SDXL GGUF, LoRAs win semi-realism

This is a workflow question more than a checkpoint question: the base model gets you into the right visual family, but likeness is usually a LoRA problem, not a GGUF problem.

• –SDXL is the sensible local baseline for semi-realistic image generation, and GGUF quantization mainly helps make it fit consumer hardware
• –Character consistency will come from LoRAs or other identity adapters layered on top, not from the checkpoint alone
• –Dark, graphic comic scenes will depend as much on prompting, negative prompting, and inpainting as on the base model
• –On an RTX 3060, lower quantization can improve speed, but the best-looking result will likely be a quality-first balance rather than the smallest file
• –If the goal is manhwa-style composition plus realism, SDXL fine-tunes and LoRAs are a better starting point than chasing a mythical "perfect" GGUF