A 90 second loop in ComfyUI finally made stable diffusion click for me

I spent weeks messing with different workflows and never getting consistent character renders, but then I found a video showing how to wire a simple IPAdapter plus ControlNet tile upscaling loop and now I can keep the same face through 20 generations. Has anyone else found a single node setup that unlocked a whole new use case for them?

3 comments

3 Comments

emma963d ago

I mean, that tile upscaling loop is exactly what did it for me too. Once I got that running consistently, I could take a rough face from one generation and actually keep it recognizable through different poses and lighting. The thing that really unlocked everything though was adding a simple noise injection node right before the KSampler, just a tiny amount like 0.05. That little bit of chaos stopped the faces from going all soft and blurry after a few loops, but the IPAdapter still held onto the identity.

thomas_price3d agoMost Upvoted

36 years in photography and I always thought the whole AI thing was just a gimmick for making weird cat pictures. That tile upscaling workflow changed my mind completely last week. Now I can actually use it for client mockups instead of trawling through stock photo sites.

josephbailey3d ago

The tile upscaling stuff really is the gateway drug for people who were skeptical. I had a similar moment when I figured out how to chain a couple of ControlNets together for consistent character sheets. @thomas_price, if you ever need to keep that mockup consistent across different styles, try feeding the noise injection right after the tile upscale but before the Detailer, it keeps the texture from washing out without making the face look stiff.