Train Your Digital Avatar LoRA
Collect the right 15–30 training photos, pick a trigger word that won't collide with normal words, and train the one LoRA everything else in this course builds on.
Everything in this course hangs off one asset: a LoRA that knows your face. You train it once and reuse it forever.
The training set (this is 90% of the result):
- 15–30 real photos of you — selfies and photos other people took.
- Vary the angle (front, three-quarter, profile), the lighting (indoor, outdoor, dim), and the expression.
- No filters, no beauty mode, no sunglasses in most shots. The model learns what you feed it — feed it filtered skin and you'll never get realistic skin back out.
- You should be the only person in every frame.
The trigger word: pick something unique that isn't a common English word — your name works if it's distinctive, or an invented token like sirio2847. Every prompt in this course starts with it. If you pick a common word, the model's existing associations will fight your face.
Training: use Fal.ai's FLUX LoRA trainer (or any equivalent FLUX.1 LoRA training service), upload the set, set your trigger word, and start the run. Default settings are fine for a first pass.
Pro tip: if later generations look like a cousin of you rather than you, the fix is usually more varied training photos, not more training steps.