With the permission of @ORAKUL_STUDIO I am posting these variants of the model located here - https://civarchive.com/models/2942239/flux2-dev-aivazovsky-xxi-rank-1280-experimental-lora-8b-parameters

These are smaller version using SVD optimization.

Three models 99.95%, 99.5% and 99% retained.

This was done as an experiment, if you'd like to donate please donate to @ORAKUL_STUDIO for as the only work done by me is running a customized script to compress the model. (I don't want donations for this.)

Heatmap image compare shows that the 99% is actually more accurate to the original than 99.5% and 99.95%.. Here are the heatmaps. The less red the better.

90.00% <- Closest matching by far.

95.00%

99.00%

99.50%

99.95%

A different analysis confirms that 99% is closer to the original.

An "Absolute difference" at 0% meaning even if it's a color difference of 1, (say original is 255x255x255 and the reduction is 255x255x254 it counts)

90.00% 22794 Error Delta
95.00% 26182 Error Delta
99.00% 24392 Error Delta
99.50% 24628 Error Delta
99.95% 25199 Error Delta

To give you an idea - There is 4,227,072 pixels on this image. And this is say that 22794 are different compared to the original. A 0.54% difference.

It's very clear that certain optimizations are more accurate, and you basically need to hunt and fish for the right one. The fact that a model that is 13GB smaller produces an image that is just 0.54% difference shows that optimization is something to consider.

I am not sure why 90% is so much more accurate than the 99.95%

It's sooo close that on visual compare, it's 100% color shade. There is no items missing like at the 99.5% it completely removed a chimney... At the 99.95% the window is shorter...

The 99% has 1 extremely small, and can only be seen when you zoom in. That red bar in the middle of the image is a 1 pixel height short run of the, I'm guess a wall. It basically straightens it out from the original. The original is not straight LOL

Recommended Settings: Parameter

Value Base Model FLUX.2 Dev (~8B)

LoRA Weight 1.0

CFG Range 0.0

(Classic/Glazing) 3.5

(Detailing)

Sampler: dpmpp_3m_sde_gpu

Scheduler: linear_quadratic

Steps | 30 – 60

Prompt Prefix oil painting,