AI Character Video Workflow: From Character Sheet to Launch Video

An AI character video workflow that holds up runs in five stages: write a character sheet, approve a set of reference stills, make short silent clips, add talking clips in a fixed voice, then cut a promo or launch video from the approved pieces. Each stage has a gate. A drifted face caught at the stills stage costs a few regenerations; caught in the launch cut, it costs the whole edit. Most of the work goes into the sheet and the reference set, not the video.

This is for anyone running a recurring character, whether a brand mascot, a host or a persona, across photos, clips and launches. Oakgen is our product, and its AI Influencers feature handles several of these stages.

The workflow at a glance

The scoring itself (traits, 0-2 points, blind checks) lives in our AI influencer drift test. Use that rubric at every gate.

Write the character sheet before you generate anything

The usual order is backwards: people generate a face they like, then try to describe it. The description ends up chasing whatever the model did, including its mistakes. Write the sheet first and the first generation has something to be wrong against.

A character sheet separates what may never change from what should change. For the fixed look (face, marks, hair, wardrobe signature, build), use the character bible template and shot pack in our consistent AI characters for ads and UGC workflow; the AI character consistency guide covers the same ground. Video needs a few more rows, because motion and sound add new ways to break the character.

Fill this in before opening any tool. The example column is an invented host called Noor, made up for this post; she's not a real person or a customer.

Keep the don't list short and checkable. Hats and sunglasses top it for a reason: they cover the hairline and eyes, which are what people recognise first. A shot where those are hidden can't be checked, so it can't be approved.

The fixed look, rewritten as one or two plain sentences, becomes the identity line you'll reuse everywhere. The variable rows become the menu you pick from per shot.

Stage 1: stills and the reference set

Everything downstream works from these images, so be fussy here. You want a reference set covering the angles your videos will need: a clear front portrait, three-quarter views both ways, profiles, and full-body shots from the front and side. If you plan walking shots, the full-body references matter as much as the face.

In Oakgen's AI Influencers, you either describe the character or start from a photo (the app asks you to "Only use people who agreed"). Describing sends the text to several image models at once (Nano Banana 2.1, Nano Banana Pro and GPT Image 2 as of October 2026), and you pick a cast take, ask for more like one, or recast. Next comes Refine: you type a change ("a bit older", "shorter hair") and compare versions on a before/after slider. Approving a version starts a 12-shot photoshoot made from that portrait: eight face shots (front, smiling, three-quarters, profiles, from below and above) and four body shots (front, three-quarter, from behind, seated). Then you choose a voice.

Gate 1: score the whole set against your sheet, not against the portrait. The photoshoot shots are generated from one image by one model, so if the portrait has the mole on the wrong side, all eleven other shots inherit it. Use the drift test's rubric on every reference. The drift test scores tools against your own reference set and never against a tool's photoshoot, because a photoshoot is an output. That still applies here: the photoshoot becomes your working set only after every shot passes against the sheet (and against your photo, if you started from one). Look hardest at:

• Marks on the correct side in both profiles.
• The wardrobe signature present in the body shots.
• Proportions in the full-body shots, which models drift more than faces.

If a shot fails, redo it or replace it with your own photo; the photo manager on the character page has both, plus "set as main photo". Don't keep a weak shot because the other eleven are good. As the next section explains, the app picks which shots to send based on your scene words, so a bad body shot will quietly show up in every walking clip.

Outside Oakgen the gate is the same: decide on the reference set, store it somewhere fixed, and always reference the originals, never last week's best output.

Stage 2: short silent clips

Short clips are where you find out whether the references hold in motion. Keep them boring on purpose: one action, one setting, one camera move, the shortest duration the model offers, no speech. A clip that does four things can't tell you which one broke the face.

In Oakgen, the character page has three buttons (Photo, Video, Talking video) and opens Create with your character mentioned by name, where you pick the model and see the price. A few rules decide what happens next, and they're worth knowing before you write prompts:

• Only models that accept reference images can use the character. A plain text-to-image or text-to-video model can't take one, and the app explains why rather than silently dropping the face.
• Only a few shots go with each generation. Up to four by default and up to six on two Seedance models, never more than the model accepts. Hovering the character chip in Create shows which shots are being sent and how many were held back.
• Your scene words choose the shots. Words like walking, full body, standing, outfit or street put the body shots first. Words like says, close-up, to camera, interview, podcast or vlog put face shots first. Everything else gets a default mix led by the front portrait and three-quarters.

Gate 2: for each clip, check the first, middle and last frames plus the worst frame you can find, using the drift test's five-clip check as the scoring sheet. Drift in motion often lasts a few frames during a head turn, and those are exactly the frames people screenshot. Keep passing clips in a folder named for the character version (more on versions below) and delete the rest so nobody cuts them in by accident.

For model-specific advice on multi-shot consistency, see the Seedance 2.5 character consistency guide.

A prompt skeleton for Oakgen

When you mention a character, Oakgen writes the identity part of the prompt for you. Your text goes first, with @Noor replaced by her name and a reference tag such as "Noor (@Image1)". Below it the app adds her saved description and an instruction to preserve her appearance from the reference images it sent. So your job is the scene, and only the scene.

@Name [one action, one verb]. [Setting]. [Shot size], [camera move or angle].
[Light]. [Wardrobe from the rotation, only if not the default].
[Talking video only: says "one short line".]

Two worked lines for the invented Noor:

@Noor looks up from her laptop and raises an eyebrow. Home desk, morning.
Chest-up close-up, static camera. Soft window light.

@Noor walks out of a glass office door onto a city street and exhales.
Waist-up tracking shot. Overcast daylight. Denim jacket and mustard scarf.

Notice what isn't there: no face shape, no hair, no mole. Restating them invites a second, slightly different description to fight the saved one ("bob" in one, "short hair" in the other). If something about her looks wrong, fix the saved description or the references, not the prompt.

Also notice the word choices. The second prompt says "walks" and "street", which tells the app to send body shots; that's what you want for a tracking shot. Put "walking" into a talking close-up and, with the default four shots, two face references are swapped for body references at the moment the face matters most (movement words win over talking words).

Stage 3: AI character talking video

Talking clips are the hardest stage. Treat them as a few moments inside a cut.

In Oakgen, a talking video happens in two steps when you put a line in quotes. First the line is spoken in the character's saved ElevenLabs voice. Then a video model performs it, with that recording sent as reference audio and the prompt asking for lips synced to it. Both steps show their price before you run them. There's a toggle for the character's exact voice; turned off, the model invents a voice.

The limits matter here:

• As of October 2026, three video models can take the spoken line as reference audio: Seedance 2.0, Seedance 2.5 Reference and MiniMax H3 Max (reference-to-video). Pick another model and the app warns that it can't take the character's voice and offers to switch to one that can.
• The app's own guidance says audio reference support "does not guarantee exact speech or lip-sync". Plan for rerolls.
• The saved voice is a voice from the ElevenLabs library, chosen during creation and stored with a short voice sample of the character introducing themselves. That sample is your reference for checking every later line.

Gate 3: listen to the line next to the saved voice sample. Is it the same speaker? Then watch the clip muted: lips should close on p, b and m, and the mouth shouldn't keep moving after the line ends. Watch it once more with sound.

What tends to help (our reasoning, not a measured result): one or two sentences per clip, a chest-up frame, the character mostly still, and no hand near the mouth. A walking-and-talking street shot is the harder version of everything.

A reliable fallback is to skip lip-sync for most of the cut. Generate the line as plain speech in the saved voice, lay it over a silent clip of the character reacting or working, and keep on-camera speech for the one or two moments that need it. Our reasoning: voiceover over b-roll can't go out of sync. Keep this clean speech file either way; laid under the matching clip in the edit, it keeps every line at the same loudness and tone.

If you only need a talking head from a single image, without a recurring character behind it, see AI talking avatar videos.

Stage 4: the promo or launch cut

By now you have a folder of approved stills, silent clips and voice lines. There are two routes from here.

Route A: cut it in an editor. Download the approved pieces and edit them with your product screens, type and music in any timeline editor. This is the most flexible route and the one most launch cuts will take. If the rest of the launch video is a type-led motion graphic, our guide to making a product launch video in one evening covers the script and timing; the character clips slot in as the hook and the sign-off.

If you plan to use Oakgen Motion for that type-led part, its "Describe your video" path builds code-made motion graphics from type, your screens and product shots, voice, captions and music; "Start from a look's video" keeps a sample's script, has no voiceover and uses only your first screen. Neither contains AI footage, so neither takes character clips. Our reading of the upload step: a still of the character can go in as one of your images, but it would sit in frame like a screenshot, not move or speak. Make the Motion video, then join it to your character clips in an editor.

Route B: Motion's Remake path. Upload a TikTok or Reel and Motion keeps its cuts and beat while casting a character you made in place of the person, with a 9:16 output. The details, including sound and pricing, are in remaking a TikTok edit with your own people.

A launch teaser shot list

This is an invented example: Noor (made up for this post) teasing the launch of Tallybird, the made-up invoicing app from our launch-video guide. 9:16, about 22 seconds. The point is the "comes from" column: every shot traces back to an approved stage.

Two talking moments at most, and one of them (shot 6) is voiceover, not lip-sync. That's deliberate.

Gate 4: before export, confirm every character shot in the timeline came from the current character version (never mix versions in one cut), and watch the cut muted on a phone with captions on.

Sound and voice handling

One voice, for good. Voice is part of identity. Changing it is like changing the face: a new version, logged.

Use the right audio models. In Oakgen, the saved voice only works with the ElevenLabs speech models. Music models can't keep it either: the app says a voice description "can guide style, but does not preserve the same singer". A singing mascot is a separate problem from this workflow.

Music goes in last. Add one licensed track in the edit and duck it under speech.

Caption every spoken line. Feeds often start muted, and captions also cover lines where lip-sync is borderline.

Version the character like a product

Log every change, or two versions of Noor end up in one ad.

Rewrite the identity line after every Refine. When you approve a refined portrait, Oakgen appends an "Updated: ..." sentence listing your refine requests to the saved description, which the character page labels "Identity · used in every prompt". That keeps prompts in step with the photos, but a string like "Updated: a bit older; shorter hair" is vague. Edit it into clean sheet language ("early 30s; chin-length bob") and update the sheet to match.

Re-check references after a change. Redoing the main photo doesn't repaint clips you've already made, and a new main photo with old body shots is a mixed reference set. Redo the shots that no longer match, then re-run Gate 1.

Keep a short change log.

The character page's "Made with" list shows what you've generated with the character, which helps fill the last column.

Track cost and time per usable asset

Rerolls, failed talking clips and redone references are where the money goes, so count spend per approved asset, not per generation. Keep the character setup on its own line. Our AI influencer cost per post breakdown has the ledger and the maths. In Oakgen, prices show in Create before each generation, and a talking clip is priced as two steps (the speech and the video). Plans are on the pricing page.

Mistakes that break a recurring AI character

• Generating before writing the sheet. The description ends up describing the model's first guess.
• Approving the photoshoot without scoring it. One wrong-side mole in the portrait becomes twelve.
• Restating the face in every prompt. Two descriptions drift apart, and the model splits the difference.
• Feeding outputs back as references. Last week's best clip is already slightly off; return to the approved set.
• Making talking video the backbone. Lip-sync is the weakest link; use it for a line or two and carry the rest with voiceover.
• Long spoken lines. Split anything over two sentences into separate clips.
• Mixing character versions in one cut. Check the folder names before export.
• Skipping the label. A photoreal recurring character reads as a real person. Read AI influencer disclosure rules by platform before the first post, and put the disclosure line from your sheet on screen.

Where Oakgen fits, and where it doesn't

Oakgen is our product, so weigh this accordingly. AI Influencers covers stages 1 to 3 in one place, and it needs a paid plan. Its own guidance says references help but "poses, scenes and model choice can change the result", so score each model you use. The limits that shape this workflow are in the stages above: reference-capable models only, a few shots per generation, three talking models with no lip-sync guarantee, and a Motion that only casts characters on the Remake path. Motion doesn't post for you either; you download and post yourself.

You don't need any of this for a one-off mascot illustration or a single ad. The image generator or AI video generator is simpler there. Use it if the character appears in more than a handful of posts.

Frequently asked questions

What should a character sheet for an AI character include?

The traits that never change (face, marks and which side they're on, hair, a signature piece of wardrobe, build), the traits you allow to change, the voice, and a short do and don't list. Keep it specific enough that someone else could check an image against it.

Should I make videos before I have a reference set?

No. Approve a set of stills from several angles first, including full-body shots. Every reference-capable video model works from those references, so a weak set shows up in every clip.

Do I need to describe my character's face in every video prompt?

Not if your tool adds the identity description for you, as Oakgen does. Describe the scene, the action, the camera and the light, and keep the face out of it so the two descriptions don't disagree.

Why does my AI character's lip-sync look wrong?

Long lines, fast speech and wide or moving shots make it harder. Keep each spoken line to one or two sentences, frame the face from the chest up, and check the clip muted for mouth closures on p, b and m.

Can I use the same AI character in a promo or launch video?

Yes, usually by cutting your approved clips and stills together in an editor with your product screens and type. In Oakgen, the Motion Remake path can also cast a character you made into an uploaded TikTok or Reel edit.

What do I do after changing my character's look?

Rewrite the identity description to match the new photos, redo any reference shots that no longer match, log the change with a date, and don't mix clips from before and after the change in one cut.