@Dr. Furkan Gözükara Is it possible to integrate joycaption as the default captioner for Supir, inst
@Dr. Furkan Gözükara Is it possible to integrate joycaption as the default captioner for Supir, instead of llava?
Nrunning a big grid right now comparing different versions of a person lora... 25, 50, and 75 epoch versions all have long hair for "man as a cosmic jazz improviser whose music causes spontaneous formation of new elements and life forms and wearing eyeglasses" even though short hair in all training images 
Fi may integrate
DSince you've mastered custom LoRA training for people, I encourage you to explore Flux LoRA training for products, objects, and art styles.
Fyep it is also in progress
FI agree, I don't understand why I'm losing the comic style I trained my lora with via comfyUI. I do have my avatar but in a realistic style, whereas I keep the comic side when I go through automatic 1111. I'd like to train a style to have consistency.
DFlux is clearly excelling at art style and product LoRAs, as evidenced by the recent surge of these types on Civitai. Also I look forward to seeing architectural and other categories too.
Fi added new config
FBest_Configs_Better_Colors
Fit may be fix for style training
Fi will do more research with it not compoleted yet
F@NicB we got error with forge
Fi am gonna email you details
D@Dr. Furkan Gözükara I'm having this errors is training with TE, the speed is lower 6s/it, do you know what this errors are?
Fthey are warnings
HWhat are the chances that you will make a one click installer for aitoolkit?
Fcan you give me github url
NGitHub
Various AI scripts. Mostly Stable Diffusion stuff. - ostris/ai-toolkit
Fit already has installation?
F
Fto be able to use it you need to know command line arguments passing
Fdont worry i am recording kohya tutorial hopefully today
Fi am testing cu118 vs cu124 atm
F12 gb training became slower after last kohya update
Fanyone noticed?
Flike 2x slow
For maybe i am wrong testing
Frtx 3090 way powerful than 3060 for sure
F3.15 times exactly matching lol
DThis style lora is really good: extract - I'm gonna answer here if other people wants more details as well.
For the captioning, here are some examples : https://imgur.com/a/hxAm3fo (used Florence2 as the base caption then rewrote them)
I've used 29 images (all of different aspect ratio, I didnt crop anything). It was with Civitai Lora training tool and I barely changed anything from the default settings :
"engine": "kohya",
"unetLR": 0.0005,
"clipSkip": 1,
"loraType": "lora",
"keepTokens": 0,
"networkDim": 2,
"numRepeats": 15,
"resolution": 512,
"lrScheduler": "cosine_with_restarts",
"minSnrGamma": 5,
"noiseOffset": 0.1,
"targetSteps": 1088,
"enableBucket": true,
"networkAlpha": 16,
"optimizerType": "AdamW8Bit",
"textEncoderLR": 0,
"maxTrainEpochs": 10,
"shuffleCaption": false,
"trainBatchSize": 4,
"flipAugmentation": false,
"lrSchedulerNumCycles": 3
For the captioning, here are some examples : https://imgur.com/a/hxAm3fo (used Florence2 as the base caption then rewrote them)
I've used 29 images (all of different aspect ratio, I didnt crop anything). It was with Civitai Lora training tool and I barely changed anything from the default settings :
"engine": "kohya",
"unetLR": 0.0005,
"clipSkip": 1,
"loraType": "lora",
"keepTokens": 0,
"networkDim": 2,
"numRepeats": 15,
"resolution": 512,
"lrScheduler": "cosine_with_restarts",
"minSnrGamma": 5,
"noiseOffset": 0.1,
"targetSteps": 1088,
"enableBucket": true,
"networkAlpha": 16,
"optimizerType": "AdamW8Bit",
"textEncoderLR": 0,
"maxTrainEpochs": 10,
"shuffleCaption": false,
"trainBatchSize": 4,
"flipAugmentation": false,
"lrSchedulerNumCycles": 3
DReddit
Explore this post and more from the StableDiffusion community
Dfor me is running 2x slow but I'm training the text encoder, maybe is it?
Fyes text encoder would slow down
Fi am yet to test it
Ffirst i am gonna test new parameter
Fstarting text encoder trainings too 
Fye total 8 new trainings starting
Fnow need to record widnows tutorial
Ffor kohya training
Didk the learning rate, to start I'm using the same 0.0001 for unet and the TE
Fi am testing them :d
Ffirst tests
F5e05 unet and clip
Fsecond is
F5e05 unet and 1e05 text encoder
Fother 6 tests about flow shift
DI went more aggressive with 1e04
Fvram usage only increased 300 mb
Fimpressive
F16 bit training
