Choosing Models for Local Image Generation: SD1.5, SDXL, and Flux Compared

If you’re putting together a rig for generating images in 2026, you’re basically choosing between these three families. SD1.5 is the old guard from 2022—lightweight, trained at 512, and the whole ecosystem grew up around it, so you’ve got the most LoRAs, checkpoints, and ControlNet support.

SDXL came in 2023, native 1024, bumped parameters to over 3 billion, and took prompt adherence and realism up a notch—it’s the sweet spot for “decent quality without insane VRAM.” Flux has a different architecture, a flow-matching transformer instead of UNet, and it’s the most powerful but also the most VRAM-hungry.

From my experience, you need at least 16GB to mess with Flux or stack LoRAs on SDXL, and 24GB+ to run Flux comfortably. If you’re on a budget, just stick with SD1.5.

SD1.5’s ecosystem is seriously complete, no way it’s getting replaced anytime soon.

SDXL is still the best bang for your buck.

Flux is great and all, but the VRAM requirement alone scares off a ton of people.

ControlNet is still the most mature on 1.5.

12G card here, barely scraping by with SDXL.

12G here too. SDXL with a bit of optimization is actually enough, but Flux’s VRAM requirement is just non-negotiable.