Do you happen to know why there's little to no improvement when using Torch 2.5? Said it a couple of
Do you happen to know why there's little to no improvement when using Torch 2.5? Said it a couple of messages ago.
UDouble-checked, I'm using Torch 2.5. [Turned off the validating requirements part in gui.bat]
Dwonder if this is a LORA trained on a single pic:
https://civitai.com/models/702897/crazy-face-peter-parker-meme-face
https://civitai.com/models/702897/crazy-face-peter-parker-meme-face
Nlong ass post was so long, converted it to image haha
Dsince you trained on 512X512 dataset, will your outputs be affected if you go for higher resolution image generation like full HD?
UAlso, does anybody have a Comfy workflow for FLUX + LORAs optimized for RTX 3090? I'm getting 15.79s/it, which strikes me as there's something wrong with my workflow/configuration.
UUsing this one.
Nall of these outputs were 1024x1024 so I don't think so? I'll try some higher resolutions now
DThanks for sharing your findings. I finished with my first character Flux Lora training today. I want to try out product lora next. Do you think it would be different from your style lora experiment? Why did you remove the word style from the target trained folder name? Does that impact the final quality?
NI removed it from the Class prompt textbox in the kohya GUI as well after I clicked 'prepare training data' and 'copy info to respective fields' and renamed the folder (not sure that textbox is used outside of preparing\copying, just wanted to be sure). I removed it because of the convos around here about training on just 'ohwx' instead of 'ohwx man'. In comparing the 25 epoch from both trainings I did, I preferred the second one that only had the instance prompt and no captions.
DThats really interesting. Also reduces the prep steps. Captioning is tedious job
Nyeah that def seems high... here's what I'm getting in SwarmUI (which of course uses Comfy as its backend) for 30 steps per 1024x1024 image in Swarm with a 2GB LoRA I trained myself...
Ndefinitely agree on captioning being tedious! I decided to try without after hearing some people say it didn't help or even hurt character likeness, and that flux apparently has its own internal captioning somehow when training?
DSharing my first batch of flux lora.
message.txt8.82KB
Ndude these are 
!! thanks for sharing - how many pics in your training dataset and how many training steps?
!! thanks for sharing - how many pics in your training dataset and how many training steps?DThanks. I had 16 pics. Rest of the configs I just imported Doc's 'rank 3' slow version
DStill let me confirm
DI think 16 pics gave me 3200 steps, took a little less than 4 hours with 1024X1024 max resolution for my base images on my 4090
USolved it! Had to update NVIDIA drivers and turnoff shared memory. Turned on the --normalvram argument on Comfy. Getting 3.2 it/s 
Dfor generating with Flux, do you guys use Forge WebUI or ComfyUI?
JForge and SwarmUI
JSwarmUI is missing in your List
Dsubstitue it with other 
D
�Check this out: https://www.reddit.com/r/StableDiffusion/comments/1f523bd/good_flux_loras_can_be_less_than_45mb_128_dim/
Reddit
Explore this post and more from the StableDiffusion community
�Pretty interesting, only trained specific layers and the Lora filesize was only 4.5mb lol
�I wish I understood this more and what tools to use to train the specific layers they mentioned here:
This LoRA was trained on two layers single_transformer_blocks.7.proj_out and single_transformer_blocks.20.proj_out
It is possible to train on blocks.7.proj_outlayer only if you don't want the dataset style transfer, the size depending on the dim can be below 4.5mb (bf16), the model is powerful enough to offer advantages such as single layer training, so let's do that.
A Flux LoRA with 99% face likeness with great flexibility can be as small as 580kb (single layer, dim 16)
Nthat's crazy
Nread the thread - no mention of how speed is affected only training on a couple layers
�I know someone like the Dr. could crack the code I bet lol
МGoogle translator.
Hi all! Successfully trained Lora for Flux. But when using it in Forge, it is ignored.
That is, with the same seed there will be the same result both when using Lora and without it. At the same time, in the console you can see that he is loading it.
Has anyone encountered the problem? How to solve it?
Hi all! Successfully trained Lora for Flux. But when using it in Forge, it is ignored.
That is, with the same seed there will be the same result both when using Lora and without it. At the same time, in the console you can see that he is loading it.
Has anyone encountered the problem? How to solve it?
NI need to reinstall because my Forge wasn't giving consistent results either. Then the console started saying something about patching the Loras? I could generate regular flux images okay in Forge but when I try to use my Lora, my 3090's dedicated memory filled and started using shared memory, so a single image was taking minutes to generate, and then it didn't look like the Lora had applied as expected - or as it does in Swarm where my images come out amazing usig Lora's I trained with Doc's configs. So I probably need a fresh reinstall of Forge. I'll test again and let you know.
M@Dr. Furkan Gözükara do i have to run accelerate config for multi gpu training?
Fit doesnt improve when using single layers methods sadly. but it improved when i was using 18 gb config
Fyes so wrong
Fswarmui way faster
Fi will look for this i think kohya can add
Falso model probably knows that person face very well already
Fjust replied this to kohya
Fcant wait to test it
Fwe set it
Fi recorded the video
Fneed to edit and publish
Vhe's awake
VDoc, when you get a chance, and it's not urgent, but I think the windows version of Instant ID gradio is currently broken. Likely do to a change in a package preventing non-safetensor files from running. Error is: "An error occurred while trying to fetch thibaud/controlnet-openpose-sdxl-1.0: thibaud/controlnet-openpose-sdxl-1.0 does not appear to have a file named diffusion_pytorch_model.safetensors.
Defaulting to unsafe serialization. Pass
Defaulting to unsafe serialization. Pass
allow_pickle=False to raise an error instead."Vthibaud does not have a safetensors version, i think only bin
Dgood news, with TE lr 9e-4 prompting "myname photo" triggers my photo its actually learning the token associated with me, I'm using regularization images from the class person, and my face is not bleeding to other men.
Dis still training
Dto accelerate the tests I increased the unet lr to 1e-4
Dmy gol is find a configuration to train multiple people on the same lora avoiding class bleeding and bleeding between the trained subjects
