Consistent character In Fooocus - kohya ss Training Lora

[@ximorro5247]

THANK YOU for taking your time to research this and teach it to us! Wow, for only 15 images you need a BIG computer... That problem about having few images in training (none in restaurant or in jeans, etc) could be solved "in reverse", I think: make the image in a general fashion and then inpaint with the LoRa to change only the face/head, so you don't generate the environment with the Lora, only the head.

[@Aininja94]

Yes, it should work. There maybe times where the body and face may not match. Maybe inpainting a larger area will work at such times. But this it's gonna get the work done.

[@nicdisan]

Great experimentation

[@indigo04r40]

What if you want to generate your own 15 AI photos/set instead grabbing them from google. Can I use those AI photos to train ?

[@Aininja94]

Yes you can, just try keeping those images identical to each other. If completely then atleast close to similar

[@Tissue_AI]

You should make FLUX workflow, no Web UI yet, since it's still on development but the branch should be working well on ComfyUI

[@Aininja94]

Flux works good on comfyUI i checked and now with fooocus as well.

[@szt3773]

SimpleSDXL2 from metercai

[@ArcosGaming]

Flux works well with forge now

[@Tissue_AI]

@@ArcosGaming yes it does, been using it since the last updates, currently still waiting for FLUX lora WebUI on kohya_ss, currently the branch one, only working with ComfyUI

[@vigiliafinal]

Nice

[@szt3773]

SimpleSDXL2 from metercai?! Hmm, nice... waiting!

[@Aininja94]

😆 you watched the video till end. Thanks, I was hoping someone does 😇

[@avobass]

Great video!! Do you know if the Config-1 file is available?

[@Aininja94]

The file is on the website, netninja.in The website is incomplete but you can get the file.

[@avobass]

Did you ever share the config-1 file and or the reg_images?

[@Aininja94]

Please watching the full video, at 21:04 I shared the files

[@avobass]

@@Aininja94 Thanks! but I noticed it after watching it over and over again. I am a novice at best. Nevertheless, I have a different problem now, after training the LORA I don't see the Safetensors file. I am seeing .json and .tomi files. How can that be?

[@Aininja94]

What are the total steps count after all the things. This could be a mistake somewhere

[@radiosiesta]

Hello, when I start Git, I always get an error in the command line. Error during Git operation. I haven't found any advice on how to fix this error. Do you have any advice? Thank you in advance.

[@Aininja94]

The error can be caused by n number of reason. Please post the error on GitHub kohya ss repository -> discussion section. I myself, and other developer will be able to see the error and could answers a solution.

[@PC_7-p4o]

when i finish it makes only ".toml" and ".json" archives, not ".safetensor" i tried with bf16 first, then fp16 and i always have the same error. My graphics are 3040 RTX. Do you know what is the problem?

[@Aininja94]

Is it a 3040 gpu? Or a typo. What are the total steps?

[@PC_7-p4o]

@@Aininja94 is a GPU 3040 rtx with 4gb. I'm using 330 training images, 40 repetitions and 6 epochs. I set 79200 steps as the maximum amount of steps.

[@Aininja94]

I never heard of 3040 rtx, as said it's a 4gb card, then there is a serious workload on the card. On a consumer card I suggest you to keep the total steps to 3000 max. 330 * 5 = 1650 then 2 epoch makes 3300 Make sure to keep the training batch 1, repetition 5. The calculation you did above is huge even for a 48gb server gpu. If not working the please decrease the images and make the total steps around 1500.

[@VendavalVendavesco]

thanks, but when i finish it makes only ".toml" and ".json" archives, not ".safetensor" i tried with bf16 first, then fp16 and i always have the same error. My graphics are 3080 RTX. Do you know what is the problem?

[@Aininja94]

What are your setting? Can you share the json file?

[@VendavalVendavesco]

@@Aininja94 Thanks a lot, but i could solve the problem. Maybe i put "many end epoch steps" thanks, have a good day.

[@PC_7-p4o]

I'm having the same problem. I'm using 330 training images, 40 repetitions and 6 epochs. I set 79200 steps as the maximum amount of steps.

[@VendavalVendavesco]

@@PC_7-p4o Did you choose "Lora"?

[@PC_7-p4o]

@@VendavalVendavesco yes

[@ebelecharles1073]

Hey man, can i do this using google colab?

[@Aininja94]

Yes you can, colab is a computer as well. Though I am not sure you will be able to do for free. Because I noticed it gets reset after an hour. The training may take more than hour.

[@giuseppedaizzole7025]

Hi there, will it work on a 12gb GPU and 16gb VRAM? thanks

[@Aininja94]

I had gone through post where people are making it work on 8gb as well but lowering down everything. It should work with one image at a time if you are talking about 3060

[@SOLOLEVELING_666]

Thx

[@CharlesFegor-ro6fb]

@@Aininja94 Hi, kindly recommend a laptop that can work

[@giuseppedaizzole7025]

@@Aininja94 Yes, thats it a RTX3060.....thanks for answering

[@Aininja94]

@CharlesFegor-ro6fb You will need a laptop with at least 12gb of vram for stuff like this. You can look for old models having rtx 3080, new models have 8gb vram or 4080 12gb will be the best but it's costly.

[@CharlesFegor-ro6fb]

please can you help with male/man regularisation images?🙏

[@Aininja94]

Need to generate multiple images of 1024p. Maybe we can ask from someone around the internet. I will try and update this thread here in 2 days.

[@CharlesFegor-ro6fb]

@@Aininja94 thank you!

[@Aininja94]

File updated

[@CharlesFegor-ro6fb]

@@Aininja94 thank you so much for your help i just downloaded, do you have a Patreon or can i reach you on whatsapp?

[@Aininja94]

I not on patreon, or any other platform. I just now started using reddit (name: No-Sleep-4069) where I try to answer questions posted by others related to AI stuff. We can connect on email if you think I can be of any help to you.

[@vigiliafinal]

what gpu are you using?

[@Aininja94]

4060ti 16gb

[@MMAHypeWatch]

Why fooocus instead of defooocus?

[@Aininja94]

The viewers I have use fooocus since my first video. You can use the lora for and user interface, it should work