How to Keep an AI Character Consistent Across Images

The process at a glance.

1. Choose and save one approved reference portrait.
2. Write an identity brief separate from clothing and setting.
3. Generate one focused change from the original reference.
4. Compare face and requested change before keeping the result.
5. Record model, prompt, source, and cost; retry from the original if identity drifts.

Separate identity from presentation

Write a short identity brief using features visible in the source: facial structure, eye spacing, nose, hairline, hairstyle, apparent age, and distinctive accessories. Put clothing, background, lighting, and pose in a separate section. A jacket is a useful continuity cue, but a matching jacket does not establish that two faces represent the same character. Change the identity brief only when you intentionally redesign the character.

Choose a portrait that can support the next scene

Use a sharp source with one fictional adult and a visible face. A tiny face in a wide shot gives you less detail to compare. A strongly turned head or dramatic shadow can make a later front-facing view harder to judge. Begin with a head-and-shoulders portrait, a natural expression, and unobstructed facial features. Keep the original file, rather than replacing it with your latest generation.

Start with a background-only experiment

Select the original as the reference. Keep facial features, hairstyle, expression, head angle, clothing, and crop explicit. Ask only for a different background. For example: keep the same adult character and portrait framing; change the background to an evening station concourse and adapt the lighting to that setting. Inspect the face before deciding whether the new backdrop is attractive. If identity changes, narrow the request and retry from the original.

Test the harder changes separately

Next compare a lighting change, then wardrobe, then expression, then a modest camera-angle change. Use the same original for every candidate. A full-body pose in a new environment changes several variables at once, so leave it until you understand the simpler edits. The generator page provides six copyable experiment prompts. They are a test plan, not a claim that the tool passed every test.

Review likeness before style

View the reference and candidate side by side at comparable face sizes. Check jaw, eye spacing, nose, hairline, and apparent age first. Then check whether the requested edit happened. Finally inspect hands, jewelry, clothing, and background artifacts. Record identity as pass, uncertain, or reject. Do not count a beautiful image as a successful character edit if the face has become a different person.

Record failed attempts and the real cost

A useful record includes the original filename, model, exact prompt, output filename, credit cost, identity result, requested-change result, and notes. Keep rejected attempts in the count. Three attempts at 2 credits each cost 6 credits, even if you keep only one image. The six-experiment plan costs 12 credits for one Raster attempt per experiment, or 72 credits with Raster Pro; retries add to that total. The three starter credits cover one Raster edit and one text generation.

Return to the original when drift accumulates

If an accepted result becomes the next reference, inspect it for small changes before continuing. Later edits can preserve and amplify those changes. When likeness drifts, branch from the original and reintroduce only the accepted presentation changes. If you need exactly identical face pixels across a sequence, use masking or compositing for those areas rather than relying on a generative edit.

Before you keep the result.

• Jaw, nose, eye spacing, and hairline still match the reference
• The character's apparent age and distinctive traits remain recognizable
• The requested change is visible
• Hands, accessories, and clothing survive a full-size inspection
• The original source, exact prompt, model, and credit cost are recorded

A few questions.