OmniHuman·ByteDanceKonuşan avatarYeni
OmniHuman 2.0
ByteDance OmniHuman 2.0: tek bir görseldeki karakteri (birden fazla kişi de olabilir) bir konuşma ya da şarkı sesiyle konuşturur veya şarkı söyletir; dudak senkronu, mimik, el-kol hareketi ve kamera hareketiyle 720p video üretir. İsteğe bağlı prompt sahneyi, hareketleri ve kamerayı yönlendirir. Ses 4 saniye ile 5 dakika arası olabilir; ücret sesin uzunluğuna göre saniye başına hesaplanır.
En uygun olduğu işler
Parametreler
Bu model yalnız prompt ve girdilerle çalışır; ayarlanacak parametre yok.
Girdiler
Prompt'a ek olarak bu modelin aldığı dosyalar.
| Girdi | Kabul eder | En fazla dosya | |
|---|---|---|---|
Görsel image | image | 1 | Zorunlu |
Ses Dosyası audio | audio | 1 | Zorunlu |
OmniHuman için LUVI Prompt Engine kuralları
TARGET MODEL: OmniHuman 2.0 (ByteDance, audio-driven avatar video). Inputs: ONE image and ONE audio file. The audio supplies the words, the lip shapes, the rhythm and most of the mood. The image supplies who is in frame, identity, wardrobe, setting, framing and orientation. The prompt is optional and only adds what neither carries: who speaks, one or two concrete visible actions, a camera move, background motion. An empty prompt is a correct answer.
Uyulması gereken kurallar
- NAME THE SPEAKER WITH A SPEAKING VERB. Every prompt that has a person in it says who speaks and how: "He talks to the camera", "She sings", "The man on the right answers". When the user named a speaker by position or a visible attribute, keep that exact identification — on a multi-person image it is what picks the moving mouth.
- NEVER PUT THE SPOKEN WORDS OR LYRICS IN THE PROMPT. They come from the audio. When the user quoted the line, replace the quotation with the manner of speaking the user implied ("tells a joke", "asks a question", "talks warmly") — keep any manner words the user gave, add none.
- TWO SPEAKERS ON ONE AUDIO FILE: keep the order the user gave and write it as "First <speaker A> talks; then <speaker B> answers." with each speaker identified exactly as the user identified them. Keep every action the user gave a non-speaker ("The masked one watches.", "She listens.") as its own sentence; never invent one.
- NO NEGATION. "No smile", "don't move his hands", "without gestures" and a "Negative prompt:" line all measured as no-ops. Delete the label line entirely. Turn an exclusion into the positive state it implies only when that state is a visible thing the subject can do ("no hand gestures" → "keeps his hands still"; "no camera movement" → "The camera stays still."); otherwise delete it.
- NO SETTINGS IN PROSE. Video length is the audio length, framing and orientation come from the image, the output size is fixed. Delete aspect ratios ("9:16"), resolutions ("4K", "1080p"), durations ("15 seconds long"; an "At N seconds" cue on the FIRST action is timing, not a duration — keep it), "seed", weights "(term:1.3)" and "--flags"; keep the words they wrapped.
- DESCRIBE MOTION ONLY. Delete sentences that only restate what the image already shows (face, hair, clothing, setting) or ask to "keep" them unchanged — they measured neutral and cost nothing to drop. Keep a description only when it identifies WHICH person speaks or acts. This deletes appearance and setting only — every body or face state the user asked the person to perform stays ("eyes closed", "hands in his pockets", "head tilted").
- SEQUENCE WITH "first … then …". Only the FIRST timecode is honoured (an opening action can be delayed onto its second); later timecodes landed about a second late. Keep the timecode the user put on the first action word for word ("At 1 second, he turns his head to the harbour; then he looks back at the camera; then he nods."); turn every later one into "then". An action tied to a spoken word ("when he says 'door', he nods") does not land on the word — rewrite it as "then he nods" in the user's order.
- ONE CAMERA MOVE AT MOST, and only if the user asked for one. Keep the move ("The camera slowly pushes in.", "The camera orbits slowly to the right."). A requested end framing ("…ends on a close-up") is ignored by the model — keep the move, drop the landing clause.
En iyi sonuç verenler
- Structure, blocks dropped when they do not apply: camera move → speaker + speaking verb → the user's actions in order, most important first → what the other people do (multi-person) → background motion.
- Short natural sentences, present tense. No tag lists, no quality or style boosters ("cinematic", "8k", "masterpiece", "ultra realistic") — delete them.
- Keep every action the user listed, in the user's order. You do not know the audio length, so never add actions and never drop one; a later action may simply not fit a short clip.
- Mood follows the audio. Keep every emotion or manner word the user wrote ("passionately", "warmly", "seriously", "with a smile") exactly where it was; never add one.
- When the user wrote in another language, translate the MEANING of each gesture, not the words: Turkish "başını sallıyor" is a nod ("nods"), not "shakes his head"; "kafasını iki yana sallıyor" is "shakes his head". If a gesture is ambiguous, pick the everyday meaning in that language.
- If nothing is left after the deletions above, return an empty string — the audio and the image already carry the performance.
LUVI Prompt Engine'in bu aile için kurallarından; çalışma alanındaki yeniden yazıcı ve LUVI'nin MCP bağlayıcısı aynı kuralları uygular.