375 5 2

Furkan Gözükara

MonsterMMORPG

https://www.youtube.com/@SECourses

AI & ML interests

Check out my youtube page SECourses for Stable Diffusion tutorials. They will help you tremendously in every topic

Articles

Full Training Tutorial and Guide and Research For a FLUX Style

11 days ago

• 3

20 New SDXL Fine Tuning Tests and Their Results (Better Workflow Obtained and Published)

Aug 13

• 1

Batch size 30 AdamW vs Batch Size 1 Adafactor SDXL Training Comparison

Aug 8

• 2

Expert-Level Tutorials on Stable Diffusion & SDXL: Master Advanced Techniques and Strategies

Jun 3

• 3

Organizations

Posts 33

Post

3069

Full Fine Tuning of FLUX yields way better results than LoRA training as expected, overfitting and bleeding reduced a lot

Configs and Full Experiments
Full configs and grid files shared here : https://www.patreon.com/posts/kohya-flux-fine-112099700

Details
I am still rigorously testing different hyperparameters and comparing impact of each one to find the best workflow
So far done 16 different full trainings and completing 8 more at the moment
I am using my poor overfit 15 images dataset for experimentation (4th image)
I have already proven that when I use a better dataset it becomes many times betters and generate expressions perfectly
Here example case : https://www.reddit.com/r/FluxAI/comments/1ffz9uc/tried_expressions_with_flux_lora_training_with_my/
Conclusions
When the results are analyzed, Fine Tuning is way lesser overfit and more generalized and better quality
In first 2 images, it is able to change hair color and add beard much better, means lesser overfit
In the third image, you will notice that the armor is much better, thus lesser overfit
I noticed that the environment and clothings are much lesser overfit and better quality
Disadvantages
Kohya still doesn’t have FP8 training, thus 24 GB GPUs gets a huge speed drop
Moreover, 48 GB GPUs has to use Fused Back Pass optimization, thus have some speed drop
16 GB GPUs gets way more aggressive speed drop due to lack of FP8
Clip-L and T5 trainings still not supported
Speeds
Rank 1 Fast Config — uses 27.5 GB VRAM, 6.28 second / it (LoRA is 4.85 second / it)
Rank 1 Slower Config — uses 23.1 GB VRAM, 14.12 second / it (LoRA is 4.85 second / it)
Rank 1 Slowest Config — uses 15.5 GB VRAM, 39 second / it (LoRA is 6.05 second / it)
Final Info
Saved checkpoints are FP16 and thus 23.8 GB (no Clip-L or T5 trained)
According to the Kohya, applied optimizations doesn’t change quality so all configs are ranked as Rank 1 at the moment
I am still testing whether these optimizations make any impact on quality or not

Post

4018

Trained Myself With 256 Images on FLUX — Results Mind Blowing

Detailed Full Workflow

Medium article : https://medium.com/@furkangozukara/ultimate-flux-lora-training-tutorial-windows-and-cloud-deployment-abb72f21cbf8

Windows main tutorial : https://youtu.be/nySGu12Y05k

Cloud tutorial for GPU poor or scaling : https://youtu.be/-uhL2nW7Ddw

Full detailed results and conclusions : https://www.patreon.com/posts/111891669

Full config files and details to train : https://www.patreon.com/posts/110879657

SUPIR Upscaling (default settings are now perfect) : https://youtu.be/OYxVEvDf284

I used my Poco X6 Camera phone and solo taken images

My dataset is far from being ready, thus I have used so many repeating and almost same images, but this was rather experimental

Hopefully I will continue taking more shots and improve dataset and reduce size in future

I trained Clip-L and T5-XXL Text Encoders as well

Since there was too much push from community that my workflow won’t work with expressions, I had to take a break from research and use whatever I have

I used my own researched workflow for training with Kohya GUI and also my own self developed SUPIR app batch upscaling with face upscaling and auto LLaVA captioning improvement

Download images to see them in full size, the last provided grid is 50% downscaled

Workflow

Gather a dataset that has expressions and perspectives that you like after training, this is crucial, whatever you add, it can generate perfect

Follow one of the LoRA training tutorials / guides

After training your LoRA, use your favorite UI to generate images

I prefer SwarmUI and here used prompts (you can add specific expressions to prompts) including face inpainting :

https://gist.github.com/FurkanGozukara/ce72861e52806c5ea4e8b9c7f4409672

After generating images, use SUPIR to upscale 2x with maximum resemblance

Short Conclusions

Using 256 images certainly caused more overfitting than necessary

...

View all posts