is it possible to convert a fine tuned flux dev model into a lora?
is it possible to convert a fine tuned flux dev model into a lora?
Dyou can extract a lora
FPatreon
Get more from SECourses: FLUX, Tutorials, Guides, Resources, Training, Scripts on Patreon
�Say i have char1.safetensors and char2.safetensors, both fine tuned flux dev character checkpoints. What will be the result of a merge, will it be some kind of average of the two looks? i.e. if char1 has blue hair and char2 has red, the merge will have purple hair or?
Fthey will bleed / mixed
�will be images look normal, or will we see monstrous looking body parts etc 
Flook normal but mixed 
like i have in this experiments
like i have in this experimentsFMedium
Thoroughly experimented with Fine-Tuning / DreamBooth training of Flux-dev-de-distill, PixelWave v03, Verus Vision and base FLUX Dev model…
�haha, ok, worth a try then, thank you Dr. 
�ah, you've done it!
Fthis was not merge but expect similar results
�Say I have a character with a certain tattoo on left arm, would a a few close ups of only the arm with the tattoo help the fine tuned model in generating the tattoo more correctly? Ofc most images will be full-body or half-body portraits, but would 2-4 close-up pics of the tattoo help? you only see the arm and the tattoo, on the other pics you also see the tattoo, but it will be less detailed. Anyone have experience doing this?
Fi didnt try tthis
�Well it applies to faces, why wouldn't it apply to other body parts too, as long as you see it on both full body and close up pic. Could improve a lot of things maybe. If it works 
�when making the grid, does this automatically reduce the quality or number of pics in the images? there's this option "RAW resolution" under "Swarm internal", to override grid. I don't fully understand this.
Fi think doesnt reduce
�say you have character checkpoint, you want to change the body type into muscular, so you make a lora using pics of Arnold Scharzenegger. How would you make the lora affect on the body of your character, and not affect the face? I was thinking <segment:body...> <lora:arnoldbody:0.9> ohwxarnold body, but I really am confused about the order, because if you do segment body at the end, you would really need high denoise to drastically change it into muscular. So unsufre how to proceed. Ideally you don't need to apply segment, just the lora being smartly trained so it doesn't change your face, or at least very little.
Fi dont think would work at all
Y@Furkan Gözükara SECourses Hi before I press 'start training' in kohya, I jsut want to ask again, unet learning rate really that low at flux??? and no TE? haha
Ythis is 48GB_GPU_28200MB_6.3_second_it_Tier_1
Y
Fyes they are very intensively researched
Fi even shared higher lower LRs on research patreon post
Yaight will check now thanks!
Yokay ngl.... it's decent, i did lower the epochs quite a lot (u had 200 lol) because i have like 1200+ images for art style
Ffor 1200 ofc you need lesser epochs
Fmy 200 epoch is usually recommended for up to 100 images
Ffor example for 256 images best epoch i find is 70
Fbut 50 is also great
DFor style and for Flux I had success with as low as 50 good quality pics. Flux is smart, you don't need that many pics. You REALLY don't
�Using the joy caption batch, I get these at end of many of the generated captions: "<|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|><|end_of_text|>" How do I avoid those?
Freduce the caption lenght
Fi can also remove them from the generation actually
Flet me update
FPatreon
Get more from SECourses: FLUX, Tutorials, Guides, Resources, Training, Scripts on Patreon
�appreciate your work so much, Dr., I happily upgraded my patreon to Gold just now. At the same time, I will cancel some other memberships which are not very helpful
YHaha it just happened I have access to 1k+ images
Ymight as well use it
Dquality>quantity
�Is there any point in using 2048px train pics? like could I get better details in the checkpoint? or should I reduce size to 1024.
Ffor person training it didnt help me
Fbut i didnt test it after initial
Xuse multires training
SHello any tips for training/finetuning flux with a logo or can I train It the same way we do for people ?
�why multires? they're all the same res, 2048px
�Question: If I launch massedcompute with 2 a6000, can I run two trainings at the same time on that machine? @Furkan Gözükara SECourses
Xno train your 2048px images at 1024x1024 with bucket
but also train them at 512x and 768x
at least for me provides better results
but also train them at 512x and 768x
at least for me provides better results
�oh, never messed with those settings
Fexactly same way
