Consistent Characters Across AI Images

Keeping one character looking the same across many images is the single most-asked problem in AI image generation, and the answer changed. The old Midjourney --cref trick is fading; today the reliable methods are reference images, image-to-image, and the strong character consistency in newer models like Nano Banana. This guide is the working playbook: the methods that actually hold a character together, in order of reliability, plus the details that make or break it.

Works in: Midjourney, GPT Image, Flux, Nano Banana, Stable Diffusion.

How to use this skill

When the user asks for an image prompt in this style, compose one using the formulas and vocabulary below. Fill each slot with a concrete choice, then return the finished prompt. Prefer naming a real camera/lens, a light, and a colour or film treatment, which is what makes the output read as a real photograph rather than an AI render.

Formulas

The consistency recipe

a fixed, detailed character description (reused verbatim) + a reference image + one changed thing (pose / scene / outfit)

Lock the description and the reference; change only one variable per image. That is how you get the same person in a new situation instead of a new person.

Worked prompt examples

_The reusable character sheet_

Character: a woman, late 20s, warm brown skin, short black curly hair, round gold glasses, small scar above left eyebrow, green utility jacket. [Reuse this block verbatim, then add the scene.] ...standing in a rainy market at night.

_Base reference_

Character sheet: a woman, late 20s, short black curly hair, round gold glasses, green utility jacket, neutral standing pose, plain background, consistent illustration style

_New scene, same character_

The same character (reuse the description + reference image), now sitting at a cafe table, warm light, same face, hair, glasses and jacket

_New pose, same character_

The same character (reuse the description + reference image), now walking through a rainy street at night, dynamic pose, same face, hair, glasses and jacket

Vocabulary

The methods, most reliable first

Reference images (most reliable)

• How: Generate one clean reference of the character, then attach it and ask for new poses/scenes. GPT Image and Nano Banana are strong at this; Midjourney uses image prompts + --cref; Stable Diffusion uses IP-Adapter / ControlNet.
• When: Any time you need more than one image of the same character. This is the default in 2026.
• Catch: Large changes (very different angle or lighting) still drift. Change one thing at a time.

A fixed character description

• How: Write a specific, detailed description (age, hair, distinctive features, clothing) and reuse it verbatim in every prompt.
• When: Always, alongside a reference. It is the cheap half of the recipe.
• Catch: Vague words drift. Name specific, repeatable features (a scar, a particular jacket), not moods.

Image-to-image / inpainting

• How: Take an existing image of the character and edit it (new background, new outfit) rather than generating from scratch.
• When: Small changes to an image you already like.
• Catch: Drifts if you push the change strength too high; keep edits modest.

Seed locking (Midjourney / Stable Diffusion)

• How: Fix the seed so the starting point is the same across generations.
• When: A helper on models that expose a seed; weakest method on its own.
• Catch: A same seed does not guarantee the same face if the prompt changes much. Use with a reference.

From God of Skills: a curated, hand-tested directory of AI skills, prompts, templates and image style guides. Source: https://godofskills.com/playbooks/ai-consistent-characters?ref=claude-skill