Consistent character prompt sheet
The hardest part of telling a story with AI images is keeping the same character looking like themselves across many shots. Models regenerate every pixel from scratch, so even a reworded prompt can change the face, hair or outfit. This sheet solves that by splitting your prompt into a locked block (identity that never changes) and a variable block (pose, expression and setting that change per image).
Why character consistency breaks — and what fixes it
Every time you call a diffusion model, it initialises from random noise and works its way to an image using only the text you provide as guidance. There is no memory of previous generations. If your description of the character changes even slightly between shots — a different word order, a synonym, an added adjective — the model has different signals to anchor from, and the result drifts: the hair lightens, the jacket changes colour, the face structure shifts.
The fix is mechanical: write the identity description once, perfectly, and then copy it identically into every prompt you make of that character. That block is your locked block. Only the scene-specific details — pose, lighting, expression, background — vary between shots. The character stays recognisable because the model always has the same anchor.
What belongs in the locked block versus the variable block
Locked block — put these in and never change them:
- Full name (if relevant to the style prompt)
- Age range and build (“athletic woman in her late 30s”)
- Specific face shape and features (“oval face, high cheekbones, warm brown eyes”)
- Precise hair description (“shoulder-length wavy chestnut hair, side-parted”)
- Skin tone description
- Signature clothing and accessories (“navy structured blazer, white blouse, small gold hoop earrings”)
- Art style and rendering keywords (“photorealistic, studio portrait lighting, 85mm”)
Variable block — change freely per shot:
- Pose (“standing”, “seated at a desk”, “walking”)
- Expression (“smiling”, “serious”, “surprised”)
- Camera angle (“close-up”, “three-quarter view”, “full body”)
- Lighting (“golden hour backlight”, “overcast natural”, “neon glow”)
- Setting (“city street”, “office interior”, “mountain path”)
- Action (“reading”, “pointing at a whiteboard”, “drinking coffee”)
How the tool works
You describe the character once in detail, and the tool freezes that into a reusable locked block. For each new image you only specify the variable tokens: pose, expression and scene. The tool outputs both the full prompt for the current shot and the isolated locked block so you can paste it into future prompts unchanged.
Tips for tighter consistency
- Specificity beats generality. “Brown hair” drifts; “shoulder-length wavy chestnut hair, side-parted” holds across generations. The model fills in any gap you leave, and it won’t fill it the same way twice.
- Never paraphrase the locked block. Copy it verbatim every time. Reordering or rewording words is the most common cause of drift, even when the meaning is identical.
- Pair with model reference features. Locked text works best in combination
with Midjourney
--cref(character reference), DALL·E reference images, or a trained Stable Diffusion LoRA of the character. Text anchors the description; reference images anchor the visual output. - Establish a canon shot first. Before generating a scene series, spend several generations on a neutral, well-lit portrait with no distracting background. Pick the best result as your canonical reference image and note the exact prompt. Use that prompt’s locked block everywhere else.
- Include style tokens in the locked block. If consistency extends to the artistic style (illustration, anime, photorealism), put those keywords in the locked block too — otherwise each shot may render in a subtly different style.