People keep asking which card to buy for running ControlNet, but the real issue isn’t how big the base model is—it’s how many preprocessors you stack. ControlNet doesn’t replace the main model, it just hooks onto it. Canny, Depth, OpenPose—each one shoves its own conditioning model into VRAM.
SDXL by itself plus one more is fine, but when you stack two or three and throw in IP-Adapter, at 1024 resolution, a 12G card starts spilling into system memory. ComfyUI starts swapping with the hard drive every time you render an image, and you know how that stutter feels.
16G is a pretty comfortable spot—you can fit SDXL, three layers of preprocessing, and IP-Adapter without hitting the ceiling. Flux is a whole other story, it’s heavier. If you’re still hanging on with 12G, upgrading to 16G makes a night-and-day difference.