SDXL is the base most self-hosted image generation still runs on. It's not the newest model in the space, but it's the most reproducible one — broad checkpoint/LoRA ecosystem, predictable VRAM cost, fast enough to iterate on a single consumer card.
What it needs: ~8 GB VRAM for base text-to-image inference.
What it's good for: general-purpose stills, style/character LoRA training and inference, fast iteration. It's the workflow to reach for first if you're not sure what you need yet.
In FlixML: `sdxl_base` (text-to-image) and `sdxl_lora` (text-to-image + trained character/style LoRA).