Two requested training videos may serve different functions rather than duplicate one another. Baidu Xiling’s dedicated lip-model capture guide distinguishes a closed-mouth action base from speaking material. Follow the exact product’s current capture specification; a general portrait recording is not automatically a complete training package.
Write a role for each source file
Before filming, identify which source establishes appearance and reusable physical behavior, and which supplies speaking-mouth evidence. Keep those purposes visible in filenames and the shot list. Do not assume that a longer speaking clip compensates for a missing base, or that a silent-looking file satisfies all audio-track requirements in the capture specification.
Inspect the actual files, including their audio streams and start-to-end behavior. A recording can appear silent while still containing background sound, and a file called “base” can contain accidental speech. Record what is present rather than inferring it from the filename. Check the latest required formats and duration ranges without copying limits from another digital-human service.
Prepare one coherent capture session
For a hypothetical training session, plan the two source roles before changing lighting or camera position. Keep clothing, framing and exposure consistent where the provider requires it. A close speaking take and a differently lit full-body base may create a mismatched package even when each looks attractive alone.
| File role | Human inspection | Failure to resolve before submission |
|---|---|---|
| Base | Intended closed-mouth behavior and reusable gesture | Accidental speech or unsuitable action |
| Speaking material | Visible articulation and usable recording | Obscured mouth or incomplete speech |
| Both together | Same approved person and compatible setup | Changed appearance or mixed session |
| Submission manifest | Correct file assigned to each role | Swapped or missing source |
This is an original capture audit, not a claim that a submitted package will train successfully. The supplier’s current requirements remain the input authority.
Review the pair before reshooting everything
Watch both files completely and note the exact defective interval. Determine whether the defect is isolated, whether editing is allowed by the specification, and whether a partial reshoot would break continuity with the other material. Do not remove sections merely to meet a duration minimum while creating an obvious jump.
When a change of setup affects both roles, plan a coherent new pair rather than mixing incompatible takes. Keep the original files until the new package is accepted. A technical submission success does not establish that the trained person’s appearance, mouth movement or gestures are suitable for the intended scripts.
Validate the trained result with separate judgments
If training is completed through the chosen service, review appearance, articulation and inherited physical behavior separately using approved test content. Use material and permissions appropriate to the depicted person. Document any limitation before making client-facing promises.
In an Imagild project, these recordings may be external production assets or planning references; this article does not claim Imagild trains Xiling models or accepts the same package. Its purpose is to prevent an avoidable role mismatch at capture and submission, before effort is spent debugging the final video.