One clear, chest-up portrait is a sensible starting image for OmniHuman 2.0. Keep the face unobstructed, leave room around the shoulders if you want visible movement, and match the photo to the shot you actually need. These are production choices, not a published BytePlus checklist or a promise that any particular photo will work.<\/p>\n
BytePlus describes OmniHuman 2.0 as generating character video from one image and an audio track, with an optional text prompt. The still image casts the shot. It sets the frame. You can’t ask an invisible shoulder to make a convincing on-camera gesture and expect the source photograph to tell the system what it looks like.<\/p>\n
For a spoken product explainer, start with a subject looking near the lens and enough space to see the upper torso. For a tighter avatar shot, put the face where the viewer’s attention will land: eyes sharp, lips visible, chin inside the frame. If the brief is a full-body performer, use a source image that actually includes the body. Don’t crop a waist-up portrait and write "full-body dance" in the prompt as if the missing legs were already established.<\/p>\n
No official pixel-size rule here. Just fewer open questions before you pay for a render. Save the original image and make a separate crop for each intended shot, so you know which choice produced which result. A horizontal crop for a desktop explainer and a vertical crop for a mobile ad answer different framing questions; don’t judge them as interchangeable photos.<\/p>\n
A mouth hidden behind a microphone, hair across one cheek, dark sunglasses, or a heavy beauty filter creates a harder starting point for speech animation. Use a relaxed expression with the lips closed or gently parted. Avoid a wide grin with individual teeth sharply exposed if your goal is natural talking rather than a fixed smiling pose. These are editing choices. They’re not tested OmniHuman 2.0 failure rates.<\/p>\n
Look at the image at the size your audience will see it. Can you identify the lip line? Is one side of the face crushed into shadow? Does a stray strand cut straight across the mouth? If yes, fix the source. Reshoot if needed. I would spend ten minutes reshooting a clean portrait before paying to rerun a clip from a photo with an avoidable obstruction.<\/p>\n
Background clutter can be a problem of direction, too. A poster with small type, a mirror that doubles the subject, or another person at the edge of frame gives the shot more visual business than a talking-head explainer needs. Pick a plain wall or a deliberate set. Keep the clothing edges and the neck visible when you can; those boundaries help the image read as a person in a space, not a face pasted onto a backdrop.<\/p>\n
The BytePlus product page says the model uses a voice track and can take a text prompt for further refinement. Audio carries the performance. The image supplies the on-camera identity. A portrait of someone visibly laughing while the narration is a solemn safety notice is a mismatched starting brief, even if the file itself is sharp.<\/p>\n
Before upload, listen to the exact audio from start to finish. Cut the unused count-in, a long tail of silence, or a breath that wasn’t meant to open the video. Check that the voice is intelligible at normal volume; don’t make the model guess through a loud music bed. This is basic editing advice, not a claim that BytePlus requires a particular loudness, codec, or silence threshold. If you need background music, decide in advance whether it belongs in the generation input or in the final edit.<\/p>\n
For a shot that should remain calm, write a modest direction such as "speaker faces the camera, small natural head movement, hands remain below frame." For a wider shot, name one visible action: "speaker turns slightly toward the product on the table." The official product page describes text-driven camera and movement control, but it doesn’t promise that a specific wording will produce an exact gesture on every run. Use prompts as directions, not choreography guarantees.<\/p>\n
Price is where a vague photo brief gets expensive. The BytePlus listing shows $0.07 per second for the film-grade tier and $0.12 per second for the avatar tier. It advertises 30 to 300 seconds of video output for film-grade. For a 30-second film-grade render, 30 s \u00d7 $0.07 = $2.10; three candidate crops at that length would be 3 \u00d7 $2.10 = $6.30, assuming each is billed at the listed per-second rate. A 30-second avatar-tier render at its listed rate is 30 s \u00d7 $0.12 = $3.60. Check the live billing terms in your account before spending; the page alone does not settle every billing detail.<\/p>\n
That 30-second floor matters when the final ad is only 15 seconds. Don’t assume a 15-second film-grade job is available just because the edit is short. The published range is attached to the film-grade offering on the product page, and the page doesn’t spell out every avatar-tier duration rule. If the brief calls for six takes, budget six billable outputs rather than one imagined perfect take.<\/p>\n
Run a controlled comparison: keep audio and prompt unchanged, then try one chest-up crop against one tighter face crop. Record the source image, crop, tier, intended output length, and what you want to inspect in the result. Judge the mouth at consonants, the edge of the jaw during turns, and whether the final framing actually works in your edit. Skip the benchmark theater. Your own shot list will tell you more than a generic "best photo" claim. <\/p>\n
Get permission to use the person’s likeness and voice before building the clip. If you are making an ad for a client, confirm their approval process and label synthetic footage where your distribution context calls for it. The BytePlus product page does not, on its own, document every commercial-use term, supported speech language, or region where access is available. Ask the provider first. Don’t promise delivery in a contract on the strength of a product page alone.<\/p>\n