they always have the leftovers haha
they always have the leftovers haha
F
Fwell my wife prefers getting clothes so she gets them
Vwomen usually love two things - clothes and jewellery
MHi all,
has anyone yet tried to be more accurate while training specific body parts, e.g. abs and pecs of a trained man, etc.
Or legs, waist and bottom? Obviously you cannot always have the face on the pictures if you dont always wanna take full body shots, especially for lower bodyparts.
I'm wondering what is the possible way to use closeup pictures to be more detailled.
Right now I'm not labelling / captioning my pictures.
But I dont know how the trainer is supposed to know that this bottom belongs to "ohwx man" or similarly.
Has anyone experience if this specific task can be improved by only labelling those specific images exactly with using the instance prompt? "bottom of ohwx man from behind". Or stuff like this?
has anyone yet tried to be more accurate while training specific body parts, e.g. abs and pecs of a trained man, etc.
Or legs, waist and bottom? Obviously you cannot always have the face on the pictures if you dont always wanna take full body shots, especially for lower bodyparts.
I'm wondering what is the possible way to use closeup pictures to be more detailled.
Right now I'm not labelling / captioning my pictures.
But I dont know how the trainer is supposed to know that this bottom belongs to "ohwx man" or similarly.
Has anyone experience if this specific task can be improved by only labelling those specific images exactly with using the instance prompt? "bottom of ohwx man from behind". Or stuff like this?
A
S@Dr. Furkan Gözükara I did not notice you were training up to 150 epoch for just 15 images. That makes me wonder what would be the ideal epoch to train up to for 4000 images?
SAre your sample images on early epochs looking fried?
Y@Dr. Furkan Gözükara , is there any LoRA model we can play with that shows the full implementation in pytorch ? I would like to see the layers , connections and everything to change the architecture and see the impact for instance
R@Dr. Furkan Gözükara testing out FullFinetune on Kohya and having issues with metadata processing. Do i have to run a script to convert my txt caption files? Do you have a video for this?
ROSError: [Errno 22] Invalid argument: 'E:\studio Dropbox\studio\ai\libs\SD\1models\1styles\gr4ph1cW_v001admw8bt_50wd14_rep5_bs6\n_20240404-145300.json'
Rthis is the error
A@Dr. Furkan Gözükara I followed your onetrainer tutorial with your preset and whenever I train a model its just broken. My current model wont even load in automatic1111
Ffor such images
Ftraing as much as you can
Fand save checkpoints every epoch
Fyou can look diffusers library i guess
Fcould be model corrupted somehow
Fi did over 15 training - different trainings - all working perfect
Fgood news massed compute increased gpu pool
A@Dr. Furkan Gözükara here are my settings, all setup but none of my models ever seem to work
Ftry to download base model
Finto disk
Fand give its path
Fi wonder if your sdxl downloaded cache file is broken
Fby the way before starting training
Fi test base model with auto1111
Fand verify it is working or not
DHi @Dr. Furkan Gözükara small question, I have a volume storage on Runpod and I wanted to reduce space but since it's not possible, I created a new one and I just copypasted my entire ComfyUI folder. But now every time I start a new pod, I try to start comfyui (installed with your script) by typing:
apt update
apt install psmisc
fuser -k 3000/tcp
cd /workspace/ComfyUI/venv
source bin/activate
cd /workspace/ComfyUI
python main.py --listen 0.0.0.0 --port 3000
But everytime it asks me to install this:
pip install safetensors
pip install einops
pip install transformers
pip install scipy
pip install torchsde
pip install aiohttp
Any idea how I can solve it? Also as a request, would be great to have more scripts for ComfyUI and SDXL. Lots of thanks!
apt update
apt install psmisc
fuser -k 3000/tcp
cd /workspace/ComfyUI/venv
source bin/activate
cd /workspace/ComfyUI
python main.py --listen 0.0.0.0 --port 3000
But everytime it asks me to install this:
pip install safetensors
pip install einops
pip install transformers
pip install scipy
pip install torchsde
pip install aiohttp
Any idea how I can solve it? Also as a request, would be great to have more scripts for ComfyUI and SDXL. Lots of thanks!
Fwell it shouldnt happen
Fbecause we install into venv
Fit must be related to your folder copy pasting
Falso how did you copy paste?
Fyou cant copy paste between pods
CI am also facing this issue, I tried first autotrain-advanced but was not able to assign captions, then I am now trying Kohya but still facing issues. Also May I know which models you are testing with?
CInteresting!
DI made a new storage volume, then I used runpodctl to transfer my confyUI folder from one volume to another.
The issue happens everytime I start a new pod and try to run comfyui
The issue happens everytime I start a new pod and try to run comfyui
Fthat is wrong
Fyou need to make a fresh install
Fyou can only transfer models
C@Dr. Furkan Gözükara
C
C
Cthis break cache progress i have to redo the cache again
Twhat settings do i adjust in supir so that it doesnt hallucinate as much. i did fidelity and v0-F
K@Dr. Furkan Gözükara Is it possible to add a checkbox or something to make it possible for CLIP Interrogator to batch process subfolders as well?!
NYes, I am doing that now on most smaller resolution datasets for concepts, however I would be very wary of any noticeable artificial enhancements it may be doing to your dataset as that "artificialness" may be enhanced in your SD outputs when you include the LoRA in your inference. I find that keeping the prompt text influence fairly low ~4 or less seems to keep the upscale/detail recovery adequate without introducing artificialness into the AI upscaled image. Your mileage may vary...
NNot sure who has responded but I train LoRAs (presently using Kohya, will attempt to transition to OneTrainer soon) that require a high amount of quality and accuracy due to celebrity reproduction. What I am finding is that approaching model sized datasets (~300 images) is the key to increased accuracy across the board of your concept. This is the easiest and most straight forward method, but not always available if you can't acquire a dataset with enough good quality representations of your concept(s). So alternatives:
1) Train multiple concepts - This from reports is a mixed bag. I've done it for a TV celebrity who plays an iconic character with OK success. If I did it again, I would increase the size of my dataset for each concept to ~250 or more images.
Further to this, you will also need to watch out when training lots of closeups to get it to learn a concept as it will lose reference with your MAIN subject (depending on what your concept). So say your concept is a male celebrity but you want the abs more accurate, doing all closeups of the abs may cause it to lose reference to the main subject you want. After all it is giving you what its learned (just abs closeups) from the trigger word you associated it with.
2) Train with color association or some other novel machine learning method - See Civitai article here: (sorry it appears he's removed the article and model). There is a reddit post still here about it: https://www.reddit.com/r/StableDiffusion/comments/1aolvxz/instructive_training_for_complex_concepts/?share_id=TyIZDSSNoYSqsqZVR0SzW
3) Train them as separate LoRAs, this would operate like 1) but remove the pitfalls people claim happens with multi-concept training in a single LoRA.
4) Learn how to train a model for ADetailer specific for your subject (tbh, I have no idea where to start with this one, but it may be what you and I both need to get body part accuracy)
1) Train multiple concepts - This from reports is a mixed bag. I've done it for a TV celebrity who plays an iconic character with OK success. If I did it again, I would increase the size of my dataset for each concept to ~250 or more images.
Further to this, you will also need to watch out when training lots of closeups to get it to learn a concept as it will lose reference with your MAIN subject (depending on what your concept). So say your concept is a male celebrity but you want the abs more accurate, doing all closeups of the abs may cause it to lose reference to the main subject you want. After all it is giving you what its learned (just abs closeups) from the trigger word you associated it with.
2) Train with color association or some other novel machine learning method - See Civitai article here: (sorry it appears he's removed the article and model). There is a reddit post still here about it: https://www.reddit.com/r/StableDiffusion/comments/1aolvxz/instructive_training_for_complex_concepts/?share_id=TyIZDSSNoYSqsqZVR0SzW
3) Train them as separate LoRAs, this would operate like 1) but remove the pitfalls people claim happens with multi-concept training in a single LoRA.
4) Learn how to train a model for ADetailer specific for your subject (tbh, I have no idea where to start with this one, but it may be what you and I both need to get body part accuracy)
Reddit
Explore this post and more from the StableDiffusion community
N@Morpheus A quick update: regarding training a custom ADetailer model: I did find this but don't presently have time to read it through: https://civitai.com/articles/1224/training-a-custom-adetailer-model
NThere is a Civitai article if I can find it that talks about manipulating the layers for LoRA optimisation, way over my head. I'll see if I can find it. here we go, not sure if it will help, but good luck LOL: https://civitai.com/articles/76/bdsqlsz-lora-training-advanced-tutorial1lora-block-training
What is Lora block training? Lora Block training involves modifying the corresponding block weights during training through U-Net processing. First...
