Drag across · scroll
0% read
The Character Lock Runbook
FREEA Koda guide
Free guide: three ways to lock an AI character so the face stops changing, from a one-image character sheet to a trained LoRA you own.
Everything you need, in one place
The sheet
ChatGPT (GPT Image 2.0)
Turns one reference photo into a labelled turnaround board you attach to every future prompt. Free tier works, Thinking mode does not.
Create your accountThe identity
Higgsfield Soul ID
Trains a reusable character on your photos in about three minutes, then holds that face across styles and presets.
Train your Soul IDThe trainer
fal FLUX.2 [dev] Trainer
The pro lock. A real fine-tune on your subject, priced per step, with a weights file you download.
Open the trainerThe portrait route
fal FLUX LoRA Portrait Trainer
Cheaper face-specific training on FLUX.1 when you want a face LoRA, not a whole aesthetic.
Train the faceThe bridge
fal API key + client
For running training and inference from your terminal instead of clicking through the UI.
npm install --save @fal-ai/clientWire it- Time
- 20 min for the sheet, 3 min for Soul ID, 30 min for a LoRA
- Level
- Beginner to pro, in that order
- Tools
- ChatGPT, Higgsfield, fal.ai
- You end with
- One character whose face survives every model you move it to
the real problem
Your character is not inconsistent. It is being reinvented every generation.
Face drift is not bad luck. It is the model guessing who your subject is, from scratch, every single time you hit generate.
You wrote the same prompt twice and got two different people. Slightly different nose. Different jaw. Suddenly younger. That is not a prompt problem, and adding more adjectives will not fix it.
Character consistency is a model's ability to hold the same person's identity across generations: facial structure, proportions, skin tone, and hair. There is a real difference between a set of images that clearly show the same individual and a set of similar-but-different faces that do not quite match.
Here is the part nobody says out loud. A reference image anchors a single generation but drifts across many separate ones. So the fix is never "better reference". The fix is giving the model an identity it does not have to guess.
Stop prompting the face. Store the face.
There are exactly three storage formats, and they are not competitors. They are three tiers of the same decision.
Free preview
The rest of this guide is free.
Create a free account and keep reading right here. Takes 30 seconds, no card.
Create a free accountAlready have one? Log in