A
Claude Code logoClaude CodeSame character or product in any setting

Train a Flux LoRA of one character and reuse it across every scene

setuproll@setuproll
88.0Overall score

A training-plus-inference pipeline that fine-tunes a Flux LoRA on 15 to 25 photos of one face, mascot or product, then loads it in ComfyUI so the subject stays identical across poses and backgrounds. For game artists, brand mascots and creators who need a repeatable character, not a one-off.

88.0Score
1.9kVotes
5Components

Install this build

Export
terminal
git clone https://github.com/ostris/ai-toolkit && python run.py config/flux_lora.yaml

Components

Model

  • Flux.1 dev base
  • Custom Flux LoRA (your subject)

Stack

  • ComfyUI
  • ai-toolkit (Flux trainer)
  • CUDA GPU 24GB

Tools

  • LoraLoader node
  • Auto-caption with BLIP
  • 1500-2000 training steps

Workflow

  • Curate 15-25 clean images
  • Train LoRA at rank 16
  • Inject at strength 0.7-0.9

How it works

  • Collect varied, well-lit photos of the subject and auto-caption them
  • Train a Flux LoRA with ai-toolkit for around 1800 steps
  • Drop the LoraLoader node into your ComfyUI graph after the model loader
  • Prompt new scenes with the trigger word and the subject stays on-model

Summary

A training-plus-inference pipeline that fine-tunes a Flux LoRA on 15 to 25 photos of one face, mascot or product, then loads it in ComfyUI so the subject stays identical across poses and backgrounds. For game artists, brand mascots and creators who need a repeatable character, not a one-off.

88.0 score 1.9k votes

0 Reviews

Your rating
Sign in to post

Loading discussion...