If the person changes between images, the cause is almost always a description that changed with the mood. A blurb rewritten as "a young professional" becomes a different young professional each time you run it.
Below is the identity block, the settings that hold a set together, and the drift checks to run before you render twelve images instead of four.
What an ai character consistency prompt actually controls
A model has no memory of your character between renders unless something carries the description. Two things carry it: a repeated text block describing the same person in the same words, and a reference image or seed where the tool supports one.
Text alone gets you close but not identical, especially across lighting and angle changes. Reference based tools get you closer, and how close depends on the tool and the version you are using. Vendor claims about how many characters a model can hold at once are marketing statements with their own test conditions, so treat them as something to verify on your own images rather than as a specification.
Two honest expectations to start from. Expect the same person, not the same pixels. And expect drift in areas you never describe, which is where the block below spends its lines.
Write the identity block once, then paste it every time
Fix the description before you render anything. Twenty five to forty words, physical and specific, no adjectives that could point in two directions.
Identity block: [age range] person, [face shape and features as you would describe them to a stranger], [hair length, colour, style, and whether it is tied], [build or frame], [skin tone as you would describe it], [one or two features that identify them, such as a small scar or a habit of wearing one earring]. Age them as [age]**, not as a vague range.
Fill in these details
- Face: the features you would name if someone asked what they look like, no mood words.
- Hair: length, colour and whether it moves in the shot.
- Identifying feature: one or two things that travel with the character across every frame.
- Wardrobe: whether the set is one outfit, or a fixed outfit plus a stated change rule.
- Do not include: the words you have been using that point at nothing, such as beautiful, attractive or cinematic.
Keep the block byte identical in every prompt. Small rewording is the most common cause of a set that looks related rather than identical.
Anchor the set: settings you never vary
Identity covers the person. The set holds together because the framing conventions are fixed too.
| Element | Fix it across the set | Vary it per frame |
|---|---|---|
| Lighting | Direction, hardness, time of day | Nothing after the first two test frames |
| Lens and height | Focal length and camera height | Angle only, and only where the story needs it |
| Colour grade | Warm, cool or neutral, stated once | Nothing |
| Background | One described location | Distance to the subject |
| Wardrobe | Same garments, same colours | Pose and expression |
Lighting is the one creators vary by accident, because a mood change feels like a creative choice. It is also the fastest route to a different-looking person, since the whole tonal map of a face changes with the light.
Worked example: one character across six frames
A short story needs the same person in a kitchen, a car park, and a doorway at night. The identity block stays fixed, and the settings table does the rest.
| Frame | Scene | Varies | Held fixed |
|---|---|---|---|
| 1 | Kitchen at 8am, reaching for a mug | Pose | Identity block, lens, light direction |
| 2 | Same kitchen, seated at the table | Framing distance | Identity block, lens, light direction |
| 3 | Car park, midday | Location only | Identity block, lens, grade |
| 4 | Car park, walking toward camera | Pose only | Identity block, lens, grade |
| 5 | Doorway at night, interior light behind | Location only | Identity block, lens, grade |
| 6 | Same doorway, closer | Framing distance | Identity block, lens, grade |
Render frames 1 and 2 first and compare them side by side. If the face already differs, the identity block is too loose or the light moved. Fix that before rendering five more frames.
How to check drift before you render the set
1. Compare the two test frames at the same size
Open frames 1 and 2 next to each other at identical scale. Look at the jaw, the eye spacing, the hairline and the nose. Those four carry most facial identity, and they are where drift shows first.
2. Check the parts you never described
Hands, teeth, freckles, earrings and clothing detail are where a model fills gaps differently each time. If those matter to your set, name them in the identity block or accept them as variable.
3. Write down what changed between two good frames
When two frames do match, note which wording and which settings got you there. That note is your template, and it is worth more than the prompts themselves.
Common consistency mistakes
- Rewriting the description for mood. Feeling is set by light and framing, not by adjectives in the identity block.
- Adding detail to the identity over time. New lines change the character. Add details once, at the start, and keep them.
- Changing the lens between frames. A wide lens and a long lens change the face, not just the background.
- Trusting the first two frames and stopping. Check frame three before you commit to a set of twelve.
- Assuming a tool setting stays fixed. If you rely on a seed or reference strength, confirm what the current version actually does before you depend on it.
Frequently asked questions
How do I keep the same character across multiple images?
Write one identity block in fixed physical terms, paste it unchanged into every prompt, and hold the lighting, lens and grade constant. Where your tool supports a reference image or seed, use it consistently and describe its strength in your own notes.
Does the same prompt guarantee the same face?
No. It produces the same description interpreted the same way, which is close but not identical. Real consistency comes from repeating the description and using reference based features where your tool supports them.
How many characters can one set hold?
That depends on the tool and the version, and published figures vary by test conditions. Render a small test set with your own characters first and see where the likeness breaks before committing to a large set.
Lock one character, then expand
Write the identity block, render two frames, compare them, and only then build the set. For the lighting and framing decisions that hold a set together, browse the photography and visual storytelling prompts.