Regarding improving the generation time

[vivalavida1977]

【Issue】
When using Hires, it takes time at 48% and 98%.
With the following settings, it takes 6-8 minutes, but sometimes it takes 3-4 minutes. Is there a way to consistently generate in around 3 minutes?

【Environment】
OS: Windows 11
Webui: 1.9.3
GPU: RTX 4080
Storage: SSD M.2
RAM: 64GB

【Settings (except seed, the same)】
SDXL/896x1024/Sampling steps: 25/CFG Scale: 7/Batch count: 1/Batch size: 4

【Generation Time】
Normal: 00:38 / 00:35 / 00:38 / 00:37
Hires (Upscale by 1.5): 6:40 / 3:50 / 8:50 / 6:00

【GPU Utilization (Hires)】
Usage: 97-100% (Dedicated GPU Memory: 16GB / Shared GPU Memory: 31.9GB)
Normal: 10.9GB / 0.2GB
48%: 15.7GB / 3.1GB
98%: 15.7GB / 6.0GB

[Soulreaver90]

I've found changing "VAE type for decode" in settings to TAESD helped me with this.

[vivalavida1977]

Thank you!!! I improved it 😭
With Hires, I can now generate stably in around 1:30!
*Also, set "Number of Lora networks to keep cached in memory" to a value greater than 0.