but you have to enable it in the advanced parameters
but you have to enable it in the advanced parameters
�Okay
Fi will update configs hopefully soon
�this right?
Mcorrect and enter this into your additonal parameters, cause i think the cpu offload checkpointing is not working.
--cpu_offlload_checkpointing
--cpu_offlload_checkpointing
�Can I DM you regarding one more thing?
Mif its feet pics no, other than that then yes 
PIs it necessary to train using resolution 1024x1024? For example, my image is a rectangle with a ratio of 768x1024, is it optimal for training or do I have to crop it to a square?
DFor the Kohya workflow, best results are achieved with the square ratio, which is the standard recommendation, as far as I know.
PIn the data train, there are many image sources with different resolutions, is that okay? Or should we just get 1 fixed resolution for all train data?
DYou can easily convert various aspect ratios to square size using outpainting. I use 'Expand' function of Adobe Firefly Generative Fill online - fastest and easiest.
Fi suggest you to first train 1024x1024
Fonce you have some base results
Fyou can try different aspect ratio multi res and compare
F1024x1024 most safe approach
Fit appears that onetrainer has dora for flux
Ftime to search it :d
Dhttps://github.com/comfyanonymous/ComfyUI/issues/4695 there have been some issues reported with loading DoRA files trained with OneTrainer in other tools like ComfyUI
GitHub
Feature Idea It looks like the Flux branch for OneTrainer is nearing completion - and the results are looking great! However, it seems as though ComfyUI cannot load the resulting .safetensors file....
Zif we have 12GB of vram, do we simply run the rank5 or 4 inside of kohya? do we still use flux dev considering that is a much bigger model? @Dr. Furkan Gözükara
Fi think he fixed already?
Fyou have to use rank 5 or below
Fuse dev model it auto cast
Fmake sure you have virtual RAM
Zoh ok gotcha, i have 12 GB of vram and 64 GB of system ram, i am watching your video right now and about to start training. I went outside and took some better pictures with the natural lighting outside
ZThank you
DHow much is quality loss in the Rank 4 fast version, compared to Rank 3 slow version?
Z@Dr. Furkan Gözükara I'm still getting OOM errors when trying to train a lora on 12GB, i'm all the way down to rank 8, i have 11.2 GB of Vram available and still getting it
TAnyone else having trouble with Forge? I'm using the same old thing in Flux and I get OutOfMemory error in Massedcompute with the 48gb
TI think it's when patching Lora, but it worked perfectly a few days ago 
Zcan we use our lora with gguf quantized models of dev?
Tthe problem is with --always-gpu @Dr. Furkan Gözükara
Zwhat's the recommended way to run inference with low gpu cards? we can't use flux1.dev
KTrained a lora with AI-toolkit and the outputs look good, but when I try to run it in Forge it doesn't at all look like the results in Ia-toolkit :/
KWhat am I missing
KI have trained on fp16 dev and are usiing dev nf4 in forge. Is that why?
KCan't trained LORA's be run in Forge? -
Fwhich config you using?
Fi havent tried yet
Fbut swarm ui should work out of box with fp8
Fnf4 is the problem
Fmost optimized
Rhey @Dr. Furkan Gözükara using swarmui grid generator, if run a grid gen with two params, and you rerun the same params but add an extra value, does it overwrite the entire previous gen? I was under the impression that simply adding a value in an existing param, just inserts the extra gens into the existing grid
Fi am not sure
Fi think it has mixed behaviour
Rits really comfusing. using on runpod. ive tested on smalll images to see the mechanism, and it seems sometimes erratic, unpredictable. i am almost sure this the case
Rbut one thing i d like to know is, if i interrupt the grid genartion, can i re-run it? wi ll it skip already generated images? Or does the interruptioon corrupt the files? Surely you've tested and experienced the behavior?
Fit can resume