How to Make an AI Avatar Look Real: A Safe QA Plan

Get clear consent for the face, voice, script, uses, places, and term.

Set the audience, task, channel, quality bar, and human help path.

Keep approved source files, versions, owners, and removal steps.

Use stable light, lens, framing, skin detail, eye line, and color.

Check speech pace, tone, names, mouth motion, gesture, and still frames.

Test varied scripts, screens, networks, crops, captions, and sound off.

Review every fact, claim, link, product view, and call to action.

Add a clear label or source record when the rule or setting calls for it.

Test drift, wrong output, correction, vendor exit, and fast removal.

Name the approved person, face, voice, dress, age range, accent, and use.

Choose reference clips for the final crop, light, pace, and emotion.

Set a stop rule for identity drift, false speech, or a misleading scene.

Match camera height, crop, background, light direction, color, and shadow to the final scene.

Record clean names, numbers, hard sounds, short lines, and long lines.

Avoid glare, harsh shade, moving light, loud room noise, and hidden face edges.

Keep the signed consent, source files, date, tool, and approved use.

Check what the plan includes and who may use, keep, or remove the source and model.

Count setup, render, review, fix, caption, export, and support time.

Use a human recording or another format when it looks or sounds better for the task.

Visual: skin detail, hair edges, teeth, eyes, shadows, clothing, and background.

Motion: lip close, jaw, blink, gaze, head, hands, loops, jumps, and drift.

Sound: pace, breath, stress, pause, emotion, accent, names, and numbers.

Delivery: phone crop, compression, captions, slow data, and final action.

Mark each fault by time, type, harm, owner, and fix.

Fail a clip for a changed fact, changed identity, unsafe cue, or lost next step.

Test with real target viewers for clear meaning, not only team preference.

Set a pass score before the review and do not move it to save a render.

Keep consent, source, version, approval, notice, publish, and removal records.

Reject output that changes age, body, accent, words, claim, or context beyond approval.

Give a human contact path for high-duty, sensitive, or disputed content.

Remove or replace the clip when rights end or a material fault is found.

Fail changed words, wrong numbers, identity drift, broken consent, or a false scene.

Mark lip, eye, hand, edge, light, speech, caption, crop, and sound faults by time.

Ask target viewers what they understood and what seemed synthetic.

Count staff time and rerenders before calling one tool cheaper.

Use the human or non-avatar route when it tells the truth better.