GTX 1060 6GB runs SDXL and Pony models — guide debunks low-VRAM myth

A user tested modern AI image models on a GTX 1060 6GB (Pascal architecture) using ComfyUI on Windows, challenging the common belief that only SD 1.5 is feasible. The key enabler is Windows' WDDM driver, which provides Shared Video Memory — system RAM automatically acts as VRAM extension when GPU memory fills. This allows models like SDXL, Pony, and Illustrious to run, though with some performance trade-offs. In contrast, Linux with NVIDIA lacks this feature, causing hard CUDA Out of Memory errors. Linux is more RAM-efficient but cannot leverage that for VRAM expansion. The guide warns: Linux NVIDIA users with low VRAM are out of luck.

For AMD users on Linux, GTT (Graphics Translation Table) memory provides up to 50% of system RAM as VRAM extension — e.g., 16GB with 32GB RAM — giving some flexibility. However, stability issues (HIP memory errors, slow generations, VAE decoding failures) and limited plugin compatibility (many ComfyUI nodes are CUDA-only) make the experience less reliable. The guide reports that AMD's ROCm support is improving but still immature. Overall, the guide offers a realistic roadmap for low-VRAM enthusiasts: Windows + NVIDIA is the sweet spot, while alternative platforms face significant hurdles.