Targetting a few layers?
Targetting a few layers?
RWhat's the size of the LoRA file?
Fnope fully layers
F128 rank
Ffp16 saving : 1.27 GB
RThat's a small file.
RResemblance is good?
ROh, and how fast is the H200?
Freally fast
Fnot too bad but lower quality here
F
Fleft ours right faster
Fyou see faster is overfit to get decent resemblance
F
F
Fi will try 2x gpu next with 1024px
.
�Thanks but I want to know exactly how are you doing this
RYeah. The original one is excellent. How many images did you use?
FPatreon
Get more from SECourses: FLUX, Tutorials, Guides, Resources, Training, Scripts on Patreon
F28 images
Fit is 170 th epoch
Mhehe
.I've written python scripts with chatGPT that will process the high resolution images for training by cutting them into 1024x1024 pieces with a bit of overlap.
TAmazing, can you upload the model to tensor art please?
.Sure!
.0 runs, 0 stars, 0 downloads. This fine tuned checkpoint is based on Flux dev de-distilled thus requires a special comfyUI workflow and won't work very well ...
.Little comparison to see the improvement side by side.
Syou recreated this? holy moly! do you have a breakdown on how you did this?
Mnot really the higher the resolution the longer it takes to render obviously - and you get better results but also diminishing returns eventually - this was resized to 1080p i believe took 5min to render
Mi sent the workflow above - input video - play with the prompt and setting, make sure to match the skip steps with the normal scheduler step to equal 20 step of genning anything under that breaks teh genning
Mand prompt basically everything betwheen the both videos i wish theres an llm that can do that for me to be more accurate
WI already test some top Flux dev de-distilled model, quality quite dissapointed, most of them only good on realistic image, for the style only the same or even worse than Flux dev although time for render is double due to CFG >1. Are there any de-distilled models you guys can recommend and maybe can be used to train lora and finetuning ?
SAwesome, thanks a ton Maki
Myep
Mand whats the best tool for lip sync?
HWhat would you guys recommend for a flux checkpoint training on Kohya for a character model.. i have 40 images on a 3090ti (24 gb, 64 gb of ram). More than batch size 1? Is current best practice not to caption our training sets and to use ohxw woman? Things change so fast in this world?
A@Dr. Furkan Gözükara I'm currently using Flux Q8 on Kaggle , is there something better I can use for multi purpose? (realism,style etc). I saw in chat something about de distilled versions?
Fnothing i know specific
Fbatch size 1 is best
Fwe have 24 gb config do fine tuning / Dreambooth
Fnot that i know
Ftesting faster training on 8x A6000 machine
Aany suggestions?
TIs Hunyuan current state of the art img2vid model?
Tin both open-source and SaaS worlds?
GAnyone know the easiest way to finetune your own VLLM for image captioning? I've already got my own dataset, but there doesn't seem to be a straightforward way to actually carry out the finetuning process itself... I know there's already some good captioners out there, but I want to finetune my own
T@Dr. Furkan Gözükara I've noticed that SwarmUI downloads t5xxl_enconly.safetensors file every time you try to use flux even if you set t5xxl model location manually.
Furthermore there is a mention in debug logs that Comfy back-end is typecasting fp8_e4m3fn model into bf16 every time flux type checkpoint loads from scratch, and that typecasting takes significant CPU time.
I looked up in swarmUi code and it seems that it's downloading model from this repo: https://huggingface.co/mcmonkey/google_t5-v1_1-xxl_encoderonly/ That mdel is 4.9 gb that is consistent with VRAM usage of swarmUI when running flux inference.
My concern is that all this time we used the inferior fp8 casted t5xxl model for generations, wich according to many reports decreases output quality much more than casting to fp8 flux itself
Furthermore there is a mention in debug logs that Comfy back-end is typecasting fp8_e4m3fn model into bf16 every time flux type checkpoint loads from scratch, and that typecasting takes significant CPU time.
I looked up in swarmUi code and it seems that it's downloading model from this repo: https://huggingface.co/mcmonkey/google_t5-v1_1-xxl_encoderonly/ That mdel is 4.9 gb that is consistent with VRAM usage of swarmUI when running flux inference.
My concern is that all this time we used the inferior fp8 casted t5xxl model for generations, wich according to many reports decreases output quality much more than casting to fp8 flux itself
TIt also checks model file hash so just replacing file with the same name would't work