your checkpoints looks like broken
your checkpoints looks like broken
Awhy?
Fyou got black output
F
Fwhich config did you use?
Ause, LR 6e-3, Text encoder 3e-6, 32 Batch Size, 10 epoch, and Gradient accumulate steps 2
Afinetunig, illustr XL Model, in format safetensorts
Feach config unique . if you use my config
Fand if you want 32 batch size
Fmultiply my learning rate size with 5.7
Athe configuration of this video https://www.youtube.com/watch?v=EEV8RPohsbw, I was guided by the last tutorial I published about flux.
YouTubeSECourses
Patreon post for config files 
https://www.patreon.com/posts/89213064
SECourses Discord To Get Full Support
https://discord.com/servers/software-engineering-courses-secourses-772774097734074388
My LinkedIn
https://www.linkedin.com/in/furkangozukara/
My Instagram
https://www.instagram.com/gozukarafurkan/
My Medium
@FurkanGozuka...
https://www.patreon.com/posts/89213064
SECourses Discord To Get Full Support
https://discord.com/servers/software-engineering-courses-secourses-772774097734074388
My LinkedIn
https://www.linkedin.com/in/furkangozukara/
My Instagram
https://www.instagram.com/gozukarafurkan/
My Medium
@FurkanGozuka...
Aok
Fthen checkpoints making black is the issue
Fdid you try model without training?
Fthat base model you are using?
Ait is an XL model, I talked to the creator, and he recommended me not to use clip skip, and to add extra parameters to Adafactor.
Jthe Mochi1 videos are really fun to create, are there any future plans for img2video?
Fcomfyui just told me they planning
Fgreat news
LHi, I want to train a rock band, it's a trio: vocals and guitar (leader), bass, drums. I plan to distribute the photos in folders: folder with photos of the singer, folder with photos of the bassist, folder with photos of the drummer, and folder with photos of the band complete. Has anyone trained Lora for more than one person? What photo distribution do you recommend?
Ldataset/
├── Band_complete/ # Group photos of the entire band
│ ├── Band_complete_001.jpg
│ ├── Band_complete_002.jpg
│ └── ... (X photos)
├── Leader/ # Individual photos of the leader
│ ├── Leader_001.jpg
│ ├── Leader_002.jpg
│ └── ... (Y photos)
├── Bassist/ # Individual photos of the bassist
│ ├── Bassist_001.jpg
│ ├── Bassist_002.jpg
│ └── ... (Z photos)
└── Drummer/ # Individual photos of the drummer
├── Drummer_001.jpg
├── Drummer_002.jpg
└── ... (W photos)
├── Band_complete/ # Group photos of the entire band
│ ├── Band_complete_001.jpg
│ ├── Band_complete_002.jpg
│ └── ... (X photos)
├── Leader/ # Individual photos of the leader
│ ├── Leader_001.jpg
│ ├── Leader_002.jpg
│ └── ... (Y photos)
├── Bassist/ # Individual photos of the bassist
│ ├── Bassist_001.jpg
│ ├── Bassist_002.jpg
│ └── ... (Z photos)
└── Drummer/ # Individual photos of the drummer
├── Drummer_001.jpg
├── Drummer_002.jpg
└── ... (W photos)
LFolder Structure and Photo Distribution
dataset/
├── Band_complete/ # Group photos of the entire band
│ ├── Band_complete_001.jpg
│ ├── Band_complete_002.jpg
│ └── ... (X photos)
├── Leader/ # Individual photos of the leader
│ ├── Leader_001.jpg
│ ├── Leader_002.jpg
│ └── ... (Y photos)
├── Bassist/ # Individual photos of the bassist
│ ├── Bassist_001.jpg
│ ├── Bassist_002.jpg
│ └── ... (Z photos)
└── Drummer/ # Individual photos of the drummer
├── Drummer_001.jpg
├── Drummer_002.jpg
└── ... (W photos) Distribution of Photos
Band_complete: Approximately 30% of total photos.
Leader: 40% of total photos (assuming the leader is more prominent).
Bassist: 15% of total photos.
Drummer: 15% of total photos.
dataset/
├── Band_complete/ # Group photos of the entire band
│ ├── Band_complete_001.jpg
│ ├── Band_complete_002.jpg
│ └── ... (X photos)
├── Leader/ # Individual photos of the leader
│ ├── Leader_001.jpg
│ ├── Leader_002.jpg
│ └── ... (Y photos)
├── Bassist/ # Individual photos of the bassist
│ ├── Bassist_001.jpg
│ ├── Bassist_002.jpg
│ └── ... (Z photos)
└── Drummer/ # Individual photos of the drummer
├── Drummer_001.jpg
├── Drummer_002.jpg
└── ... (W photos) Distribution of Photos
Band_complete: Approximately 30% of total photos.
Leader: 40% of total photos (assuming the leader is more prominent).
Bassist: 15% of total photos.
Drummer: 15% of total photos.
Fyou dont need
Fyou can use .txt caption
Fand have all in same folder
Fif you gonna use repeat 1
Lgreat, so I only differentiate them with the captions? eg. leader(1).txt : nameofrockbandxyz, j0hnxyz leader vocalist, holding a microphone on a stage....
LI would use a single folder, 1 nameofrock bandxyz people (1 repeat_triggerword class)
WHi @Furkan Gözükara SECourses have you tried training a style with dreambooth. I wonder what is different when a checkpoint learning a style compare with lora. Does it learn faster so we can train with lower epoch
Hhi, I trained this sdxl checkpoint in the past https://civitai.com/models/570998/cocaxl-01 but it was impossible to generate character sheets with it. So I trained another checkpoint based on https://civitai.com/models/293331/wildcardx-xl-turbo, the previous one had been trained on juggernaut xl. Not character sheets work better, but there is a problem.
Hfor the same prompt the new model outputs this
Hthe old one outputed this
HI don't mind the output, but it doesn't render the clothing as well as the old model
HI am talking about this
Hfor example for the footwear, when I input "man wearing ((opinci-like moccasins)) " the old model creates pretty accurate opinci, the new one, not as much, also it creates 2 people where as I only asked for one. And in general it doesn't stick to the prompt as well.
HSo my question is: is there something I did wrong with the training? can I do anything to make the new model be more like the old model with respect to accurately represanting the clothing
Fyes but i wont read folder caption if you use txt captions
Fimpossible to know without looking everything it requires full time debug. you can also try flux did you try?
Hit takes way to much time, I have 340 training images scanned from old children's story books, it took 27 hours to train on my wife's hp omen
WBut this is lora not a full model, I already read your writing many times but it's only about lora training
Fit is exactly same for fine tuning
Fyou only change config
WYes but i wonder how about learning speed does full model require less epoch to learn style than lora.
Fit should be around same
Hif I increase the text encoding rate and train again could that improve the results? I used the low vram settings for your tutorial, I just increased the repetitions from 40 to 100 on the first screen
Fit can cook text encoder
FHi Boss @Furkan Gözükara SECourses , for SDXL training I can more or less use the same settings for flux, just changing at the beginning ? Or you recommend different tutorial?
Fwe have sdxl config
Fyou should use it as a base
Fyes, but you dont have like sdxl config file by any chance?
