How to train your virtual model

With the pack's training set you create your own LoRA: a small file that teaches an image generator the model's face. Then you write a sentence and get new photos of the same person, in any scene, outfit or trade.

1. Prepare the set

Unzip the pack. The entrenamiento/ (training) folder has the photos at 2048 px and one .txt caption per photo. Every caption starts with the model's trigger word (for Lucía, lucia_fpn): that word "calls" the person. You don't need to change anything: the trainer resizes the photos itself.

2. Choose where to train

Online, no graphics card (easiest)

Services such as fal.ai, Replicate or Civitai train in 10 to 30 minutes and cost 1 to 3 dollars per run (October 2026 prices). Upload the zipped folder, type the trigger word and wait.

On your computer, free

With ai-toolkit (by Ostris) or OneTrainer. You need a 16 GB NVIDIA card for Flux (24 GB is better) or 12 GB for SDXL.

3. Recommended settings

Photos	All of the set (ideal: 15 to 30)
Resolution	1024 px
Rank	16 for Flux · 32 for SDXL · alpha equal to rank
Learning rate	0.0001 (1e-4)
Steps	1,500 to 2,000
Save	Every 250 steps

4. Check the result

Look at the samples every 250 to 500 steps. If they all look almost identical to the training photos, it went too far: use an earlier checkpoint. If the face doesn't match, train more steps or raise the rank to 32.

5. Use it in ComfyUI

Copy the .safetensors file into ComfyUI/models/loras/ and refresh the list.
Add a "Load LoRA" node between the model and the sampler, strength 0.8.
Put the trigger word at the start of the prompt and generate 4 to 6 photos to check.
If the face looks distorted, lower the strength to 0.6–0.7.

Online services work the same way: pick your LoRA, type the trigger word in the prompt and generate.

Sample prompts

photo of lucia_fpn, a pharmacist behind the counter of a bright pharmacy, natural light
photo of lucia_fpn, a real estate agent handing house keys to a young couple, sunny afternoon
photo of lucia_fpn, a yoga teacher in a sunny studio, smiling, wide shot
photo of lucia_fpn, a bakery owner holding a tray of fresh bread, warm light
photo of lucia_fpn, a dentist talking with a patient in a modern clinic, candid

Good practice

Check the license of the base model you use: some allow commercial use of the images but not of the model itself.
Label "AI-generated image" wherever the law or the platform requires it.
Don't use the model to impersonate a real person, or in sexual, political or minor-related content.

The LoRA you train is yours for your campaigns and your clients'; it may not be resold as a file.