medvram argument test

Today, I will try the medvram argument test

My computer is a laptop (MSI GP76) with a 3070 TGP140w and 8GB of VRAM

So I cannot just use SDXL. (The model itself cannot be loaded.)

The test model is protovisionXLHighFidelity3D

(It is not a specially selected model.)

After adding the medvram argument, it was successful in loading.

The prompt is the same as the previous test photo, with the full-body outfit wildcard applied.

a 4:5 aspect ratio of 512*640

Hires 1.5x was successful 

CUDA error occurred at hires 1.8x.

Tiled VAE started to be applied. Started successfully with default settings.

Hires 1.9x was successful 

CUDA error occurred at hires 2.0x

Lowered the Encoder Tile Size to 1344

Hires 2.0x was successful 

Hires 2.5x was successful 

The laptop started to lag from hires 2.5x. It seems impossible to do other additional work at the same time.

Stable Diffusion WebUI 1.5 version was meaningless to run more than hires 2.5x (I had to increase the denoising to ensure image quality, but the distortion became severe from denoising 0.5. I will blog the test when I have a chance next time.) I was curious about how SDXL would be, so I tried it again

Hires 2.6x, denoising 0.55 success to run, but it took over 8 minutes to draw a single image. Stable diffuion start to use Shared GPU Memory (guess).

Encoder Tile Size reduced to 1200

Returned to the normal speed of the 2-minute range

Attempted hires 2.7x

CUDA memory error occurred. Test terminated

I was curious how much distortion would occur with denoising 0.55, so I tried it again, but

Even at hires 2.6x with Encoder Tile Size 1200 , Stable diffusion used GPU memory and took a long time.

Even if I give up 2.6x, I decided to check only the number that can be run stably

Encoder Tile Size 1104

took a long time.

At Encoder Tile Size 1072  was low speed.

??LOL

Decreasing the Encoder Tile Size does not seem to be the answer, so I stopped here.

Decided to use the same process as sd1.5