GTX 1060 6GB runs SDXL and Pony via Windows shared memory—not just SD 1.5

A GTX 1060 6GB can run SDXL, Pony, and Illustrious models despite having only 6GB of VRAM, according to a detailed practitioner guide posted this week. The author—who started with AI image generation to translate Japanese text on visual novel CGs—was initially told by Claude Sonnet that only Stable Diffusion 1.5 would fit the card. Testing proved otherwise.

The key is Windows NVIDIA drivers' Shared Video Memory feature, part of the Windows Display Driver Model (WDDM). When VRAM fills up, Windows automatically spills into system RAM as a seamless extension. Task Manager displays this as "Shared GPU Memory." The guide explicitly targets Windows + NVIDIA users; Linux NVIDIA setups lack this fallback and hit hard CUDA Out of Memory crashes when VRAM fills. AMD ROCm support exists in ComfyUI Desktop v0.7.0 but remains unstable, and many plugins are CUDA-only.

Platform compatibility

Windows NVIDIA is the only fully tested path. Linux NVIDIA fails because the driver doesn't implement shared memory—a limitation that creates an ironic gap: Linux's lower OS overhead would make it the better low-VRAM platform if NVIDIA had ported the feature. macOS with 8GB unified memory performs worse than the GTX 1060 because the OS shares the same memory pool; higher-end Macs work but aren't the target audience.

What stands out

1. 01Model compatibility on 6GB VRAM. The author reports running SDXL, Pony, and Illustrious checkpoints via ComfyUI, contradicting common advice that SD 1.5 is the ceiling for that hardware tier.
2. 02Practical workflows over pure generation. Testing centered on image-editing use cases (text translation overlays on game CGs) rather than txt2img alone, suggesting the workflows lean on inpainting and controlnets.
3. 03ComfyUI as the platform. All testing was done in ComfyUI; no mention of Automatic1111, Forge, or InvokeAI compatibility.