Ok LoRA research is in progress, and Full Fine Tuning research is planned?
Ok LoRA research is in progress, and Full Fine Tuning research is planned?
Fafter lora
RAre you now fully just using swarmUI and no more A1111?
Fcurrently for flux yes
Fthe realism is just great and i am not even cherry picking or best config yet
R@Dr. Furkan Gözükara, how fast is FLUX with your graphics card? Which graphics card is good – is it the one with more VRAM?
I am trying to decide between SDXL Turbo or FLUX...
I am trying to decide between SDXL Turbo or FLUX...
FKohya SS GUI FLUX LoRA experiments - The realism is just great and i am not even cherry picking (these are from my x/y/z grid testings for research) or best config yet (still lots of test to do and evaluate) For stylization it is still super overfit. Trying to fix that as well
F
Fyou need 24 gb gpu atm
For cloud
FPatreon
Get more from SECourses: Tutorials, Guides, Resources, Training, FLUX, MidJourney, Voice Clone, TTS, ChatGPT, GPT, LLM, Scripts by Furkan Gözü on Patreon
Ftesting flux cfg now
Jthis it happened for me too in SwarmUI, the Comfy guys are aware of this problem
Fwow this is a big bug
Fit is certainly a bug
Jyes sure, it happpens if you create alot of pics without stop
DI changed to a different workflow and that fixed it
Fnice
CRegarding style overfitting
what did you tag your photos with? did you tag taken on smartphone or something like amateur photography?
Also have you tried adding back in generated photos of yourself (the ones with good likeness) that dont have that standard style. Or are you only experimenting with how good a model can do if you have very basic dataset
Finally another thing i found to fix style is generating first few steps with 0.4 Lora weight, then generating the remaining with 1.0 weight. Since the first few steps are dictating colors and style it gives it much better style that is closer to what original model would have produced
All the above has helped my loras be much more versatile being able to generate much more natural looking images while retaining likeness
what did you tag your photos with? did you tag taken on smartphone or something like amateur photography?
Also have you tried adding back in generated photos of yourself (the ones with good likeness) that dont have that standard style. Or are you only experimenting with how good a model can do if you have very basic dataset
Finally another thing i found to fix style is generating first few steps with 0.4 Lora weight, then generating the remaining with 1.0 weight. Since the first few steps are dictating colors and style it gives it much better style that is closer to what original model would have produced
All the above has helped my loras be much more versatile being able to generate much more natural looking images while retaining likeness
Fonly ohwx man atrm
Fhow do you do that=? i am gonna ask swarmui :"steps with 0.4 Lora weight, then generating the remaining with 1.0 weight"
JDr.Furkan i can not find the menue in swarmui for the extra args settingsm any quick hint?
Fdisplay advanced
Ji meant the ExtraArgs --fast you just posted under tips-info
Fit is in backends
F
Fi think speed up fp8 significantly
Jaaah thank you!
Flol some configs are pure noise
Fthis one nice but resemblance reduced
CI'm using comfyUI and the advance sample lets you choose start and end steps, and so i have two samplers in series, one takes lower model weight and passes the output latent to the other higher model weight sampler
Fi just told this to swarmui
Cnever used swarm ui, didnt realize it can take natural language configs
Fno i mean the developer :d
Fout of 7 new testing 3 of them are noise
Fjust got a better config out of new tests amazing 100 epoch 1500 steps
Freg images killed likeliness hugely
Ni was noticing that as well - bottom row of Dr. Furkan's grid, 2nd from right - best, 3000 steps, with reg images - bears the worst resemblance
Fyep
Fi am gonna update patreon with best new configs
F
Jthe problem happens not only on SwarmUI even on Forge for me. Other reporting it too, like here: https://github.com/mcmonkeyprojects/SwarmUI/issues/192
GitHub
Expected Behavior Subsequent 16-bit images generated with a LORA in Flux dev should not degrade into fuzz. Actual Behavior When using any LORA, the first image generated after switching to Default ...
Freg images improves environment but likeliness is gone in some cases - 2225 step
Vdrop your prior weight loss to 0.10
Vand more steps
Fnice idea should test
Vreg images not reccomended for flux, but 0.10 and lower prior loss and higher lr ideal
Fjust testing atm. so far got great results and getting even better with finding better params
Vsame as well, been getting amazing results with finetune/dreambooth as well
Vlora is sort of crap because of the different alphas and conv needed