Character Consistency in Multi-Shot Image Generation

Notes from Krea 2 Turbo

These notes come from repeated multi-image character work on Krea 2 Turbo. I was feeding multiple prompts, one after another, over the API, like a story. Parts of each prompt stayed static: the style description, the character description, and the character reference image. I had 2–3 main characters I wanted to keep consistent while the environment, locations, and what they wore changed from scene to scene.

The goal was simple: keep the same person recognisable from shot to shot without locking the pose, clothes, lighting, or room into every frame.

Most failures were not model failures. They were prompt failures that became obvious once a sequence was generated instead of a single hero image.

Separate Permanent Identity from the Current Shot

Treat every character as two layers:

Identity layer (what never changes)

• Age range, sex/presentation, face structure, hair color and style, skin tone, permanent marks, distinctive body features.

• This stayed in a short, stable text description plus a clean reference image.

Shot layer (what changes every image)

• Clothes, accessories, pose, action, expression, gaze, room, lighting, and camera.

If something can change between images, it does not belong in the identity description. Putting a specific outfit or room into the permanent block was the fastest way to get the same pose and location forever.

Reference Images

A good identity reference is a clean face/head (or upper body) shot on a plain background. It should not contain a strong pose, room, or scenery.

This is the exact reference instruction that worked:

REFERENCE: use image1. Keep her face, hair, body, and clothes from image1. Ignore background, pose, and setting.

Without an explicit “ignore pose and setting” line, Krea 2 frequently copied the reference pose even when the rest of the prompt asked for something different.

For wide or distant shots where the face is small, the text identity alone was often enough and avoided pose leakage.

Every Prompt Must Be Stateless

Krea 2 does not remember the previous image. A later prompt cannot rely on “still,” “again,” “the same dress,” “continues,” “downstairs,” or “after the last shot.”

If a fact must be true in the current image, write it fully in that prompt:

• Who is present

• What they are wearing (full outfit)

• What they are holding

• Where they are

• What they are looking at

• How many hands are visible and what those hands are doing

This feels repetitive. It was also the difference between a coherent sequence and a collection of almost-the-same people.

Camera, Bodies, and Hands

Camera is empty space.
Do not write the camera as a person. Phrases like “someone over her shoulder,” “a figure at the edge,” or “leans in from the side” often created an extra person. Describe the lens position directly: height, angle, distance, and whose back or shoulder is in the foreground if any.

State orientation separately from camera height.
“Low angle” plus a readable face frequently collapsed to a waist-height shot. If you want the camera on the ground, prove it in the frame description (nearest sole large, sky above the head, etc.).

Count hands.
Two people = four hands unless some are out of frame. Uncounted hands became extra arms or extra people. Name who is holding what.

Exactly the number of people you intend.
Name them. “No extra people” was weaker than “exactly two people: [Name A] and [Name B].”

Gaze needs a real target in the frame.
“Looks around” or “examines the area” was weak. “Looks down at the tablet in her hands” or “watches the door on the left” gave the model something concrete.

Prompt Structure That Worked

Static parts (style, character description, reference) stayed the same across shots. The variable parts changed every scene. A typical structure looked like this:

STYLE
[style LoRA keywords and additional style description]

CHAR1 is PingPong
KEEP the same woman from image1, her name is PingPong.
REFERENCE: use image1. Keep her face, hair, body, and clothes from image1. Ignore background, pose, and setting.

SHOT: Medium side action shot.
CAMERA: Where we stand, where we look. Both eyes hidden. Does not face the viewer.
LOCATION: Place, ground, weather or interior.
ACTION: Caught in the instant of [verb].
GAZE: Eyes on a named thing in the frame, never the viewer.
HANDS: What both hands are doing.
MOTION: Body, boots, pack, debris, the beat.

Leave out sections that do not apply. Empty template language rarely helped.

Common Failure Modes and Fixes

Problem Usual Cause Fix

Same pose every shot - Reference winning - Add explicit “ignore pose and setting”

Wrong room or location drift - Location unnamed or too vague - Name the room and key objects every time

Extra person or arm - Camera written as a body, or uncounted hands - Camera = empty space; count hands

Object changes between shots - Prop not named or not held - Name the prop and say who is holding it

Person melts or becomes abstract - Words like shape, silhouette, blur, form - Describe a complete person with a readable face when the face matters

Camera stays at waist height - “Low angle” without proving the viewpoint - Describe what the camera sees from the intended height

Face stays frontal when you want profile - Restated smile or mouth, or “seen from under the chin” - Specify one side only; do not restate the front of the face

Style Notes

Style (medium, line weight, colour palette, film grain, etc.) belongs in a short style block, not mixed into character or action descriptions. Avoid parking generic ground, horizon, or sky language in the style block unless every image in the set needs it.

Practical Testing Approach

• Generate the identity reference first and lock it.

• Write each shot prompt as if the model has never seen the previous image.

• When a shot fails, change only the camera, action, or hand description and reroll that shot. Do not rewrite the entire sequence.

• Strong permanent features (large tattoos, cybernetic limbs, horns) can leak across characters when multiple references are active. Prefer putting those features in text and keeping references to face/head only.

These are observations from Krea 2 Turbo work only. Different models have different biases around anatomy, text, and composition, so the same phrasing may behave differently elsewhere.

The value here was in making the prompt constraints explicit so it was possible to see which failures were prompt issues and which were model issues on this specific system.

If you have tried similar multi-shot character work on Krea 2 (or elsewhere) and found different patterns that hold up, or counter-examples to any of the above, I would like to hear about it. The interesting part is what actually survives real multi-shot use rather than single-image demos.