umbral rivet
I have been experimenting with training faces as LoRA's for SDXL and I am happy with the results.
I want to try some more examples that I don't have control over to get an idea what end users are able to provide.
If you have a friend/family you want to build a lora for, I would like to offer my services as an exercise. I only ask for feedback as you use the LoRA as to how well it recreates the likeness.
What I need:
- Preferred SDXL model to train against
- MIN 20-25 good quality images of face from various angles. Should be good lighting with all features exposed. No heavy shadows on face. Front, angled, side profiles are all good.
- NICE TO HAVE: 4-6 good quality half body images (not required for good results)
- BONUS: 4-6 good quality full body shots (not required for good results)
- a trigger concept (name of person or unique thing, ill recommend a trigger based on tokenization rules )
Guidelines:
- images should be as varied as possible for all details that should not be part of the lora. so anything you want the lora to learn should be the same in every image
- high quality mean at least 1024 x 1024 and preferably higher, in good sharpness/focus with good lighting and minimal shadow noise
- avoid images where the face varies greatly over time (e.g. dont include photos of someone at 20 years old and then photos of them at 40 years old
Terms:
- You must articulate reasonable ownership of images and relation to the face.
- I will use the images to train the lora and nothing else
- I will not share images created with the lora except with you or someone you identify
- I will create samples for whatever prompt or themes you supply
- I will provide recommendations for pos/neg prompt tokens to accentuate the quality of the lora
- You will provide constructive feedback as to the quality of the matching of the appearance
If you are interested, please DM me. I will select participants as my capacity permits.
warm zenith