Do I need a specific image type for best results?

Use a front-facing portrait with good lighting and the mouth unobstructed (no hands, mics, or heavy objects covering lips). Center the face and use a reasonable resolution (e.g., 512–1024 px on the short side). Avoid extreme angles and busy backgrounds.

What audio format works best?

Clean, speech-focused audio produces the most accurate mouth movements. WAV or high-bitrate MP3 typically work; mono voice tracks are preferred. Minimize reverb and background noise to improve alignment and realism.

Will the exported video include the audio track?

That depends on your SaveVideo node build. Some variants can mux audio; others export a silent MP4. If yours is silent, combine the output video with the same audio in an editor or a separate muxing step.

How can I improve lip sync and overall quality?

Start with a clear, centered portrait and clean audio. Match FPS to your target platform (e.g., 24–30). If a prompt field is available, describe subtle motions (blinks, small head turns) to avoid a static look. Try different seeds for variety, and consider cropping the face tighter if the mouth region appears too small.

LTX-2.3: Image Audio to Video

Back

This ComfyUI workflow turns a single image and a voice recording into a lip-synced talking video using the LTX-2.3 model. You load a portrait with LoadImage and provide speech via LoadAudio or capture it live with RecordAudio. Both streams feed the LTX-2.3 generator node (98ee9e5b-467b-40aa-a534-36033f27d0b4), which synthesizes a sequence of frames where the subject speaks in time with the audio. The resulting frames are encoded to an MP4 using SaveVideo.

Under the hood, LTX-2.3 conditions on the visual identity from your reference image and the temporal features of the provided audio to drive mouth shapes and subtle facial motions over time. The node typically exposes settings like output resolution and FPS, and many builds also include a seed for reproducibility and an optional text prompt to guide motion or style. The MarkdownNote in the graph documents quick tips and links to the official Lightricks model repositories so you can download the required weights.