V3: fix for black screen some people are getting
V2: Rank 32 -> lower file size, more epochs
V1: Rank 128 -> large file size but better performing in my testing
This model was designed to be very flexible and to enhance the nudity of Krea 2 with many different body types. It works with men as well but was trained on female pictures. Examples are using Turbo 8 steps DDIM BETA. It was trained at 1696x1024 but you can use other wide resolutions such as 1920x1088
The base prompt is
A realistic photograph of a [DESCRIBE SUBJECT, e.g. blonde 21 year old woman], she is posing in 5 different [DESCRIBE VIEW, e.g. full body] poses, the first pose on the left she is wearing clothes and facing the viewer, the second pose she is naked nude and facing to the right, the third pose she is naked nude facing the viewer, the fourth pose she is naked nude facing to the left, the fifth pose she is naked nude facing away from the viewer showing her butt, the background is [DESCRIBE BACKGROUND, e.g. photo studio backdrop]The subject, view, poses, clothing/nudity, and background should all be able to be changed to what you want but if you change the base prompt too much you might need to cherrypick.
Description
FAQ
Comments (32)
why its so big? what rank? fp32?
128 dim / 64 alpha, yeah its pretty big which helps with this concept, i can try to see if it still works with lower ranks but that would be after I get my other models retrained
@Nugus You should try reducing the rank before retraining. At least see how well it works at Rank 64 and half the file size.
Looks great BTW
@Murf_ I tried rank 64/32 and its giving 50% errors, either more poses than there should be or people wearing clothes when they shouldn't, ive had similar problems with klein and z-image lower ranks, maybe its my setup but this concept has been one of the harder ones ive tried to keep consistent
@Nugus Interesting, I'm still new to natural language, but I'm wondering if it might be that your trigger "phrase" using too many tokens? Couldn't you have condensed your paragraph into a single short sentence like, "Posed in a five-angle view" or "she is posing in 5 different poses"? Probably a difference of 20ish tokens vs your 80ish tokens. Just an idea.
@Murf_ yeah it takes a lot of experimentation and every model is different, i never posted the z-image and klein loras because I never got it quite right, at first I tried it with a single trigger word for z-image and klein and it didn't go well so I tried these longer prompts, but I didn't try that the short or single word triggers with Krea 2. Krea 2 seems great and so it might respond better to that
@Murf_ I added a rank 32 version trained longer, go ahead and test it if you are still interested and post some examples here please :D
Thank you very much, amazing lora!
👍🏻👍🏻👍🏻
no idea why i get black output with this lora
same issue the only lora on civitai for krea2 that has done this i use krea2TurboNSFWAIO_v10 with basic setting no special workarounds SwarmUI and even at lora 0.15 strength with your recommended default prompt and resolution it will fully black out the entire image what a shame i was looking forward to this one
Strangely if i lower the lora strength to 0.1 it seems to work (4090 64gb ram pc not the issue default standalone SwarmUI also not likely the issue)
Anyone else that is having this problem, do you use SwarmUI?
@Nugus Same issue on comfyui, turbo fp8 scaled
I have the same issue. I've tried different models\checkpoints, workflows and got the same result. It works at lora str 0.1, but what the reason for usage Lora in this case? BTW thank you for awesome prompt. It works awesome even without lora. Could you please share your workflow?
@bugtest919811 @Vinnnn @Agent_Smth I dont know what is causing the problem but I posted a new lower rank version and maybe it solved something
@Nugus Sorry but it does not work as well: just a black screen. I even tried on the default KREA2 template with recomended model/vae/clip - no changes.
@Nugus will test bro
ok tested the new one same issue bro no idea why this is a pain because i love the Czech casting style of photo too are you using ostris to train them? using the turbo or the raw? if the turbo are you using the training adapter?
Im using this to simplify training and it works really well https://github.com/bongobongo2020/krea2-character-lora-trainer
@Vinnnn @Agent_Smth @bugtest919811 @Wallbreaker Well, I tried the lora stripper that reduces krea2 lora sizes to try to make a stripped down version 3 and it told me "no txtfusion keys found / not krea 2 architecture", Im using musubi-tuner to make my loras so maybe after the next update it will work better.
@Nugus you do you bro but the one i linked you can also use musubi-tuner in the background and train on raw and its fronted is soo simple i legit love this trainer i may even start releasing some stuff maybe if you can/ want too try running your dataset though that one and see if that fixes it, that said thanks for your effort to fix this i cant be the only one that appreciates your effort
It would help a lot if someone gets it to work to post a ComfyUI workflow. All I get is a black screen, even at 0.1 strength.
Come to think of it though reference to image seems to be garbage with this model anyway.
@jenniferhoustonpancakes @Agent_Smth @Wallbreaker @bugtest919811 @Vinnnn I added a ALT version, check the tabs at the top, maybe it will fix the problem, idk
@Nugus Will test and get back to you and this is very appreciated mate
EDIT* @Nugus Yes Bro this fixed it awesome!!!!!
Thanks for your prompt, I found that it works perfectly
This should be extremely useful for consistent characters.
Could you share your exact prompt and settings (weight/sampler) please?
prompt is just the base prompt using the examples provided in the description such as 21 year old blonde woman, weight is 1.0, using clownsharkksampler, 0.5 eta, 8 steps, DDIM sampler, beta scheduler, but it works fine with other settings and samplers
@Nugus Thanks !
Interesting.. can it do actual full side profile poses? In all your samples the woman's face is at 3/4 view in the side poses.
the training data is at face 3/4 view so it will be difficult, I just did a test after reading your comment and after lowering the strength of lora to 0.5 and adding looking to the left showing side face profile, I was able to get some side profiles after cherrypicking but the lower lora strength also lowered the prompt adherence of the other poses as well
@Nugus I see, any hopes of a V3 where side profile views are trained in?
@Nugus : for a version V3, could you also fix the problem that the feet are cut off?
But besides that a really great lora!
Thanks for that!