Software Engineering Courses (SECourses)•11mo ago

Would it be possible for you to perhaps share any results? I am also looking into simpletuner and wo

Would it be possible for you to perhaps share any results? I am also looking into simpletuner and would love to see if someone has achieved some reasonable results. Would really appreciate it, I won't share ^^

Ssmoke Would it be possible for you to perhaps share any results? I am also looking int...

�

𝗡𝗶𝘁𝗶𝗻 𝗣𝗮𝗿𝗲𝘃𝗮•2/24/25, 6:04 PM

I have not yet tested it but I plan to do that next month. I have heard many people talk good about it. Here is a video that goes deep into it and compares it with ai toolkit - https://youtu.be/se3qpLkJnrk?si=7MhyT8i3DALGK9L5

YouTubekasukanra

Kasucast #26 - FLUX.1 [dev]: Training LoRA with SimpleTuner and AI-...

#flux #simpletuner #bghira #ComfyUI #LoRA #finetune #stablediffusion #style #styletraining

This is an FLUX.1 [dev] LoRA training log for art style and concepts. I document my thought process, experiments, mistakes, and analysis of quantitative and qualitative results. Hopefully, this can be a good starting guide for people looking to train LoRA...

�𝗡𝗶𝘁𝗶𝗻 𝗣𝗮𝗿𝗲𝘃𝗮 I have not yet tested it but I plan to do that next month. I have heard many peo...

smokeOP•2/24/25, 6:35 PM

So your opinion "Way better than ai toolkit btw" is based on a video? You haven't tried it yourself yet?

smokeOP•2/24/25, 6:36 PM

Interesting

Ssmoke So your opinion "Way better than ai toolkit btw" is based on a video? You haven'...

�

𝗡𝗶𝘁𝗶𝗻 𝗣𝗮𝗿𝗲𝘃𝗮•2/24/25, 6:39 PM

See the video you will know. I have done a lot of trainings with ai toolkit btw

novsky•2/24/25, 8:44 PM

can some explain how to reduce flux lora size with no (almost) quality loss? Because I've made bunch of them and 1gb loras is eating my storage) or maybe there is video with explanation?

novsky•2/24/25, 10:37 PM

never mind, found in patreon posts

Nnovsky never mind, found in patreon posts

Furkan Gözükara SECourses•2/24/25, 11:34 PM

saving as fp8

FFurkan Gözükara SECourses saving as fp8

novsky•2/24/25, 11:40 PM

yep, reduced to 260mb instead of 1GB, thx

Nnovsky yep, reduced to 260mb instead of 1GB, thx

Furkan Gözükara SECourses•2/24/25, 11:43 PM

you are welcome

Guy Bertental•2/25/25, 12:22 PM

Hi @Furkan Gözükara SECourses , do you have an idea how to create a consistent character for image generation from a single image?

GGuy Bertental Hi <@205854764540362752> , do you have an idea how to create a consistent charac...

Furkan Gözükara SECourses•2/25/25, 12:24 PM

yes 2 step training

Furkan Gözükara SECourses•2/25/25, 12:24 PM

1st traing with single image

Furkan Gözükara SECourses•2/25/25, 12:24 PM

add cropped version zoom out etc

Furkan Gözükara SECourses•2/25/25, 12:24 PM

then generate images

Furkan Gözükara SECourses•2/25/25, 12:24 PM

then train again

Guy Bertental•2/25/25, 12:42 PM

Interesting idea, I'll try it. Thanks

Guy Bertental•2/25/25, 12:45 PM

I saw this tutorial, I wonder if they do it the same way behind the scenes

https://youtu.be/UY_aRiOK6gY?si=6U4TAXFPa0lxPPU0

Guy Bertental•2/25/25, 12:48 PM

What about this process using different posing on a single image?

https://youtu.be/MbQv8zoNEfY?si=oufZgsSpJJSBMGD9

GGuy Bertental Interesting idea, I'll try it. Thanks

Furkan Gözükara SECourses•2/25/25, 2:53 PM

same logic. if you can get more poses from single image can help

Furkan Gözükara SECourses•2/25/25, 2:54 PM

but still i recommend twice training

Aikage•2/26/25, 5:03 PM

Single Image training is pretty good but will work exponentially better if you can get at least three from different angles from my experience

Aikage•2/26/25, 5:03 PM

you will save yourself a lot of time and headache

Aikage•2/26/25, 5:04 PM

If you have Comfy there are some workflows floating around which will attempt to create these angles from the image - you can randomly hit on some with a few rolls of the dice

FFurkan Gözükara SECourses add cropped version zoom out etc

Luiz xd•2/26/25, 5:22 PM

under this logic, not having so many images, could you recommend some configuration to train with a minimum of 5 or 10 images, how many steps, etc.

LLuiz xd under this logic, not having so many images, could you recommend some configurat...

Furkan Gözükara SECourses•2/26/25, 5:26 PM

it is always same the config

Furkan Gözükara SECourses•2/26/25, 5:26 PM

only number of epochs may change

Furkan Gözükara SECourses•2/26/25, 5:26 PM

like do more epoch

FFurkan Gözükara SECourses like do more epoch

Luiz xd•2/26/25, 5:33 PM

Perfect! To test then... Thanks

Guy Bertental•2/26/25, 7:32 PM

I'm looking for way to install all needed dependencies with a single script or wrap it up in a docker container for fine tuning flux model, I'm planning to do so on Azure GPU inside Azure ML training clusters. If you've got anything ready which might help me achieving this goal it would be great @Furkan Gözükara SECourses

GGuy Bertental I'm looking for way to install all needed dependencies with a single script or w...

Furkan Gözükara SECourses•2/26/25, 7:37 PM

sadly i dont have

Furkan Gözükara SECourses•2/26/25, 7:38 PM

but we have installer for runpod and massed compute

Furkan Gözükara SECourses•2/26/25, 7:38 PM

you can look runpod

Mathew•2/27/25, 2:13 PM

Has anyone encountered this error while trying to train on an A100 or H100 in RunPod? I've started four machines, and all of them give this error.

raise ValueError("Attempting to unscale FP16 gradients.")
ValueError: Attempting to unscale FP16 gradients.

MMathew Has anyone encountered this error while trying to train on an A100 or H100 in Ru...

Furkan Gözükara SECourses•2/27/25, 2:41 PM

yes people had it

Furkan Gözükara SECourses•2/27/25, 2:41 PM

i think you need to get a new pod

Furkan Gözükara SECourses•2/27/25, 2:41 PM

i trained on h100 a100 h200

Furkan Gözükara SECourses•2/27/25, 2:41 PM

didnt have that error

Furkan Gözükara SECourses•2/27/25, 2:41 PM

which template you used?

FFurkan Gözükara SECourses which template you used?

Mathew•2/27/25, 3:22 PM

This one

FFurkan Gözükara SECourses i think you need to get a new pod

Mathew•2/27/25, 3:23 PM

strange, because I've tested on 4 pods already. I was checking if GPT could help me, but the solutions it provided didn't work

Mathew•2/27/25, 3:24 PM

I ran it on my personal machine without any issues; it's only on RunPod that I encountered this error

MMathew This one

Furkan Gözükara SECourses•2/27/25, 3:34 PM

ye this is correct one

Furkan Gözükara SECourses•2/27/25, 3:35 PM

i recommend massed compute people never have errors there :d

FFurkan Gözükara SECourses i recommend massed compute people never have errors there :d

Mathew•2/27/25, 3:40 PM

i'll try, thanks

JotaTerrasa•2/28/25, 8:07 AM

Hi there! Just wanted to share with everyone some results of my latest trained Finetuning.

If you want to try it! Here is the repo: https://huggingface.co/JotaTerrasa/BZ4X/tree/main

Trigger word is: BZ4X

101 Image dataset
100 Epoch

Hardware used for training and inference:

GPU: RTX 4090
RAM: 128GB DDR4
CPU: Ryzen 9 5900XT
Motherboard: MSI MPG B550 GAMING PLUS

If you have any tips for improvement, would be greatly appreciated.

Thanks for sharing your knowledge on this @Furkan Gözükara SECourses, you rock!

MetaKan•2/28/25, 10:58 AM

I've been running tests with various training approaches for days without successful results, and I'd like to share my goal to see if someone can help me.

I'm training a model for interior design. Essentially, I need a flexible model that allows me to train a variety of products in different environments. That is, multiple furniture designs applied to different spaces such as kitchens, bathrooms, dining rooms, etc.

My idea is to use captioning to specify what type of product appears in the image.
Then, I want to be able to generate images of that specific product applied in different settings.

For a single product, it's easy—I train a model with that specific item. But for multiple product families, do you think it's feasible to group them all into a single model?
Do you think that through captioning, the model would later be able to identify each product family correctly?

Thanks a lot!

JotaTerrasa•2/28/25, 12:36 PM

I would say that lora training would be suitable for this situation, for each product and stack them with Lora stack node & divide weights proportionally.

Then in your prompt you’ll need to call to each product trigger word.

JotaTerrasa•2/28/25, 12:38 PM

https://youtu.be/hZCTiZi-EPw?si=fTg65vTJCP2tqdxX the Lora training can be done easily from this workflow via ComfyUI, if not you can use Doc’s Lora Tutorial via Kohya GUI

JJotaTerrasa I would say that lora training would be suitable for this situation, for each pr...

JotaTerrasa•2/28/25, 12:40 PM

For the stacking, you can follow this idea by Mickmumpitz: https://youtu.be/MbQv8zoNEfY?si=gP9b6zPSW5_eBy3y

JJotaTerrasa For the stacking, you can follow this idea by Mickmumpitz: https://youtu.be/MbQv...

JotaTerrasa•2/28/25, 12:40 PM

Minute 11:47

Would it be possible for you to perhaps share any results? I am also looking into simpletuner and wo

Similar Threads

Similar Threads

Similar Threads