OmniHuman 2.0 prompts: official fields and limits
KIE documents the prompt as optional, for scene, character motion and camera movement, with a maximum of 300 characters. It lists Chinese, English, Japanese, Korean, Spanish (Mexican) and Indonesian for this text field. That is the documented prompt-language list; it should not be presented as an independent accuracy guarantee for speech in every language. The workspace keeps the same character limit and shows a counter while you type.
The museum-guide sentence above is the provider’s request example, not a prompt invented for this page. We quote it with its source so that you can inspect the original context. The guidance below explains how to use that documented field in this workspace. It does not introduce a hidden negative-prompt syntax, a model training command or an undocumented way to force a particular output.
Separate the spoken words from the visible direction
The required audio supplies what the character performs. The optional text describes how you want that performance to look. The example’s museum topic should therefore match a recording you have prepared; simply typing a subject here does not supply a spoken explanation. This generator does not replace the required audio file with text-to-speech. If you have only a script, create the recording separately and listen to the export before uploading.
Use an audio editor for a changed sentence, a removed pause or a shorter ending. Use the direction field for the visible behavior that the provider documents. Keeping those decisions separate makes a failed expectation easier to diagnose: a wrong word belongs to the source recording, while an unwanted visible gesture belongs to your performance review. Neither problem is solved by assuming the prompt is a complete editing timeline.
Organize direction around scene, motion and camera
Those three categories come from the provider’s prompt description. Before writing, decide which category actually needs direction for your clip. If the portrait already establishes the scene, use your limited space for the performance choice that matters most. Avoid adding unrelated instructions simply to fill the field. The character counter is a ceiling, not a target length or a reason to repeat the model name.
Read the finished direction back alongside the portrait and audio. Check whether it describes one coherent performance or several incompatible ideas. This is practical editorial advice based on the field’s purpose, not a tested formula for improving a quality score. We do not claim that a particular adjective, word order or prompt length has been benchmarked to outperform alternatives. Judge the generated result against your intended use.
Prepare inputs before refining the wording
The prompt is only one input. This website accepts one still JPEG, PNG or WebP portrait, up to 10 MB and four megapixels, plus a supported MP3 or WAV recording. The image must pass the dimensions and aspect-ratio checks shown in the workspace. These are this site’s validation rules; do not substitute a larger limit from a general provider page and assume the same file will be accepted here.
Use Preview cost to validate the actual files. A recording outside the supported duration range or an image that fails decoding cannot be fixed by more detailed performance direction. If validation rejects a file, export a supported copy and check it again. Keep the original material so that resizing, cropping and audio editing are reversible. The generator requires your permission to use the submitted image and recording.
Review a direction with comparable inputs
When you review a new direction, keep a note of the portrait and audio used for it. If you change the crop, recording and wording together, it becomes difficult to explain which change addressed the problem. A simple review log can record your intended behavior, what you observed and what you will change next. This is a suggested workflow for your own evaluation, not evidence of deterministic output.
Look at the whole performance rather than checking whether one requested gesture appears. Consider the mouth, expression, posture, framing and the transition into and out of the clip. The official field description establishes what direction you can provide; it does not guarantee exact choreography or timing. The examples page links to the provider’s demonstrations and supplies a review checklist without claiming independent benchmark results.
Check the quote, then confirm a new attempt
You can explore the direction field before committing to a generation. Preview cost measures the uploaded recording; final confirmation requires a signed-in account and sufficient eligible credits. The planning calculator on the pricing page explains the duration rule, while the validated quote is what you should review for the actual task. Reading or editing a prompt is not itself a paid generation.
After submission, follow the existing attempt in your private library. If a network interruption occurs, check that attempt before making a separate request. A different confirmed creative attempt uses credits again, and a provider-confirmed failure returns the credits for the failed task. Prompt wording cannot override purchase terms, input validation or the model’s output specification. Use the linked documentation for the current API contract and the site’s policies for the service you are using here.