23
Found a way to cut my AI image generation time in half
I was waiting forever for stable diffusion to churn out images on my old laptop, like 5 minutes per batch. Then I tried lowering the steps from 50 to 25 and using a different sampler, and the results came out in under 2 minutes with barely any quality loss. Has anyone else found a trick that sped up their workflow a ton?
3 comments
Log in to join the discussion
Log In3 Comments
black.jake3d ago
That CFG thing is interesting, I never touched that setting. When you dropped it that low did you have to mess with other stuff to get the composition right or did it just work?
6
the_diana17d ago
Oh I totally feel you on the waiting struggle. "5 minutes per batch" is brutal, I used to just set it and go make coffee. What really helped me was switching from SD to ComfyUI and playing with the CFG scale - I dropped it from 7 to like 4 or 5 and suddenly my images were done in half the time with way less weird artifacts. Also found that using a lower resolution like 512x512 first and then upscaling with a separate model was faster than trying to generate at 1024x1024 from the start.
0
drew553d ago
That CFG scale trick is legit. I went through the same frustration with SD before switching to ComfyUI and realized my settings were totally wrong. Dropping CFG from 7 down to around 3.5 made a massive difference for me too, not just in speed but in actual image quality. The artifacts just vanished. Also found that using a lower res like 448x640 first then upscaling with a RealESRGAN model cut my render time by almost 70 percent. The secret is batch processing the upscales later so you don't have to wait per image. Honestly, once I dialed in those two things, the whole workflow felt way less painful.
3