SDXL
Stable Diffusion XL: the image model for about 1024 pixels with two text encoders, and the base of Pony, Illustrious and countless fine-tunes.
What it is
Released in 2023, SDXL draws at about one megapixel with much better composition than SD 1.5, and runs on cards with 8 GB or more. Its ecosystem of fine-tuned checkpoints (realistic, illustration, anime), LoRAs and ControlNets is the largest after SD 1.5. The original release had an optional refiner model; fine-tuned checkpoints rarely need it.
Files and nodes
One checkpoint (about 6.5 GB) in models/checkpoints, loaded with Load Checkpoint. Two text encoders inside (clip_l and clip_g). Empty Latent Image.
Settings that work
About one megapixel: 1024×1024, 832×1216, 1216×832, 1344×768. 20 to 35 steps; cfg 5 to 7; dpmpp_2m with karras, or euler. Prompt: tags and short phrases. Speed-up versions (Turbo, Lightning, Hyper, DMD2) use 4 to 8 steps and low cfg.
What fits
SDXL LoRAs and ControlNets (also on Pony and Illustrious, with varying results). SD 1.5 and Flux add-ons do not work. A refiner pass can be any SDXL model.
How the app helps
Shown as SDXL on model cards; names with a separate XL word are recognised as SDXL. The lesson SDXL done properly sets the size and the sampler the family was trained for, and checks that the checkpoint is SDXL.