How to keep the same face in AI images without LoRA training

If you have ever tried to make an AI character for a series of posts, you know the problem. The first image looks great. The second one, with a new outfit, looks like her cousin. By the fifth image you have a whole family.

The usual advice is to train a LoRA. That works, but it means collecting 20 or more consistent images (which is the very thing you can’t make yet), captioning them, and running a training job. For most people that is a weekend gone before they post anything.

There is a simpler way, and it is how the workflows in the AI Character Kit are built. This guide explains the method so you can do it yourself in ComfyUI.

Why does the face drift when you reuse the same prompt?

Text to image models don’t remember anything. When you run the same prompt twice, the model starts from new random noise and draws a new person who happens to match the words. “24-year-old Indian woman with long wavy black hair” describes millions of faces. The model picks a different one every time, and the seed only fixes it as long as you change nothing else.

The moment you change “cafe” to “beach”, the composition shifts and the face shifts with it. That is why prompt-only consistency always breaks at the second image.

What is the alternative to training a LoRA?

Flip the process around. Instead of asking the model to describe the same person again, give it a picture of the person and ask it to change everything except the face.

That is what instruction-based image editing models do. You hand them an image and a sentence like “put her on a beach at golden hour”, and they redraw the scene while keeping the subject. The face comes from the pixels, not from luck.

The method has two steps:

1. Make one strong base image. This is the master copy of your character. In the kit this is workflow 01, using Krea 2 Turbo.
2. Edit that base image into every new scene. In the kit this is workflow 02, using Qwen-Image-Edit 2511 with a 4 step Lightning LoRA.

Every later image, and every video, starts from that one picture.

How do you make a base image worth copying?

Spend most of your time here. Every flaw in the base image gets copied into every scene, so a little care now saves hours later.

A prompt structure that works well with Krea 2:

[photo type] of [name], a [age]-year-old [ethnicity] [woman/man] with [hair],
[eyes], [one or two unique marks], [skin detail], wearing [simple outfit],
[place] in [light], [camera detail]

Here is the one used for Aanya, the fictional sample character in the kit:

candid iPhone photo of Aanya, a 24-year-old Indian woman with long wavy black hair,
warm brown eyes, a small gold nose stud, natural skin texture with visible pores,
light freckles, wearing a cream knit sweater, sitting by a cafe window in soft
morning light, shallow depth of field, slightly grainy

The settings used in the kit’s Krea 2 workflow:

Setting	Value
Model	Krea 2 Turbo fp8
Steps	8
CFG	1
Sampler / scheduler	euler / simple
Size	896 x 1344 (2:3 portrait)

On an RTX 3090 Ti this takes about 32 seconds per image once the model is loaded. Run it several times with random seeds and pick your favourite.

What makes a good base image?

• The face is clear and looks near the camera, with no hands or hair over it.
• The outfit and background are simple. You will change them later anyway.
• One or two unique marks: a nose stud, a mole, freckles, a hair ribbon. They help both you and the model notice drift.
• Even, soft light. Hard shadows on the face get copied into every scene.
• Waist up or upper body. In full-body shots the face is too small to copy well.

Avoid words like “perfect skin”, “flawless”, “8k” and “masterpiece”. They push the model toward a plastic, airbrushed look. “Natural skin texture with visible pores” and “slightly grainy” do the opposite.

Save the winner as character.png and copy it into ComfyUI/input/.

How do you put the same character in a new scene?

Load character.png into an image edit workflow and write one instruction. With Qwen-Image-Edit 2511, these settings give clean results in about 15 seconds on a 3090 Ti:

Setting	Value
Model	qwen_image_edit_2511_fp8mixed
LoRA	Qwen-Image-Edit-2511-Lightning-4steps-V1.0
Steps	4
CFG	1
Sampler / scheduler	euler / simple
Image size	scaled to about 1 megapixel

The models are free on the official Comfy-Org Qwen Image Edit page and the lightx2v Lightning LoRA page. If you want the full node by node setup, read the Qwen Image Edit 2511 beginner guide.

The scene prompt formula:

Show the same [woman/man] from the image, same face, same hair and [unique mark],
now [action] [place] [time/light] in [outfit], [camera style], natural skin texture

For example:

Show the same woman from the image, same face, same hair and same nose stud, now
walking on a beach at golden hour in a white linen dress, candid phone photo,
natural skin texture

Which rules keep the face locked?

These five habits make the biggest difference:

1. Always say “same face”. It sounds too obvious to matter, but it does.
2. Name the small details you want kept. The nose stud, the tattoo, the ribbon. Edit models drop small things unless you mention them.
3. Always start from the original base image. Never edit an edited image. Edits of edits drift a little each time, like a photocopy of a photocopy.
4. Make one big change per run. New place plus new outfit is fine. New place, outfit, hairstyle and art style at once is asking for drift.
5. Reroll before you rewrite. If the face drifts, press Run again with a new seed before you touch the prompt.

Can you change the art style and still keep the face?

Yes. The same edit workflow can turn your character into anime, a 3D cartoon or a cyberpunk poster. Keep “same face, same hair” in the prompt and describe the style:

Turn the woman into a 2D anime character, same face shape, same hair, same outfit
and same background, clean line art, soft cel shading

Stylised versions lose some identity by nature, so treat them as a fun extra rather than your main content.

What about very realistic, “camera roll” photos?

Ask for a specific phone and a specific everyday situation. Mess sells realism:

...taking a mirror selfie in her bedroom with a burgundy iPhone, flash off,
slightly messy room, natural skin with pores

When is a LoRA still worth it?

If you plan to make hundreds of images of one character, or need extreme close-ups where tiny facial details matter, a LoRA can lock things even tighter. The good news: the edit method gives you the training set. Twenty to thirty images from your edit workflow, all with the same face in different scenes and lighting, is exactly what a LoRA trainer wants.

For most creators, though, the base image plus edits is enough, and you can start posting today.

What comes after still images?

Once you have a set of scenes, each one can become a 5 second clip. The Wan 2.2 image to video guide covers the settings that fit on a 24 GB card, and the Wan Animate 2 character swap guide shows how to move your character with a driving video you filmed yourself.

If you would rather skip the wiring, the AI Character Kit has both of these workflows ready to run, with the settings above already set, plus 64 tested prompts and a sample character to test with.