1. Finetuned FLUX models are incredible, much higher quality than Stable Diffusion ones.

2. To create your own avatars, first collect around 10-15 images, a mix of headshots and full body portraits. Your dataset should look something like this. All of the images should be 1024x1024.

‍

3. Rename these images to “a_photo_of_TOK_1.png, a_photo_of_TOK_2.png…”
4. Then go to Replicate:

https://replicate.com/lucataco/ai-toolkit/train‍

5. Create a new model

6. Upload your images as a .zip file
7. Add your Huggingface token

(if you don’t have one create it here: https://huggingface.co/settings/tokens, add READ permissions)

8. Define resolution

9. Before hitting Create training, go to Huggingface and accept the T&C’s on the model page. https://huggingface.co/black-forest-labs/FLUX.1-dev
10. Wait for training (around 30 mins).
11. Once training is finished download the weights.
12. Upload it into a new Huggingface model repo (make it public)

13. It should look like this:

14. Go to this replicate page: https://replicate.com/lucataco/flux-dev-lora
15. Change hf lora:

16. Make sure you add “TOK” to your prompt.

‍