AI Runner v1.17.6 fixes a major bug with LoRA.

In previous versions LoRA was in beta release. It wasn’t generally available in the Huggingface Diffusers library which AI Runner uses, and as such I had to hack an implementation together. Huggingface released official LoRA support in the latest version of the Diffusers library, so I was able to strip out the hacky solution and replace it with one that works as expected.

To give you an idea of how big of a fix this is, I have generated some test images to demonstrate what you can expect.

This first image is generated without LoRA applied.

This next example is the same generated image, this time with the LoRA beautifulEyesLikeness_halfBody in AI Runner v1.17.5 - as you can see there’s a slight difference, but not by much and the image hasn’t really improved at all.

Finally, here is the same image generated using the previously mentioned LoRA, this time in v1.17.6 with the LoRA fix applied.

As you can see, the image quality is vastly improved.

In other example we examine a 1.16 version of AI Runner

Without LoRA applied

With LoRA applied in 1.16 we get a dramatically different result

And in 1.17.6 with the beautifuleyeslikeness_halfBody LoRA applied we see a more expected result

In addition to improvements in the overall results with LoRA, v1.17.6 also addresses interface issues in which the trigger words would be lost.

Additionally in previous versions the entire model would have to be reloaded when disabling or applying a LoRA, this is no longer the case. You will notice that I have also removed the weight sliders - this is one trade-off to the new way LoRA are being applied. These may return in the future, but for now they are removed in favor of the improved functionality of this feature.