Photos to character · One-tap training
Create a consistent AI character from your photos
One-off prompts give you a different face every time. Training a small model on your own photos gives you a character that stays the same person in every image — and the model is a file you download and keep, not an account-locked preset.
What works today
Upload 8–40 photos of the character — a real person with permission, or a drawn or generated character — name it, and press Train. NiceVois trains the model on an isolated cloud GPU, generates new images of the character right on the page, and delivers the model file with everything needed to use it elsewhere. Your first model is free.
Varied photos make the likeness stronger — the trainer will tell you when the set is thin.
Train a character modelNext: the same character, in video
The open-source video ecosystem now trains character models too — the same photos-in, character-out idea, carried into generated video. NiceVois is opening this as its next workflow. If you want it, say so — the count of people on this page is literally how we decide.
Not live yet. Leave your email and you will be told the day it opens — nothing else is sent.
Under the hood
For the technically minded: the live trainer builds a rank-16 LoRA on Flex.1-alpha (Apache-2.0) via ai-toolkit, 250–2000 steps with identity subjects budgeted higher than styles, delivered as a portable .safetensors with its trigger word and exact training manifest — it loads in ComfyUI, Forge, or any FLUX-compatible runtime. The video step targets Wan 2.2 character LoRAs (the dual high-noise/low-noise pair trained together), which ai-toolkit already supports; the same dataset rules and the same consent-first contract apply. Only train people who gave you permission.