i replied you solution
i replied you solution
Fthis is torch 2.4 error on win 11
UWhat solution? Haven't seen it.
Fopen this post
FPatreon
Get more from SECourses: Tutorials, Guides, Resources, Training, FLUX, MidJourney, Voice Clone, TTS, ChatGPT, GPT, LLM, Scripts by Furkan Gözü on Patreon
![Dazzastrous [4090]](https://cdn.discordapp.com/avatars/871389071686135838/c4a5f4d81d9eeb18d856295d692e2c07.webp?size=40)
Dfigures win 11 is crap
Fread Windows Requirements
Fi think they gonna fix it with newer torch
UI'm on windows 10.
Fwow weird
Fstill solution should work
Fi bet it is related to your folder path : "0 - KOHYA"
Ti was gonna say that, spaces in folder name
UOkay, lets try changing names.
Fmassed compute trains 3000 steps in 1 hour 
1.25$
1.25$Frecording video
Fyou need to reinstall
Finstall into
Fc:/kohya_ai
UGreat! Changed names and installed said the DLL and it works. Thanks guys.
Di have win 11 and didnt have any probs with installing it
Fgreat. i always say that dont use space char in paths
Dyeh no voids allowed
Djust facefusion
Dlol
F
D@Dr. Furkan Gözükara 20 epochs 0 resemblance to the training characters (multiple people) using regularization class person (people of all ages and genders) TE and unet lr 1e-4, Epoch 45 I'm getting some resemblance, so I increased now lr of the TE to 9e-4, lets see
Fyou need to train like 200 to see more resemblance . did you look my grids on patreon for reg images?
Frecording runpod and massed compute tutorial atm
DYes I did, mixed faces, I'm playing with the TE lr to see its effects, I want to know if something fix or minimize the bleeding, also I'm not using class on my characters
DI'm just experimenting
DTE lr 9e-4 is quite aggressive but may be is what it needs idk
Fnice let me know
UDo 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