The whole purpose of inpainting models is to provide at least somewhat better results when inpainting or outpainting an image. When e.g. outpainting an image, with the normal checkpoint models you tend to get a very visible seam between the original part of the image and the newly extended part -- this model helps eliminate that seam! Also, using a specific version of an inpainting model instead of the generic SDXL-one tends to get more thematically consistent results.

Here is an example of a rather visible seam after outpainting:

The original model on the left, the inpainting model on the right. (Yes, I cherrypicked one of the worst examples just to demonstrate the point)

Disclaimer: you definitely can get good results even without, but it's easier with an inpainting model. No need for any offensive comments about how I'm bad and should feel bad.

How to use:

Basically, the same general guidelines apply as for the original model, ie. DPM++ 2M Karras as the sampler and 30+ steps for generally good results. I like to go all the way up to 80 steps, myself, but YMMV and all that. CFG scale I usually set between 2-6 and denoising strength I usually begin with at 0.8, though I may bring it down, if I feel the model being a little too creative in the stuff it generates. May also need to bring it all the way up to 0.9 -- denoising strength has a large effect on the outcome and it may require a lot of playing around, so if you're not happy with your results, keep everything else the same and just give this setting a go.

Other stuff:

Created in Automatic1111 by merging the base SDXL inpainting model (by wangqyqq), Juggernaut XL (the non-lightning model), SDXL base and the fp16 fix VAE.