Improving Lip Sync Quality

Audio-video misalignment typically stems from one of three causes: sync mode configuration, audio quality, or duration mismatch. First, check your sync_mode setting — cut_off mode trims audio that extends beyond the video length, which works well for most cases. If audio and video durations are significantly different, the remap mode adjusts video playback speed to match, but large speed changes can look unnatural. Try cut_off if you are seeing drift. Second, ensure your audio is clean — background music, overlapping speakers, and heavy noise make it harder for the model to align lip movements to the correct speech patterns. Use a noise reduction tool on your audio before submitting. Third, verify that the audio language matches what the model expects; Sync supports all languages, but audio with mixed languages in a single track can cause alignment issues.

Long videos are internally divided into 30-40 second chunks for processing. If chunk boundaries fall at points where the face is partially visible, moving rapidly, or absent, the output quality can degrade at those transitions. For the best results with long-form content, consider splitting your video into segments under 2 minutes using the Segments API and processing each segment separately. Use lipsync-2-pro for long-form content where quality is critical — its diffusion-based super resolution handles transitions between chunks more gracefully. Also ensure the speaker’s face is consistently visible and well-lit throughout the entire video. Rapid scene changes within chunks are a common cause of processing timeouts and quality drops.

Visual artifacts around the face region usually result from challenging input conditions. Start by ensuring the speaker’s face has good, even lighting without harsh shadows — uneven lighting causes the model to produce inconsistent skin tones across frames. Verify the face is front-facing or near-frontal, as extreme angles produce distortion. Remove any obstructions covering the mouth area, including hands, microphones, or hair. If obstructions are unavoidable, enable occlusion_detection_enabled in your generation request for better handling of partially hidden faces. For persistent artifacts, try lipsync-2-pro, which uses diffusion-based super resolution to produce cleaner output around teeth, beards, and fine facial features. Check that your input resolution is at least 480p; very low-resolution faces make detection and generation less reliable.

When a video contains multiple faces, Sync’s default behavior selects the most prominent face in the frame. To target a specific person, use the Speaker Selection feature. Speaker selection lets you identify the correct face using automatic detection (auto_detect) or by providing a bounding box or frame number reference. In the API, set the active_speaker_detection option in your generation request. In Sync Studio, use the speaker selection tool in the video player controls. Note that speaker selection is available for lipsync models only — react-1 does not support this feature and requires a single visible speaker. For videos with multiple speakers taking turns, use the Segments API to define time ranges and assign speaker selection per segment.

Watermarks appear on videos generated with free or Hobbyist accounts. To remove watermarks, upgrade to the Creator plan or higher — watermark removal is included on all Creator+ plans. Existing videos generated on a free or Hobbyist plan will retain their watermarks; you need to regenerate the video after upgrading to get unwatermarked output. See the Billing page for plan details, pricing, and upgrade instructions. Note that the watermark is applied during generation, not as a post-processing overlay, so it cannot be removed from already-generated videos without re-running the generation on a plan that includes watermark removal.