- 01
Define the identity and consent boundary
Begin with the people and assets, not the effect. Identify the likeness owner, the person operating the workflow, the intended audience, the destination, and the exact permission covering capture, transformation, storage, and publication. Distinguish a fictional character, a consenting performer, a private person, and a public figure. Keep consent records with the source asset and specify whether the permission covers training, previews, commercial use, paid advertising, and later edits. Never infer consent from public availability. At handoff, preserve the delivery checklist, and ask the continuity editor to record camera logic before the production checkpoint.
- 02
Separate local context from vendor claims
The word Vietnam can describe the audience, the production location, the account region, the language, or the applicable review team. Record which meaning applies. Verify current first-party documentation for service location, supported languages, account restrictions, storage, deletion, and export. A page mentioning Vietnam does not establish local hosting, local compliance, local support, or a permitted use case. Escalate jurisdiction-specific questions to qualified counsel instead of converting a workflow guide into legal advice. At handoff, preserve the handoff draft, and ask the model evaluator to record disclosure clarity before the editorial approval.
- 03
Test the visible controls with a safe sample
For an authorized evaluation, use a low-risk sample and a written acceptance checklist. Observe face selection, alignment, expression handling, occlusion, lighting, hair and accessories, temporal consistency, audio or lip-sync behavior, masking, preview quality, export resolution, and failure handling. Change one meaningful variable at a time. Record input provenance, settings, date, account context, output identifiers, and reviewer notes. Do not use a successful preview as proof that the same result is repeatable across people, scenes, languages, or destinations. Before moving on, preserve the brief version, and ask the model evaluator to record source fidelity before the rights check.
- 04
Review provenance and disclosure
A transformed face can be mistaken for an authentic recording. Preserve the original, the transformed version, edit history, consent record, and delivery decision. Decide whether the audience, platform, client, or internal policy requires disclosure. Check the source and destination terms for synthetic media, impersonation, advertising, political, biometric, and identity-related restrictions. A technical result is not sufficient evidence that a publication is safe, permitted, or acceptable to the person represented. At the next gate, preserve the evidence table, and ask the release approver to record destination fit before the workflow transfer.
- 05
Compare workflows by control and reversibility
When comparing providers or local workflows, use the same authorized source and the same acceptance criteria. Score controllability, identity protection, provenance retention, preview and export behavior, reviewer access, deletion, cost, latency, and recovery from a failed result. Treat missing documentation as an evidence gap. Prefer a workflow that keeps the original untouched, limits access to the minimum team, and allows the transformation to be withdrawn without losing the source or audit trail. At handoff, preserve the continuity note, and ask the model evaluator to record visible continuity before the reversible handoff.
- 06
Set a release gate for Vietnam-facing delivery
Before delivery, name the accountable reviewer and check the current first-party platform rules, client requirements, consent scope, localization, captions, audio, disclosure, and destination format. Recheck any changing product or policy statement immediately before publication. If identity, authorization, provenance, or jurisdiction remains unresolved, stop at the research or internal-review stage. This guide supports a careful decision; it does not certify a vendor or a legal outcome. For a controlled test, preserve the failure note, and ask the claims reviewer to record human approval before the final sign-off.