Local FLUX, SDXL Stack Eyes 16GB GPUs

Hot take: 12GB is a “can make it work” tier, not a “stop thinking about VRAM” tier. If you want FLUX to feel usable instead of constantly offloading, 16GB is the first sensible buy, and 24GB-class cards are where local image gen starts feeling genuinely roomy.

• –ComfyUI’s FLUX docs describe the full model as VRAM-heavy, while fp8 checkpoints reduce memory at a quality cost; the FLUX.1-dev model card discussion also points to roughly 21.5GB VRAM for the full 12B model.
• –Z-Image-Turbo is explicitly positioned for consumer hardware, with its own docs calling 12GB the native BF16 floor and 16GB the recommended comfort zone.
• –SDXL is the easy part here; the real pressure comes from FLUX plus encoders, LoRAs, ControlNet-style additions, and your desire to queue 2–3 jobs without the system thrashing.
• –A 12GB card can be fine if you accept queue-first behavior and lighter model variants, but it is not the right choice if “real-world usage” means frequent FLUX runs with minimal friction.
• –If budget is tight, 16GB is the pragmatic midpoint; if you know local image gen will be a long-term hobby or tool, a 4090-class card is the boring-but-correct headroom play.