Software Engineering Courses (SECourses)•16mo ago

I’m going to take a 1-minute clip from a movie and use Joy Caption to generate captions for each fra

I’m going to take a 1-minute clip from a movie and use Joy Caption to generate captions for each frame. Then, I’ll edit those captions into prompts that fit my AI model to hopefully optimize my workflow.

Xxlads I’m going to take a 1-minute clip from a movie and use Joy Caption to generate c...

Furkan Gözükara SECourses•9/13/24, 4:48 PM

sounds like a plan

FFurkan Gözükara SECourses yep overfit

dsienra•9/13/24, 5:48 PM

yes but maybe the key is the class on this case, it learned and associated the token to the background and not me, may be training T5 I must specify the class man, so it learns me rather than the background, is like it needs guidance to what to learn, just a theory.. hehe

slipkn67•9/13/24, 5:54 PM

in the <segment:face,0.7> part what happens if I don't reduce it to 0.7? Won't it bring better result?

xladsOP•9/13/24, 5:57 PM

so my caption joy just complete batch processing progress and says that the output folder is same as the input folder but when i check i can see anything

xladsOP•9/13/24, 6:01 PM

statred processin progress again this time with oput folder pasted in ui

Vinmi•9/13/24, 6:09 PM

Anyone has any idea why this workflow is not working? It gives me an error, but it was fully working with the original Flux-Dev(8fp)

Vinmi•9/13/24, 6:09 PM

Original message was deleted

Narva•9/13/24, 6:26 PM

theres 17 gpu aliviable

Maxivi•9/13/24, 7:17 PM

@Dr. Furkan Gözükara is the lasest kohya broken for you?

DDreamer@TDFilmStudio I have trained 3 custom Flux loras till now and everytime, I am getting scan lin...

Luciani•9/13/24, 8:15 PM

I have exactly the same issue with vertical lines appearing on the results. One point is to restrict generating to 1024 and do not go above. However, often the lines are still visible on that smaller images. Did you find a solution yet? You may also send me a direct message!

3DAI•9/13/24, 9:02 PM

Is there a way to use 2 loras of 2 different people together without them blending?

Sslipkn67 in the <segment:face,0.7> part what happens if I don't reduce it to 0.7? Won't i...

Furkan Gözükara SECourses•9/13/24, 9:02 PM

i,t is 0.6 by default

Furkan Gözükara SECourses•9/13/24, 9:02 PM

as far as i know

MMaxivi @Dr. Furkan Gözükara is the lasest kohya broken for you?

Furkan Gözükara SECourses•9/13/24, 9:03 PM

he fixed it

FFurkan Gözükara SECourses he fixed it

Maxivi•9/13/24, 9:04 PM

yep i saw he replied to me

Maxivi•9/13/24, 9:04 PM

seems you can only fix the bleeding with lokr

Maxivi•9/13/24, 9:04 PM

but the face quality is not good

LLuciani I have exactly the same issue with vertical lines appearing on the results. One ...

Furkan Gözükara SECourses•9/13/24, 9:05 PM

i almost never get them with 1024

Furkan Gözükara SECourses•9/13/24, 9:05 PM

but when i go 1280 1536 very noticeable

LLuciani I have exactly the same issue with vertical lines appearing on the results. One ...

Dreamer@TDFilmStudio•9/13/24, 9:41 PM

I'm beginning to suspect that overtraining might be the cause, but I'm not entirely sure anymore. Flux training can be quite unpredictable, especially in the early stages, and the guides often give conflicting advice. In my personal photo set, vertical lines keep showing up, even after three attempts. I’m starting to think there may be an issue with my training image sources as well. Whether I train for 1,000 steps or 3,500, the lines persist. Oddly enough, I trained two other Loras using different subjects, and those turned out perfectly fine, which makes the situation even more confusing.

Furkan Gözükara SECourses•9/13/24, 10:19 PM

looks like hugging face having huge internal issues

Pelolisu•9/14/24, 12:18 AM

https://huggingface.co/spaces/KingNish/Realtime-FLUX

FLUX Realtime - a Hugging Face Space by KingNish

PPelolisu https://huggingface.co/spaces/KingNish/Realtime-FLUX

Furkan Gözükara SECourses•9/14/24, 12:19 AM

i like your new photo :d

Pelolisu•9/14/24, 12:20 AM

haha

Pelolisu•9/14/24, 12:20 AM

is was an sd 1.5 photo realistic trainning

PPelolisu is was an sd 1.5 photo realistic trainning

Furkan Gözükara SECourses•9/14/24, 12:20 AM

nice

Pelolisu•9/14/24, 12:20 AM

but was try and error, because i never generated other like haha

Pelolisu•9/14/24, 12:21 AM

so i not have spend much time with training

Pelolisu•9/14/24, 12:29 AM

hi guys! i and @Dr. Furkan Gözükara is working in a new web ui for improve sdxl image generations, it will bring the most desired feature InstantID with InstantStyle, so stay tuned!