👍 *** If you love it, like it! ***👍
workflow: https://civarchive.com/models/1088678
👑 Dominique Swain (Face/Off, movie, 1997) 🎬
About my celebrities loras
90% of the dataset used to build my loras only use head images. That really help the blend with other lora or model as there is no hands, feet, that may or will interfere in the final image render. When you get distorted hands with a person lora, it's because there is info on hands in the dataset used to train the lora, but that will not happen with my loras.
I've trained on Flux.1 Dev so other merged or trained checkpoint may not work well with my loras.
The drawback side of that is that the body may not be reflecting the reality. It may not be a drawback tho.
This is a lora for Flux.1 Dev. Work with other model but you must drop some simple bloc (good start 19-32).
Trained with ai-toolkit, so merging it is not easy.
To get the best result
Guidance: 2.2-3
Steps (dev): 30-40
daemon detailer (laying Sigma Sampler): factor: -0.02, start 0.06, end 0.7
Resolution: Upscale the latent by 1.25 or 1.5 you'll get awsome result. (take longer time but worth it)
Trigger word is (may work better in certain context): ohwx
Enjoy!
Description
V2, improved likeness
V1, initial release, not shared
FAQ
Comments (2)
i would say this is a really good lora, as other so many others you made, but i would also say i hate that so many other creators can make a lora without using trigger words, but people like you require one. is it like branding? is that why its important to you? keeping track of an arbitrary trigger word for every lora is something NO ONE wants to do
You can use the trigger word, or not, you wont see big difference. The big difference will be when the context of the prompt try to bring someone else known by the model. If you say: photo of Yoda standing with his light saber. you may not have any result from the lora. If you use the trigger word with that prompt, you may have a mix of both subject... I always use the trigger word, so when i got a mix result, i know my prompt was wrong.