I’ve been playing a lot with Stable Diffusion. A lot.

I got started like most of us toying around, with Automatic1111. But soon I found ComfyUI. The control, the flexibility. chef’s kiss

Now like anyone else I found limitations and extremely painful solutions. But one thing I discovered when idly testing a workflow was this: Stable Diffusion (SDXL Turbo, specific to my testing) is REALLY GOOD at generating cats.

A cat sitting in grass, facing the camera.

This distinguished little gentleman is just the first image I happened to grab from my output folder. I can easily notice a few things that seem “off” here… Mostly how his poor lil’ spine is getting along back there. But hey, it’s a (non)random sample.

A portrait of a cat walking toward the camera outside.

Ok… I’m noticing a pattern here.

A portrait of a cat with its head tilted.

Yep, ok. All their lil’ ear tips have this fluffy black end.

When I saw these results my first thought was “Oh I get it, we fucking LOVE cat pics!” so — as I understand it anyway — naturally, it’s easy to train on a plentiful and diverse data set.

The quality of the cats is really generally GOOD. My guy up here may have ‘ole righty looking off to the distant future a bit, but it’s much harder to get a weird, mutated result out of a stable diffusion model whereas I sometimes spend an hour or more trying to get a new subject really perfected in prompts, strengths, and other little node weirdness.

I only tried this on two models: DreamshaperXL_Turbo and ultraspiceXLTURBO (Careful! Those models aren’t “safe”, I hate restricted models and that goes with the territory). I thought the slope of this guy’s eyes was off, and it may be a little extreme, but I searched around a bit and found that this angle is actually pretty normal. I’d say the very bottom of his left eye pupil looks pinched. But, otherwise, pretty solid. Yet the TUFTS remain.