Software Engineering Courses (SECourses)•15mo ago

it's just.... i trained same 10img lora of myself once without captions and once with minimal captio

it's just.... i trained same 10img lora of myself once without captions and once with minimal captions. and it's so hard to say which is better side by side... subjective.
i need some more xy grids

Ttazztone it's just.... i trained same 10img lora of myself once without captions and once...

Furkan Gözükara SECourses•10/18/24, 11:03 AM

true

FFurkan Gözükara SECourses well flux has internal captioning mechanism

tazztoneOP•10/18/24, 11:15 AM

did BFL tell you? ^^

Ttazztone did BFL tell you? ^^

Furkan Gözükara SECourses•10/18/24, 11:31 AM

well it is discovered by several scripts developers

Furkan Gözükara SECourses•10/18/24, 11:31 AM

by analyzing its architecture

Furkan Gözükara SECourses•10/18/24, 11:31 AM

it generates image embeddings

Furkan Gözükara SECourses•10/18/24, 11:31 AM

which makes captioning effect

tazztoneOP•10/18/24, 11:31 AM

interesting. thanks for sharing

ParadoxNeon•10/18/24, 12:40 PM

@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?

PParadoxNeon @Dr. Furkan Gözükara Hello everyone, I have a question about custom lora trainin...

Furkan Gözükara SECourses•10/18/24, 12:54 PM

fine tuning prevents this

Furkan Gözükara SECourses•10/18/24, 12:54 PM

but this usually happens when you generate high resolution

Furkan Gözükara SECourses•10/18/24, 12:54 PM

ah yes you said 1024+

Furkan Gözükara SECourses•10/18/24, 12:54 PM

yes fine tuning reduces this a lot

Furkan Gözükara SECourses•10/18/24, 12:54 PM

also you can use SUPIR to upscale

Furkan Gözükara SECourses•10/18/24, 1:11 PM

supir face enhance difference

Furkan Gözükara SECourses•10/18/24, 1:11 PM

https://imgsli.com/MzA4NzQx/1/2

Furkan Gözükara SECourses•10/18/24, 1:11 PM

original vs supir https://imgsli.com/MzA4NzQx/0/2

FFurkan Gözükara SECourses get your training set better :d

deepblue_1337•10/18/24, 1:11 PM

is there a recommendation for a good training set? (background, composition....)

Ddeepblue_1337 is there a recommendation for a good training set? (background, composition....)

Furkan Gözükara SECourses•10/18/24, 1:13 PM

different background and clothing

Furkan Gözükara SECourses•10/18/24, 1:13 PM

will yield better results

Furkan Gözükara SECourses•10/18/24, 1:13 PM

so model wont memorize

Furkan Gözükara SECourses•10/18/24, 1:16 PM

fine tuning tutorial this will take forever to edit :/

IWishIWereGood•10/18/24, 2:08 PM

Hi 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

IIWishIWereGood Hi everyone, I'm not sure which chat to post this in, but can someone explain to...

Furkan Gözükara SECourses•10/18/24, 2:11 PM

swarm internalk

Furkan Gözükara SECourses•10/18/24, 2:11 PM

check those parameters

Furkan Gözükara SECourses•10/18/24, 2:11 PM

IWishIWereGood•10/18/24, 2:15 PM

Hmmm I will look for it

IWishIWereGood•10/18/24, 2:15 PM

Thank you

FFurkan Gözükara SECourses you can convert model to fp8

Nick44•10/18/24, 5:04 PM

that worked thanks

dsienra•10/18/24, 5:08 PM

Update 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-

nyanko7/flux-dev-de-distill · Hugging Face

Ddsienra https://huggingface.co/nyanko7/flux-dev-de-distill Unet LR 0.0001 vs 0.00003 the...

�

𝗡𝗶𝘁𝗶𝗻 𝗣𝗮𝗿𝗲𝘃𝗮•10/18/24, 5:11 PM

Can you share some images from your trainings comparing the normal dev vs de distilled one?

�𝗡𝗶𝘁𝗶𝗻 𝗣𝗮𝗿𝗲𝘃𝗮 Can you share some images from your trainings comparing the normal dev vs de dis...

dsienra•10/18/24, 5:18 PM

Not 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.

Ddsienra Not from my training because is from family and friends I can't share, The quali...

�

𝗡𝗶𝘁𝗶𝗻 𝗣𝗮𝗿𝗲𝘃𝗮•10/18/24, 5:19 PM

Okay thanks for all your updates here

�𝗡𝗶𝘁𝗶𝗻 𝗣𝗮𝗿𝗲𝘃𝗮 Okay thanks for all your updates here

dsienra•10/18/24, 5:20 PM

you welcome.

Ddsienra Update https://huggingface.co/nyanko7/flux-dev-de-distill training. LR 0.0001 vs...

druhl•10/18/24, 5:35 PM

My 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.

druhl•10/18/24, 5:36 PM

My training on dev2pro with the same parameters (same concept) would also be completed soon.