How to keep a character consistent in AI video: character sheets and storyboards for Veo

Veo does not understand “the same man from the previous scene”. Neither does Sora 2. Neither does Grok Imagine. But all three understand “45-year-old man, grey hair tied in a low bun, thick beige wool sweater, round glasses, sitting in a worn green armchair”. The difference between an AI video where the lead changes face every 8 seconds and one where they hold an identity is methodological, not technical. This guide is that method — character sheets, text storyboards and Sequences mode — run in batch with Grok AutoKuak and Flow AutoKuak.

Why the problem exists

Generative models have no memory between prompts. Every call is an independent event: the model gets your text plus any reference image you attach, produces an output, and forgets everything. It does not “remember” that five minutes ago you generated a specific character. If you want that character to look the same tomorrow, you have to say so, every time, with the same precision.

What happens if you ignore that: shot 1 shows a young man with short hair; shot 2 a young man with mid-length hair; shot 3 an older man with a beard; shot 4 a woman. A different universe every 8 seconds. Useless for narrative.

The three levers of consistency

Lever 1: an ultra-detailed text description

The character description in every prompt has to be identical, word for word, across all 8-12 prompts in the batch. If shot 1 says “45-year-old man with grey hair” and shot 2 says “older gentleman with greying hair”, the model reads two different people. The working rule is to copy and paste the same description, never to rewrite it.

The levels of detail that matter, ordered by impact on consistency:

1. Approximate age (not “young” or “old” — use ranges: “45-50 years old”).
2. Hair colour and style (“grey, tied in a low bun”, “short messy red hair”).
3. Specific wardrobe (garment type + colour + texture: “thick beige wool sweater”).
4. Distinctive accessories (“round thin-metal glasses”, “silver necklace”).
5. Ethnicity or skin tone if it is central to the identity.
6. Build (“slim”, “stocky”, “athletic”).
7. Unique details (a scar, a tattoo, a mole) if they are central.

A working description for multiple scenes ends up looking like this:

45-50 year old man, grey hair tied in a low bun, round thin-metal glasses, thick beige wool turtleneck sweater, slim build, introspective expression

That same 20-word string goes into all 8 storyboard prompts. What changes is the action and the context, not the character.

Lever 2: a visual reference frame

Text alone gets you to about 70 % consistency. The other 30 % comes from the model receiving a reference image of the character on every call. That image — the character sheet, or reference frame — anchors the visual identity.

The traditional, manual way: you generate the first image of the character, download it, and upload it in the next prompt as a reference. The batch way: you pass the character sheet (or the previous frame) automatically with each prompt, using Grok AutoKuak’s Sequences mode or Flow AutoKuak’s image-to-video mode.

Lever 3: Sequences mode automates delivery frame by frame

This is the piece that turns theoretical consistency into consistency you can actually run in batch. Without an extension, assigning a different image to each of 8 prompts by hand is tedious and error-prone. Sequences mode solves it programmatically: you load 8 images in order and the extension injects them one by one with each prompt.

Without that layer, the workflow degrades quickly and most creators end up giving up on visual consistency.

Building a character sheet from scratch

A typical character sheet has three views of the same character. How to produce it:

Step 1: define the character in text

Apply the seven attributes above. Rewrite until you have a description a human illustrator could use to sketch the character with no ambiguity. Example:

35-year-old woman, long wavy brown hair swept to one side, pale green eyes, fair skin with light freckles on the cheeks, short-sleeved floral dress in pastel blue and pink, slim build, warm slightly smiling expression, thin silver ring on the right hand

Step 2: generate three views from the same description at different angles

Three short prompts that share the character description but ask for different angles:

• View 1 — front portrait: “Front portrait of [full description], looking straight at camera, neutral grey studio background, soft natural lighting, realistic photographic style”.
• View 2 — profile: “Right side profile of [full description], looking left, same neutral grey background, same lighting, same photographic quality”.
• View 3 — medium-long shot: “Medium-long shot (waist up) of [full description], showing the whole floral dress, relaxed posture, same background and lighting”.

Generate all three in Grok Imagine with Grok AutoKuak (3 prompts in Text only mode, or Image + text if you already have a starting image as reference). Download the three resulting files. If one view comes out odd, regenerate just that one.

Step 3: check coherence across the three views

The three images are your character sheet. Put them side by side and check:

• Same skin tone?
• Same hair colour and length?
• Same wardrobe (same fabric, same colours)?
• Consistent build?

If there is drift (say the profile has darker hair), add detail to the description and regenerate. This phase rarely takes more than 15-20 minutes and saves hours later.

Applying the character sheet to a storyboard

With the sheet ready, you can produce the real storyboard. The key move is to build every prompt on the pattern: character description (identical) + scene description (varies).

Template:

[SHOT TYPE] of [FULL CHARACTER DESCRIPTION], [ACTION AND SPECIFIC CONTEXT FOR THIS SHOT], camera [MOVEMENT TYPE], lighting [DESCRIPTION], style [VISUAL REFERENCE]. Keep exactly the character from the attached reference image. Duration: 8s. 16:9.

Worked example (shot 3 of an 8-shot storyboard about a woman in a bookshop):

Medium-long shot of a 35-year-old woman, long wavy brown hair swept to one side, pale green eyes, fair skin with light freckles on the cheeks, short-sleeved floral dress in pastel blue and pink, slim build, warm slightly smiling expression, thin silver ring on the right hand, holding an open book near her face as she reads with concentration, camera slow dolly-in from medium-long to close-up, soft natural light coming through a side window in warm golden tones, intimate cinematic style like Lost in Translation with subtle grain. Keep exactly the character from the attached reference image. Duration: 8s. 16:9.

Half the prompt is the character description, copied verbatim. The other half changes from shot to shot. For shot 4 you change the action (“takes a book down from a high shelf”) and the camera (“wide shot from a distance”), but the character description stays untouched.

The full batch workflow

1. Generate the character sheet (3 views) with Grok AutoKuak. Time: 15-20 min, 3 prompts.
2. Write an 8-shot text storyboard in Google Docs, copying the character description verbatim into each one. Time: 25-30 min unaided, or 10 min using ChatGPT AutoKuak to pull out the shots.
3. Produce the 8 static frames with Grok AutoKuak in Sequences mode, assigning a different character-sheet view to each prompt in a logical order (view 1 → shots 1, 4; view 2 → shots 2, 5; view 3 → shots 3, 6, 7, 8). Time: 15-20 min, 8 prompts.
4. Turn each static frame into a Veo clip with Flow AutoKuak in image-to-video mode, using the frame as reference plus the shot’s text prompt. Time: 30-40 min, 8 Veo prompts. Prompt detail in cinematic Veo 3.1 prompts.
5. Edit the 8 clips in CapCut or DaVinci with audio and transitions. Time: 30-45 min.

Total pipeline: about 2 hours for a 60s Short with a fully consistent character, against 5-6 hours by hand without a character sheet.

---

Common mistakes that break consistency

• Changing the character description between prompts. Even swapping “young” for “young and attractive” in one shot gets read as a variation. Always paste it exactly.
• Not anchoring the reference image in the text. Add explicitly to every prompt: “Keep exactly the character from the attached reference image”. Without it, the model sometimes ignores the reference.
• Mixing visual styles between shots. If shot 1 is “realistic 35mm” and shot 2 is “anime”, character consistency collapses even with an identical description. Keep the visual style constant.
• Skipping the character sheet. “I generate shot 1 and use that frame as reference for shot 2” works on the first iteration but accumulates drift. An independent sheet with three views is structurally sounder.
• Drastic lighting changes. If shot 3 is cyan neon and shot 4 is warm gold, the character changes enough to read as someone else. Keep lighting shifts gentle between consecutive shots.

When perfect consistency is not needed

Being straight about it: not every video needs a consistent character. Cases where you can skip the sheet:

• Tutorials with no visible presenter (voiceover plus visual assets).
• “Types of people doing X” compilations where each shot deliberately shows someone different.
• Short inserts of landscapes, products or nature with no human lead.
• Memes and absurdist content where the inconsistency is part of the joke.

Where it is critical: narrative storytelling, a recurring AI presenter for the channel, and campaign visual branding.

---

Frequently asked questions

How many views does a character sheet need at minimum?

Three is the solid minimum (front, profile, medium-long). For very detailed characters you can add more: three-quarter, back, specific expressions (laughing, surprised). More than six is overkill for video up to 60s.

Does it work with non-human characters (animals, fantasy creatures)?

Yes, identically. You describe the key attributes (species, colour, size, distinctive accessories, typical expression) and run the same workflow. For fantasy creatures (dragons, robots) the description may need more technical detail (materials, design style).

Grok Imagine or Veo for the character sheet?

Grok Imagine for static images (cheaper, faster, good enough). Veo to check how the character looks in motion if your narrative needs it. Typical flow: sheet in Grok, validation in Veo Fast (20 credits) with one test shot before committing Veo standard to all 8.

Can I keep TWO characters consistent in the same scene?

Yes, with more effort. You need separate character sheets for each plus a double description in every prompt (character A + character B). Veo handles two better than three. More than three at once nearly always breaks.

Do I need Pro to produce storyboards with a consistent character?

Not for an occasional storyboard (it fits inside Free’s 30 prompts per 24 h on Grok plus 30 on Flow). Pro makes sense if you produce 2+ storyboards a day, or if you want to iterate aggressively without limits. Detail in AutoKuak Free vs Pro.

Does this method work for Sora 2 as well as Veo?

Yes, almost identically. Sora 2 accepts image-to-video with a visual reference and respects detailed text descriptions. The prompt structure barely changes (Sora 2 tolerates longer prompts). Batch it through ChatGPT AutoKuak.

---

Start with a single character

Recommended first exercise: define one character (any of them — the lead for a test Short), generate their three-view character sheet, write a four-shot text storyboard with the description copied verbatim into every prompt, and produce the four frames with Grok AutoKuak in Sequences mode. Put the four frames side by side: if the character reads instantly as the same person in all four, you have the technique. If not, keep iterating the text description until it does. Once four shots behave, scaling to 8 or 12 is trivial. To go deeper on each piece: Sequences mode in Grok is the operational lever. Cinematic Veo 3.1 prompts covers the camera language. The script-to-Veo pipeline connects it all end to end.