did BFL tell you? ^^
did BFL tell you? ^^
Fwell it is discovered by several scripts developers
Fby analyzing its architecture
Fit generates image embeddings
Fwhich makes captioning effect
Tinteresting. thanks for sharing
P@Dr. Furkan Gözükara Hello everyone, I have a question about custom lora training on kohya using flux
I trained my face with 21 photos and 142 epochs as you mention on the equation on your video, which is... 3.000 / (number of photos) = number of epochs.
When I write the prompt on SwarmUI using Flux.1-dev-fp8. with 1024x1024 resolution the picture looks with low quality, also when the size is 1024+ seems to appear some artifact vertical lines on the picture (see the image) . I'm thinking to train again from zero a new lora with more images or use finetuning. What do you recommend?
I trained my face with 21 photos and 142 epochs as you mention on the equation on your video, which is... 3.000 / (number of photos) = number of epochs.
When I write the prompt on SwarmUI using Flux.1-dev-fp8. with 1024x1024 resolution the picture looks with low quality, also when the size is 1024+ seems to appear some artifact vertical lines on the picture (see the image) . I'm thinking to train again from zero a new lora with more images or use finetuning. What do you recommend?
Ffine tuning prevents this
Fbut this usually happens when you generate high resolution
Fah yes you said 1024+
Fyes fine tuning reduces this a lot
Falso you can use SUPIR to upscale
Fsupir face enhance difference
F
Foriginal vs supir https://imgsli.com/MzA4NzQx/0/2
Dis there a recommendation for a good training set? (background, composition....)
Fdifferent background and clothing
Fwill yield better results
Fso model wont memorize
Ffine tuning tutorial this will take forever to edit :/
IHi everyone, I'm not sure which chat to post this in, but can someone explain to me how to properly use wildcards in swarm ui? I have a weird issue with. I have a wildcard file with 10 lines for shirt color in it but when I try to generate 10 images, it will choose one color from those 10 and generate 10 images with that color
Fswarm internalk
Fcheck those parameters
F
IHmmm I will look for it
IThank you
Nthat worked thanks
DUpdate https://huggingface.co/nyanko7/flux-dev-de-distill training. LR 0.0001 vs 0.00003, much better results with lower LR, may be a little undertrained. Less class bleeding, but the class on training captions seems to decrease resemblance in non de-destilled models, I think that adding the class is a problem, I will remove the class from the captions I think that it will eliminate class bleeding completely. You can always add the class on inference if you wish, but removing it from the captions will protect the class from bleeding. flux-dev-de-distill learns very different than regular flux-dev it learns the caption tokens much better. I will update tomorrow-
�Can you share some images from your trainings comparing the normal dev vs de distilled one?
DNot from my training because is from family and friends I can't share, The quality on inference is identical, the advantage is that you can train multiple people on the same lora, I trained 11 people in one lora. I'm perfecting the technique to avoid concept bleeding completely not only between the people I'm training that is fixed, but fix the class bleeding to a minimum too.
�Okay thanks for all your updates here
Dyou welcome.
DMy training on de_distill just completed (dreambooth on a single person). The ability to control cfg helps you create a far greater image than CFG 1 @ flux_dev. The results are absolutely amazing. You can create more noisier (and less plasticky images) which seem more realistic. The ability to use negative prompts and the prompt adherence is very, very good.
Only downside is: Time! It sure takes helluva lot of time to generate (Minimum Steps 70 imo), and there are extra inference parameters to deal with.
Only downside is: Time! It sure takes helluva lot of time to generate (Minimum Steps 70 imo), and there are extra inference parameters to deal with.
DMy training on dev2pro with the same parameters (same concept) would also be completed soon.
Dgreat!!!
Ddev2pro was made to train loras and then use them on regular flux I think, is not really de-distilled
DI never try it, but I can try and see if it helps with concept bleeding.
DIt is not truely de-distilled true.
DI don't like what it does to hands so far lol
DWith thresholding, it is able to generate nicely at CFG 3.5.
DIt's in a strange place b/w flux dev and de-distill. 
NI still use jegernaught and which models in Super now give the best result for realism for CG graphics? 3D, 2D graphics and drawing for example anime
Fi didnt test for it actually
Fnow i need to record intro part of the tutorial
Ftime to edit
F
Aanyone know if its possible to increase the vocab of Flux Dev ? Like to train the clip models to add new vocab ?
DYes, you can train the clip or T5 through captions
Aboth at the same time ?
Fno way to increase as far as i know. everything still will get tokenized
