yes dataset is king
yes dataset is king
Fupdated the post
FPatreon
Get more from SECourses: Tutorials, Guides, Resources, Training, FLUX, MidJourney, Voice Clone, TTS, ChatGPT, GPT, LLM, Scripts by Furkan Gözü on Patreon
Fwill update 1 more time
WHi @Dr. Furkan Gözükara for training person you said that no caption is better than with caption, however when I read your 3d lora style training section you also said if the dataset is inconsistent, caption is better. So does it apply to training person too
WMy dataset is something like this
Wmultiple size with 1024x1024, 512x768 or 768x512
WDoes it count as inconsistent dataset
WDo I need to train with caption
Phavent tried Swarm UI .. versy used to SD Forge . But maybw worth looking into ?
Мflux allows to create in images inscriptions in English, is there a tool to add inscriptions in Russian?
AHi Everyone
Awho got experience with using AI Toolkit for fine tuning flux?
AI need some guidance please!
Fgood question. it is a same person so may not apply
Funless person is very different in photos
Fyes swarm is really good
Fnot that i know but fine tuning may add that capability
Fwhy dont you use kohya?
МWhat do you mean? I'm not very good at this...
Fi mean model need to be trained on such dataset
�@Dr. Furkan Gözükara Can I set the weight of lora used during generating samples while it's training or anyway to know what weight it used to generate those samples?
ASure I can give try
AI want to replicate a meme actually - naming Kwen
AFor e.g. this character
A
AI took this from net but lets assume this is not there in the net
Aand the model havent seen the character at all
Aso what configuarations will give me the best results
AHow many pictures I will need
Fi dont use samples usually not working
For breaking training :d
ASo model has no information about Kwen it doesnt know anything but I want to generate such that it hangs out with Mickey Mouse
Fwith dataset you can make
ASo how many pictures will I need? LoRA ranking? Captioning of image? There is florence 2 built in with ai toolkit but isnt great at captioning
Fmore better
Fwe have full configs on patreon use according to your vram
Fi dont suggeest captions
Ayes I have subscribed to your patreon
AWill I be able to use it with ComfyUI?
Aafter fine tuning with Kohya?
Fof course
Fi use with swarmui it uses comfyui as back end
ANice!
�@Dr. Furkan Gözükara What to do about bleeding? I have tried lower network and lower learning rates but can't stop bleeding
KNo. You only specify one value for vertical AND horizontal resolution. The training is on square images. You should specify 1024 in your case. The images will be bucketed if they ar not square or if they are a different (higher) resolution than what's being trained.
K@Dr. Furkan Gözükara https://huggingface.co/blog/MonsterMMORPG/ultimate-flux-lora-training-tutorial <- Is 404
SDr Furkan there is new settinsg that improve the velocity of trianing for a 4070 ti super 16 vram actualli is 10 its i don rmeember the nomenclaure in rank 5
Fi couıldnt stop yet either
Fye they silently
Fdeleted