How to keep an AI character consistent in ComfyUI

A diffusion model does not remember a person. Every generation samples a new face that fits the prompt, so a new pose, lens or light produces a slightly different woman. A fixed seed only helps while nothing else changes.

A LoRA is a small add-on model trained on images of one subject. Loaded at a strength around 0.7 to 1.0 with its trigger word in the prompt, it pulls every generation toward the same face and body.

It costs a training run and a good dataset, but it is the method that holds up across hundreds of posts, new outfits and extreme angles.

IP-Adapter feeds a reference image into the model as a prompt. Face-specific variants and newer identity nodes like InstantID and PuLID lock the face harder while leaving composition free. No training is needed, which makes them ideal for building a LoRA dataset from a single reference.

The trade-off is control: strong face weights can stiffen expressions and copy the reference’s lighting.

In wide shots the face is only a few dozen pixels and drifts first. A face detailer pass crops the face, regenerates it at higher resolution with the LoRA or reference, and blends it back.

Most creators combine them.

1. Start with a face-reference workflow to make a varied dataset from one chosen portrait.
2. Train a character LoRA from the best 20 to 40 images.
3. Generate with the LoRA, and add a face detailer for full-body and group shots.

In Musetta a character is defined once and chosen by her face in every tool. The studio loads her LoRA and trigger word into whatever workflow you run, and every output records which version of her it used.

Face-reference and detailer workflows lean on custom nodes more than any other kind. When one will not load, the guide to missing ComfyUI nodes covers the fixes, and the ComfyUI alternatives page covers what to use if you would rather not run a node graph at all.