My Notes on Running AI Image Generation Locally (Stable Diffusion)
I bought a graphics card and have been tinkering with it for about two weeks. I switched to local because I didn't want to keep paying for cloud, and overall I'm satisfied.
My specs are a 4070 Ti 12GB and 32GB of RAM. At 512x768, it takes about 3–4 seconds per image. Once you go past 1024, it gets noticeably slower.
1. Keep only two or three checkpoints
At first I downloaded all sorts of them and ended up with more than 30, but in the end I only use two. They just eat up storage.
2. Write prompts in English, and order matters
Tags that come earlier have a stronger effect. I found it easiest to write in the order style > composition > details.
3. Test LoRAs one at a time
If you stack several, you can't tell what ruined what. In the end, testing them one at a time while adjusting the strength is faster.
4. Upscale with a separate tool
If you force the generation model to upscale, faces get smeared. It's better to use a separate upscaler.
My electricity bill seems to be about 20,000 won more per month. It's less of a burden than I expected.
9 answers
A 4070Ti 12GB is just right for getting started with local models, agreed.
Isn’t an extra 20,000 won on the electricity bill a bit optimistic? It doesn’t say how many hours a day it’s based on, so it’s vague. I often leave mine running overnight, so I’ve had months where it came to 30,000–40,000 won. How did you measure it?