with 0.0003 LR adamw_8bit do you think 200 epoch for 20 image will make it over trainned?
with 0.0003 LR adamw_8bit do you think 200 epoch for 20 image will make it over trainned?
Fi cant comment based on other configs. but my config wont
KI just found that if you put "from text file per sample" , and dont provide any caption, it still does the lora training. what will it actually do?
KI tried that and it was working without any trigger word. like textual embedding
Fit will use internal caption it has
Fso will work still true
DI've used your Flux dreambooth training presets and the results are amazing. Would training at 1280 x 1280 increase the quality even further or is that useless to try?
Dwell I'm trying it anyway lol but it's going to take awhile
Fsadly not improving
Flet me know results. but i didnt get better 
NI usually do LoRAs over full dreambooth trainings. did one LoRA at 1536x1536. on my 3090 i did 150 epochs on 20 images and it took 22 hours 
same dataset and settings at 1024x1024 had only taken 7 hours. (i think the configs and kohya itself have both been optimized a lot since then) i do feel like the 1536 LoRA is better, more detailed. but after that i stuck with 1024 - didn't feel like the improvement (it's usually subjective by then anyway with character LoRAs from only epoch 100-125 turning out awesome as well) were worth triple the training time.
same dataset and settings at 1024x1024 had only taken 7 hours. (i think the configs and kohya itself have both been optimized a lot since then) i do feel like the 1536 LoRA is better, more detailed. but after that i stuck with 1024 - didn't feel like the improvement (it's usually subjective by then anyway with character LoRAs from only epoch 100-125 turning out awesome as well) were worth triple the training time.
Fyes doesnt help much. but you can try tiled training
Fmay work for objects
DI trained 59 images at 1024 x 1024 190 epochs (of a complex artstyle) and it works great even at high resolutions like 2000x2500 or something without upscaling. So I'm really curious to see if training at 1280x1280 would offer any improvements at high resolutions. I'll let you know whenever it finishes.
Fif you test let me know please ty
DI'll be able to check in a few hours
DSome things look better, some things look worse. I think it's not really worth it...
It seems to add more detail to small text or backgrounds, but the characters look worse
It seems to add more detail to small text or backgrounds, but the characters look worse
Fthanks a lot for info
CI rented massed compute first time after few month off since i have rtx 4090... I though generating wan2.1 will be much faster....lol.. definitely not on L40S lol slow.. Maybe H100, but to expensive..
Ftrue. 4090 is still a beast
Fand china selling 48gb and 96 gb . 48 gb is literally l40s
Kman I'm trying to do dreamboothsdxl training to the likeness of an individual and I've tried so many times and the sample output images continue sucking. I subscribed to the Patreon, can anyone point me towards the most useful pages re: optimal parameters and so on?
KIf your training data set is 59 images and each is being repeated 10 times, how many regularization images is ideal? Also, is repeating 10 times the best number of repeats?
Fhello sorry for late reply
Fi recommend you to train flux and we dont use reg images there
Fhere best tutorial
FYouTubeSECourses
If you want to train FLUX with maximum possible quality, this is the tutorial looking for. In this comprehensive tutorial, you will learn how to install Kohya GUI and use it to fully Fine-Tune / DreamBooth FLUX model. After that how to use SwarmUI to compare generated checkpoints / models and find the very best one to generate most amazing image...
KThanks apprecaiate it, I have since switrched to flux and am working on that currently.
KWould you say FLUX is the current SOTA for photoealistic images trained to a particular individual?
KI bought the personal consultation thing at Patreon, how do we set that up?
Fwhat is your username?
Fyou can private message me here
Fok i see it
Fmessage me here so we can setup. we can do now too
SI did too- Gold membership
Fhello please message me details . did you subscribe gold? private message me
MAre there any special instructions or 'Additional parameters' to enter into the Kohya_ss GUI when trying to "Resume from saved training state (path to "last-state" state folder)?"
I have entered the path/workspace/stable-diffusion-webui-forge/models/Stable-diffusion/fruit_realismByStableYogi_v5XLFP16/model/ohwx-woman_Fruit_Full-finetune_adafactor_realismIllustriousBy_v35FP16_292-Img-200-Instances-step00032000-state
I have verified the contents within that folder include all of the expected, model_1, model, optimizer, random_states, scheduler.
The training starts, however, it starts at step 01, versus step 32001 as I intend to continue traning.
Can anyone offer insight or help here? Thank you in advance!
I have entered the path/workspace/stable-diffusion-webui-forge/models/Stable-diffusion/fruit_realismByStableYogi_v5XLFP16/model/ohwx-woman_Fruit_Full-finetune_adafactor_realismIllustriousBy_v35FP16_292-Img-200-Instances-step00032000-state
I have verified the contents within that folder include all of the expected, model_1, model, optimizer, random_states, scheduler.
The training starts, however, it starts at step 01, versus step 32001 as I intend to continue traning.
Can anyone offer insight or help here? Thank you in advance!
Qdoes kohya work with rtx 5090? 
Fyes
Fwe have updated our installers just use it
Fjust use your last checkpoint
Fi never used last state
Fwith our config exactly same
QThanks! i've just upgraded my setups from 4090 to 5090 and noticing that I have problems running pretty much everything I tried before thanks again! Now i think im getting it
i've just upgraded my setups from 4090 to 5090 and noticing that I have problems running pretty much everything I tried before thanks again! Now i think im getting it
HHey, I wonder, has anyone tried / have article or any sort of content about full checkpoint finetuning on Wan ? I really want to get my hands into it but it'd be great to have some references to start working on it
Has far as I know only DiffSynth offers the possibility to full checkpoint training on Wan, right ?
Fgood point. i havent checked that yet though. currently musubi is good for lora
�What are the optimal sdxl fine tune and lora configs? Having a hard time finding them. Using massedcompute. Onetrainer or dreambooth? Was thinking of using big love xl 2.5 as base checkpoint. Like the realism and quality of pics there.
�