OmniHuman2
OmniHuman 2.0 lip sync starts with a readable face and a clean voice track. BytePlus describes its avatar tier as making a realistic talking video from one image and audio, but that is not a guarantee for every photo. Keep the mouth visible, trim dead air, then inspect the render against the words before paying for another take.
Start with a mouth the model can see
Pick a chest-up, front-facing image with even light across both cheeks. You want a closed, relaxed mouth, a defined jawline, and enough pixels around the lips to spot their shape at normal viewing size. A hand over the chin, a heavy mustache hiding the mouth, a sideways face or a glass held in front of the lips gives any image-driven animation less visual information to work with. Those are production heuristics, not input rules published by BytePlus.
Don’t confuse a sharp face with a suitable one. An extreme smile exposes teeth and sets up a hard transition into speech; a neutral expression gives the animation a quieter starting point. If you can take a new picture, shoot two or three variations at the same distance instead of enlarging a tiny head from a group photo. Read the reference image guide for the rest of the framing choices.
BytePlus says OmniHuman 2.0 accepts a single image with audio and can use an optional text prompt. It also lists native 720p output. That makes a tight face shot useful for checking lip shapes, but don’t promise yourself that a poor source image will become a crisp close-up after rendering. Judge the reference at the size your final viewer will see.
Fix the track before fixing the prompt
Listen to the audio on headphones. If the first consonant is clipped, rerecord or replace the line; don’t ask a video prompt to recover a syllable that isn’t there. Remove long silences before the first word and after the last one, then leave a little breathing room so the performance doesn’t feel cut off. If you’re making a talking explainer, write the script for speech, not for a page: short clauses with natural pauses. Pick words you can pronounce without rushing.
Keep one voice prominent. Music or sound effects mixed over the voice can make review harder because you won’t know if an apparent mouth mismatch is a vocal cue you can barely hear. For a first pass, submit the clean spoken track and add music in an editor afterward if your workflow allows it. That’s an editing recommendation, not a BytePlus specification for supported audio mixes.
BytePlus’s product page doesn’t document a supported-language list. It doesn’t publish a word-level timing accuracy target there either. Don’t promise a client phoneme-perfect sync in a particular language on the strength of the product card. Use a short approval render using the actual speaker and language. Keep the script style you will use in the campaign.
Make a controlled first render
- Choose one clear portrait and one clean speech recording. Keep copies of both so you can identify which change mattered later.
- Start with a simple speaking shot. If your interface offers a text prompt, describe framing and a restrained delivery, such as: “Chest-up, facing the camera, calm delivery, small head movement, steady framing.” BytePlus says text prompts can guide camera and body motion; it doesn’t say they can manually force a precise mouth shape on a given syllable.
- Generate a sample using the access route you’ve confirmed. The step-by-step setup guide covers the general image-and-audio workflow, and the API access guide covers the developer path. If your console labels or endpoints differ, check the current BytePlus interface or API documentation rather than guessing from an old screenshot.
- Watch it once with the sound on and once with it off. Note where the mouth starts moving, where it stops, and whether the face holds together on wide vowels or quick consonants. Review at normal speed first. Frame-by-frame scrubbing can turn a perfectly acceptable short ad into an endless repair job.
Change one thing per rerender. Start small. If the lips are obscured, swap the photo; if the syllables feel late, check the edit and the start of the audio first. Don’t rewrite the camera prompt, change the image and rerecord the voice at the same time. You’d have no way to tell which input helped.
Check the right kind of mismatch
A mouth that’s visibly moving before the first spoken word points you toward the clip boundary or the animation itself. A mouth that opens roughly on cue but looks mushy during speech points you toward face visibility, crop size or the limits of that render. These are diagnoses to test, not documented fixes built into OmniHuman 2.0.
Watch for two non-mouth problems as well. A fast head turn can hide one side of the lips right when you need to judge them; a hard camera move can make a tiny mismatch look worse. Keep the approval shot simple. No whip pans. You can try a livelier version afterward with the camera and gesture prompt examples. The mouth has to sell the sentence before the camera sells the scene.
Compare the export against the original audio, not against an edited social clip with music and captions. Put both on a timeline, find the first clear word and the last one, then note any drift you can actually see at playback speed. If one syllable looks odd but the whole line reads naturally, decide whether the viewer will notice. That’s a better editorial test than magnifying a single frame until it looks strange.
Spend on the failure you need to solve
BytePlus lists its lip-sync-focused avatar model at $0.12 per second. A 30-second output priced at that rate would be 30 s × $0.12 = $3.60 before any other charges or attempts. BytePlus lists the film-grade tier at $0.07 per second and describes camera and body movement plus a 30-to-300-second output range for that tier. The same 30-second arithmetic there is 30 s × $0.07 = $2.10. The pricing guide breaks down longer durations.
For a plain head-and-shoulders spokesperson, I’d budget a lip-sync-focused avatar pass before spending time on elaborate motion prompts. The rate difference on a 30-second example is 30 s × ($0.12 − $0.07) = $1.50, and another failed take can cost more than that choice. This isn’t a published head-to-head quality result; it’s a way to decide what you are buying. Check the billing terms shown in your account before ordering production volume.
If you’re still deciding whether the input is workable, make the portrait and voice recording first. You can make a video on OmniHuman2.net when those assets are ready, but keep the untouched audio file: if the output seems off, it is the only reliable reference for finding the first bad syllable.
