Kinda, but likeness isn't as good or as flexibile on small LoRAs. This LoRA was trained by first training a checkpoint then ripping off the changed data into the LoRA you see here. I've tried making good 18mb but they really don't hold a candle. I have one 18mb on this account that was a dress LoRA, something simple and vague. Not a realistic character that doesn't butt chin and has almost a perfect likeness (as good as flux can produce). Besides the true ultra checkpoint rip worked best (640dim) and I'm not posting that here, that was a 5.7gb file :)

@LOREVER Penses-tu que je n'ai pas fait du bon travail sur les légendes avant de laisser le modèle s'entraîner pendant plus de 24 heures ?

@LOREVER I'm trying to tell you that I have already deeply explored this and this current iteration is the absolute best quality I can achieve with Flux. I am not using the civitai.com trainer and do not want to go any smaller because they fall apart in likeness and in realism. I don't think you understand why or what I'm doing but I can reiterate that this is training directly onto the flux checkpoint aka I have a full 23.7gb file that I pull this model from. It is a completely different process than a normal lora which I gave up making due to its obvious bad likenesses and inflexibility. You can force a close likeness with a basic lora but when it finally trains to the point where the butt-chin disappears it is way way less flexible than this current iteration which also does not butt-chin, has an excellent representation of the character across all sorts of picture types and is also incredibly flexible as you can see in the images posted. If you can you should compare some of your bests with this or my others so that you can see the direct difference between the direct lora training and my method/size. This is still not perfect but I think it as close to perfection as Flux can get with current tools. Lady Gaga was a 25 image model (lots for flux) they were the absolute best I could find, no weird people in them or strange hands or bad film grain or noise, or dis-likeness of the character, I caption with cogVLM2 then go back through my self to edit and make them as good as possible. Then the model bakes for about 24 hours on my pc and I retest using my original captions to see how close it is to original. After it looks finished then I start working on the variations. This is the basic run down not talking about settings but if you would like to see some of them I did a post on reddit here.

https://www.reddit.com/r/StableDiffusion/comments/1gtpnz4/kohya_ss_flux_finetuning_offload_config_free/

french translation for fun, how bad is claude at translating by the way hehe?

J'essaie de vous expliquer que j'ai déjà exploré ce sujet en profondeur et cette itération actuelle représente la meilleure qualité que je puisse obtenir avec Flux. Je n'utilise pas le trainer de civitai.com et je ne veux pas réduire la taille car cela compromet la ressemblance et le réalisme. Je ne pense pas que vous compreniez pourquoi ou ce que je fais, mais je peux répéter que j'entraîne directement sur le checkpoint Flux - j'ai un fichier complet de 23,7 Go duquel je tire ce modèle. C'est un processus complètement différent d'un Lora normal, que j'ai abandonné à cause de ses mauvaises ressemblances et son manque de flexibilité.

On peut forcer une ressemblance proche avec un Lora basique, mais quand il finit par s'entraîner au point où le menton proéminent disparaît, il devient beaucoup moins flexible que cette itération actuelle qui, elle, n'a pas ce problème de menton, offre une excellente représentation du personnage à travers différents types d'images et est aussi incroyablement flexible, comme vous pouvez le voir dans les images publiées.

Si possible, vous devriez comparer certains de vos meilleurs résultats avec celui-ci ou mes autres travaux pour voir la différence directe entre l'entraînement Lora standard et ma méthode/taille. Ce n'est toujours pas parfait mais je pense que c'est aussi proche de la perfection que Flux peut atteindre avec les outils actuels.

Le modèle de Lady Gaga utilisait 25 images (beaucoup pour Flux), c'étaient les meilleures que je pouvais trouver, sans personnes étranges, mains bizarres, grain de film ou bruit parasites, ou mauvaise ressemblance du personnage. Je fais les légendes avec cogVLM2 puis je les reprends moi-même pour les éditer et les rendre aussi bonnes que possible. Ensuite, le modèle cuit pendant environ 24 heures sur mon PC et je reteste en utilisant mes légendes originales pour voir à quel point il est proche de l'original. Une fois qu'il semble terminé, je commence à travailler sur les variations. C'est l'aperçu basique sans parler des paramètres, mais si vous voulez les voir, j'ai fait un post sur Reddit à ce sujet.

@LOREVER resemblance sure, when I look at your lora's they sometime resemble the subject but look closely at the lines in the face and how they sit in the environment of the picture. They stick out in a weird way like they aren't a part of the image as much, and their skin is like smooshed in a certain way. There is a lot of noise that is learned in the photographs that make them look extra real. With this method this size lora is the the medium size (1.1gb). It does a good representation but it is not the end of the realism spectrum. When I go lower say with this model specifically, say I rip a 64dim lora off the checkpoint (a 500mb or so), the likeness clearly starts to fall apart. This is not the same in animation which takes less to learn and get good representations and works much better at a lower dimension, even a 32dim (200mb) works with that. The lora's you are doing on this site are all only training a couple of layers of a low dimension lora, it is how the site is built. Why is your family relevant? Don't use your family as an excuse or crutch, it's ok to not have the money, I personally buy used as a rule and the cool thing about checkpoint training is that even a 8gb card can do it as the kohya_ss guy posted on reddit. You would just use the config that I linked you and then turn up that block transfer like I talk about. In case you didn't know there is also a third option. You can use masscompute they have a hidden option to rent 48gb a6000 for only 30c an hour and it only takes like 9 or less hours to do this on one of those.