it to the source video, or the words land late.
• video_resize_limit - default 1280; drop it to 768/512 on 12 GB cards.

Outputs: images (the synced frames, for your video encoder) and audio (the input track passed back, for muxing).

Install and gotchas

Same pack install as everything else here: ComfyUI Manager search LipSync GAP, or clone + install.bat, then restart. No separate model setup - the alias reads the same UNet, Whisper checkpoints, and face_landmarker.task as the current node.

The failure modes are the shared pack ones: artifacted or gray frames (lower video_resize_limit, debug with tiny Whisper), weak articulation (bigger Whisper in the loader), and VRAM errors (fp8, resize, shorter clips, in that order). One alias-only wrinkle: if an old workflow references this name and it comes up unknown after install, the node registry is stale - restart ComfyUI once and it resolves. It's the oldest name in the pack, but it's also the same honest workhorse.