Hello! whats the news on Flux Dreambooth training in Kohya on 24 vram ? possible already?
Hello! whats the news on Flux Dreambooth training in Kohya on 24 vram ? possible already?
DCreate it.
Nare there any other samplers?
Nit's Flux
N
NAnd which upskellers are the best if you apply the Hires fix at the moment
Mfor generating is there a reason to use the fp8 model instead of the 23.8gb dev model? is it faster to use without loss of quality?
Mbtw this link in the kaggle notebook is not working https://huggingface.co/Kijai/flux-fp8/resolve/main/flux1-dev-fp8.safetensors so I'm downloading "flux1-dev-fp8-e4m3fn.safetensors" and " flux1-dev-fp8-e5m2.safetensors" separately from here https://huggingface.co/Kijai/flux-fp8/tree/main
NI tested it, 0 difference and moved to fp8
NErrors fail due to memory on the regular version of Dev
NI'm just training the lora on he
Moh you mean for lora training?
NYes, on instructions @Dr. Furkan Gözükara
DI have trained 3 custom Flux loras till now and everytime, I am getting scan lines in images. I use Rank 3 config (slow) with Kohya. Anyone else facing this issue? Not sure what I am doing wrong. I updated to latest Kohya version yesterday, but problem persists. Keeping Lora weight 0.8-1
DYou mean those vertical lines? Can it be an issue with something you're using to process the initial dataset?
DEven the bottom edge of the generation doesn't look clean, which makes me wonder.
DI have used 2 different datasets for custom Lora training till now. Each of 20 images
DBoth are producing such lines
DEven when I use different versions of same Lora saved at various epochs, getting scan lines
Not sure if this is because of overtraining. All images are 1024X1024
Not sure if this is because of overtraining. All images are 1024X1024
Dhere another example. Faint vertical lines are visible if you zoom in
M@Dr. Furkan Gözükara https://huggingface.co/OpenGVLab/InternVL2-8B/tree/main
this is the best captioner i found for flux, tested with 20 images, got my best results with them
this is the best captioner i found for flux, tested with 20 images, got my best results with them
M26B also nice, think could work on 4bit
DThhis is the top ranked in VLM models benchmarks
MYep even the 8b model tops many of the other ones
DDid you use any app or tool to load this model?
Mi manually used the huggingface space
Dnice, hope @Dr. Furkan Gözükara can create a set up like Cogvlm 2
DWhy does my terminated instance keep rising back up on Massed Compute? 
DGuess it finally died this time lol.
MHey any advices for training a Flux style Lora with 100 images? For now I have been using Ostris Colab.
Muse onetrainer now
Mits fixed and the loras work on forge and comfy
MAny rec. Settings?
TI can't find a solution how to generate a queue of 100 different promts in order. The context window allows you to enter a lot of text with promts where they are placed in one line, but only the last one is launched
DWhat are you using? SwarmUI generator?
DUse seperator || to separate the prompts.
D
Fyes . you can use hugging face for backup which i use
Fyou need to generate if folder missing
Fshared the prompts here , will put into our zip file as well : https://gist.github.com/FurkanGozukara/12c837939ca3fc59abc04d8bd44876ab
Gist
100 flux prompts.txt. GitHub Gist: instantly share code, notes, and snippets.
Fi think yes and hopefully gonna start tests today
Fdid you notice that in my grids?
Fi saw such thing when only i used 2x upscale of swarmui
Fdo you use system prompt? if so what is it
Mjust caption the image with details, colors, makeup, items, objects, emotions, art style, drawing style and objects but do not add any description or comment. do not miss any item in the given image, very simple sentance.
Mthey have a demo interface
M@Dr. Furkan Gözükara its quite nice actually, able to capture details better as well. i've tried longer prompts too works quite nice
Mi think it would be better than joy and cog for style training too
Xjust create a folder titled Umet in the model files inside of swarmui?
D@Dr. Furkan Gözükara Yes please see if InternVL2 is possible
