AI Image & Video

You Locked the Character but the AI Draws Someone Else: A Three-Link Self-Check for Identity Consistency

Identity consistency is not a switch. It is a chain of three links, and all of them must hold. When it breaks, find which link failed before you swap models.

10/09/2026 ｜ 4 min read ｜ Works for any flow that locks a character with a reference image — about a 15-minute read

The one idea that matters

Identity consistency is a chain of three links, not a switch. The links are AND, not OR: if one is misaligned, the other two being perfect won’t save you. That’s why swapping models usually doesn’t help: you were probably fixing the wrong place.

The three links: ① the reference image itself → ② the instruction to the generation or edit model → ③ the AI that writes your prompts.

Check backward, it’s fastest

Step 1　Check the reference image. This is the most common trap. Open the image you used to lock the character and ask:

• Is it a frontal view where the whole face is visible?
• Is the face sharp, not blurred or covered?
• Is it a clean portrait with the person as the subject, or a scene where the person is tiny?

Back views, heavy profiles, blurred faces, or a person occupying a tiny corner: replace it. The model copies faces by looking at them; with no clear face, it invents one, differently each time. I really did this once: I used a distant back view of a samurai on a castle wall as the reference. The result was a different person every time, often a woman.

• Good reference: frontal or three-quarter view, upper body, clear face, simple background.
• Unusable: back view, wide shot, blurred face, group shot, artistic framing where the face can’t be seen.

Step 2　Check whether you told the generation model to hold the identity. Edit-style models that generate from a reference (Qwen-Image-Edit is one example) will repaint the whole image, people included, if you give them only a description of the new scene. Add an explicit identity-preservation line:

Keep the same person as in the reference image: face, features, gender, hairstyle and clothing all unchanged. Change only the scene and the action.

Then add a negative prompt to push the drift directions away:

different person, different face, gender swap, wrong gender, changed identity

Step 3　Check the AI that writes your prompts. This is the most hidden link. If your flow first uses an AI (ChatGPT, Claude, a local LLM) to turn the script into image prompts, that AI doesn’t know you locked a character elsewhere. It invents a lead and easily writes your male lead as “a long-haired woman.” The prompt says female, the reference is male, the two fight, and you get the woman.

Fix: tell it the character sheet before it writes, with a hard rule. Copy this:

This film has locked characters. Their looks are fixed as below, and in every shot where they appear you must depict the same person, never changing gender, age, hairstyle or appearance: Male lead: 30-year-old man, short black hair, square face, navy jacket. Now turn each line of narration below into one image prompt (describe the picture only, no dialogue text): …

Step 4　Generate one image after all three are aligned. Check one thing only: is the face and gender right? If yes, repeat for the other shots. If not, go back to Step 1 and look at each link again.

Traps you will hit

Trap 1　Using a scene image as the reference. You feel it looks like the character, but the model needs a face, not a mood. The smaller, more sideways or blurrier the person, the worse the lock. Better to generate a clean frontal headshot as the reference.

Trap 2　Assuming a strong model locks faces automatically. A text-only image model and a reference-taking edit model are different things; the former never looks at your reference. Confirm you are on the path that reads reference images.

Trap 3　Writing the prompt and the reference separately. You lock the lead in one place and let a different AI write prompts freely in another. Two parts that never talked will clash.

Update (added in October)

For long-form work, my main local multi-reference image-to-video flow is now MiniMax H3’s Ref2VA (character sheets plus scene images as references), no longer the Qwen-Image-Edit path in this checklist. The engine changed, but the three links still hold: the reference must show the face, you must say clearly who is who, and the AI writing upstream prompts must know who is locked.

When you’re stuck, ask an AI like this

I’m new to this, using <your tool, e.g. ComfyUI / an AI image app> to make AI video or comics. I want the same character to look identical across images. Here’s my situation: <e.g. I locked the character but the gender changed>. Please explain step by step in beginner-friendly terms: 1) what a good reference image looks like, 2) which feature or node in my tool actually “reads the reference image,” 3) how to write the prompt so it doesn’t override the person in the reference. If you need screenshots or settings, tell me exactly where to look.

In one sentence

When the person comes out wrong, don’t swap models first. Check backward through the three links: a frontal reference → an instruction that holds identity → a prompt-writing AI that knows who the lead is.