Custom AI Avatars: A Practical Workflow for Video Teams

A custom avatar is a presenter, not a story. Teams that get real value from synthetic presenters treat the avatar as one layer inside a production pipeline — the layer that introduces, explains, and signs off — while the rest of the video does the persuading. Get that division of labor right and one recording session can power a year of onboarding modules, product updates, and localized training. Get it wrong and you end up with a talking head that is almost right, which is the most expensive outcome of all.

This guide walks through the whole chain: choosing the right job for an avatar, shooting footage that survives training, reviewing the first output honestly, writing scripts a synthetic voice can deliver, building a shot list around the presenter, keeping a series consistent, handling consent and disclosure, and knowing when cinematic AI video should take over.

Start With the Job Your Avatar Has to Do

Before you book a studio or write a script, decide what the presenter is for. Custom avatars earn their keep in a narrow band of formats:

• High-volume explainers where the same face appears across dozens of videos.
• Onboarding and compliance training that gets revised every quarter.
• Product walkthroughs where the presenter narrates over a screen recording.
• Localized versions of one message across several languages.
• Personalized follow-ups where the prospect's name appears in the opening line.

They struggle in formats that depend on performance: brand films, comedy, anything where a raised eyebrow carries the meaning, or scenes with fast physical action. A synthetic presenter can hold a close-up while it talks. It cannot land a joke.

A practical decision rule: if the video's value comes from information the viewer must retain, an avatar is a sound choice. If the value comes from a feeling the viewer should catch, use real footage, real actors, or generated B-roll instead. Volume matters too. Forty short modules justify the setup cost of a high-fidelity custom avatar; a single two-minute launch video usually does not.

Write the job down in one sentence before anything else: "A ninety-second explainer that teaches new hires how expense reports work, refreshed every quarter, in three languages." That sentence decides your footage, your script length, and whether you need a full-body or a waist-up avatar.

The Footage Checklist That Sets Your Avatar's Ceiling

Every custom avatar is a compression of the footage you feed it. No amount of post-processing recovers detail that was never captured, and no platform setting fixes a hard shadow under the nose.

Camera and light

• Lens height at the subject's eye level. A camera that looks up at the chin produces an avatar that always seems to be looming.
• A moderate focal length — roughly 50–85mm equivalent — so facial features are not stretched at the edges.
• Soft key light at about 45 degrees, a fill on the opposite side, and a subtle hair light to separate the subject from the background.
• No hard shadows across the eyes. Reviewers notice dead eyes before they notice anything else.
• A plain, mid-tone background at least 1.5 meters behind the subject. A busy office shelf becomes visual noise in every future video.

Wardrobe and grooming

• Solid mid-tone colors. Fine stripes and tight checks shimmer during the motion synthesis pass.
• Hair kept away from the face, no dangling earrings, no reflective pins.
• Minimal retouching. Keep skin texture; over-smoothed plates produce a waxy result that reads as artificial on a phone screen.

The read itself

Record five to eight minutes of speech, not sixty seconds. Cover a range of sounds, sentence lengths, and emotions: a neutral explanation, a warm greeting, a serious compliance warning, a list of numbers. Read a script full of plosives and sibilants so the training material includes difficult mouth shapes.

Keep your hands relaxed and out of frame unless you specifically plan a waist-up avatar. Sit or stand the same way for the entire session — slouching halfway through creates two different body languages in one model.

Finally, capture clean audio separately if you plan to clone a voice, and get a signed release from the person on camera before you start. Both of those are far easier to arrange before the session than after.

Training, Review, and the First Quality Gate

Once the footage is processed, resist the urge to approve the first output because the likeness is impressive. Likeness is the easy part. Review at full resolution on a monitor and then again on a phone, because most of your audience will watch on a phone.

A workable review checklist:

1. Lip closure on m, b, and p sounds. Failing here is the single most common tell.
2. Sibilance — does the s sound produce a smear around the mouth?
3. Teeth and tongue during wide vowels and th sounds.
4. Blink cadence. Roughly fifteen to twenty blinks per minute reads as human; a fixed stare or a machine-gun blink does not.
5. Head micro-motion. Natural drift at sentence boundaries, no slow mechanical sway.
6. Seams at the jaw, ear, hairline, and collar, especially when the head turns.
7. Shoulder and torso drift during long pauses.

Build a twenty-second stress-test script with a name, a number, a long compound word, and a question. Run every new avatar through it before you write a single line of production script. If the stress test fails, the fix is almost always upstream: better key light, closer framing, or a re-record of the difficult sounds rather than a different platform.

Keep the approved stress test as a reference clip. When you revisit the project months later, it tells you instantly whether an output regression is your script or the model.

Writing Scripts a Synthetic Presenter Can Deliver

Synthetic presenters do not improvise around awkward phrasing. They read exactly what you give them, which means the script has to be written for the ear.

Keep sentences short. Twelve to eighteen words. If a sentence needs a comma to survive, split it.

Spell out anything ambiguous. "Q3" becomes "the third quarter." "$1.2M" becomes "one point two million dollars." Product names get a phonetic spelling in a pronunciation note. Acronyms that read fine on a slide are landmines in speech.

Mark the pauses. A presenter that never breathes sounds like a phone menu. Add a beat before a key claim and after a list. If your platform accepts pause markers or SSML-style breaks, use them; if not, write short paragraphs and let sentence boundaries do the work.

Read it aloud and cut fifteen percent. Every script benefits from this, and avatar scripts benefit more than most because there is no performance to cover a flabby sentence.

Separate spoken text from on-screen text. What the presenter says and what appears in the lower third should not be identical. The viewer reads faster than anyone speaks.

Time the call to action. If the ask lands at second eight, it will be forgotten by second sixty. Put it in the final ten seconds and repeat it in the description.

Building the Shot List: Presenter, B-Roll, and Motion Graphics

A video that stays on one synthetic face for ninety seconds is exhausting regardless of how good the avatar is. Plan cuts the same way you would for a human presenter.

A simple, repeatable structure for a ninety-second explainer:

Time	Visual	Purpose
0:00–0:06	B-roll or product shot, no presenter	Hook the viewer before the face appears
0:06–0:20	Presenter, waist-up	Establish who is speaking and why
0:20–0:40	Screen recording or diagram with voice-over	Deliver the detail
0:40–0:48	Presenter	Re-engage and pivot
0:48–1:15	Mixed B-roll, close product shots	Show the thing being described
1:15–1:30	Presenter	Close and call to action

A rough 40/40/20 split — presenter, demonstration, graphics and transitions — keeps attention without turning the video into a slideshow.

This is also where cinematic AI video earns its place. Abstract concepts, product macro shots, and scene-setting transitions are expensive to shoot and trivial to generate. The Orelon AI video generator can produce the establishing shots and B-roll that surround the presenter, and the video templates library is a fast way to find a structure before you commit to a script.

Consistency Across a Series: Face, Voice, and Look

Episode twelve should look like episode one. That does not happen by accident.

Create a look bible with three reference stills: one presenter frame, one B-roll frame, one graphics frame. Record the color values, the font, the lower-third position, and the target loudness — around minus fourteen LUFS for web delivery is a reasonable default. Store the approved voice settings and speaking rate alongside it.

For visual continuity, keep the presenter's framing identical across episodes. Same crop, same background plate, same wardrobe palette. If backgrounds vary, generate them from a small set of approved plates rather than inventing a new one each time; a consistent AI image generator workflow with saved prompts makes that easy.

Name files in a way that survives a handover: series-ep04_presenter_v03_approved.mp4. Future you, and the editor who replaces you, will be grateful.

Finally, watch two episodes back to back before publishing the third. Mismatched color grade and loudness jumps are the two continuity errors audiences feel without being able to name.

Language, Localization, Consent, and Disclosure

If a video will appear in more than one language, decide early whether you will dub the existing presenter or regenerate the performance in each language. Dubbing is faster and keeps one visual performance; regeneration gives better lip sync per language but multiplies review time. For most training content, dub first, then regenerate only the languages that carry the most viewers.

Expect language-specific friction. German compounds and Spanish syllable density change how long a sentence takes to say, so a script that fits sixty seconds in English may run seventy-five in translation. Localize the script, not just the subtitles.

Add captions to every video. They improve retention, they make the content usable in silent autoplay feeds, and they are an accessibility requirement in many organizations.

On consent: a voice or face model is personal data. Get written permission that specifies scope — which projects, which channels, how long — and include a revocation clause. Store the source footage somewhere access-controlled and document who can generate from it.

On disclosure: a short on-screen label such as "presented by an AI avatar" costs you almost nothing and protects you in regulated industries, internal communications, and anywhere the audience might reasonably assume a human is speaking. Check your organization's policy before the first publish, not after.

Mistakes That Get Caught in the Final Review

• Over-polished script. Every sentence perfectly balanced sounds like a brochure. Leave one plain, slightly informal line in each section.
• Presenter overuse. More than forty percent of screen time on a synthetic face invites scrutiny of the face rather than the message.
• Grade mismatch. Avatar footage and B-roll graded separately, then cut together, produces a visible seam at every transition.
• Loudness drift. Voice-over normalized by ear, B-roll music normalized by a different tool. Normalize everything to one target.
• Mispronounced names. The one thing viewers will comment on. Always stress-test proper nouns.
• Uncanny smiling. A presenter grinning through a compliance warning undermines the warning.
• Missing captions. Half the audience is watching with the sound off.
• No version control. Two people generating from different saved settings, then arguing about which output is canonical.

Where Cinematic AI Video Fits Around the Presenter

The presenter carries the argument; cinematic footage carries the feeling. Splitting those jobs is what separates a watchable explainer from a recorded slide deck.

Use generated footage for the moments that are expensive or impossible to shoot: an establishing shot of a city at dawn, a slow macro push across a product surface, an abstract visualization of a data flow, a transition that bridges two chapters. Keep the presenter for the claims, the caveats, and the call to action.

A hybrid workflow that works well: draft the script, mark the beats where the viewer needs a visual breather, generate those shots in a batch, then assemble around the avatar footage. The prompt library is useful for keeping the visual language consistent between batches, and the Seedance examples show how far motion quality has come for short cinematic inserts.

If you are still choosing a production stack, the Orelon blog covers workflow comparisons, and the alternatives pages break down where avatar-first platforms and cinematic generators each make sense.

FAQ

How much source footage does a custom avatar need?

Five to eight minutes of varied speech is a practical minimum for a high-fidelity result. More important than length is variety: different sentence rhythms, numbers, questions, and at least one emotionally different read.

Can I use one avatar across several brands or clients?

Technically yes, visually risky. Different brands need different wardrobe, background, and grade. Build a separate look bible per client and re-render the presenter plate rather than recoloring the same footage.

What is the fastest way to fix bad lip sync?

Change the script before you change the tool. Shorten the sentence, spell out the number, and remove the consonant cluster that is causing the smear. If the problem persists across a stress test, the footage is the issue.

Should the avatar look directly at the camera?

For training and explainers, yes — direct address reads as instruction. For narrative or documentary-style pieces, a slight off-axis gaze is more natural.

How do I keep an avatar series from feeling repetitive?

Vary the visuals, not the face. New B-roll, new graphics, new pacing, and a consistent presenter is a formula audiences accept. A new face every episode destroys the recognition you paid for.

Do I still need a human presenter?

For anything where trust is the product — a founder's message, an apology, a sensitive announcement — yes. Audiences forgive a synthetic presenter in a how-to video and do not forgive one in a resignation letter.

Bring Your Next Presenter Video to Orelon

A custom avatar solves the face. It does not solve the ninety seconds around the face, and that is where most videos succeed or fail. Build the B-roll, the establishing shots, and the transitions that give your presenter somewhere to stand — start with the Orelon AI video generator, keep your visual language consistent with a saved prompt set, and treat the avatar as the narrator of a film rather than the film itself.