To put even more context on that…when I first used a diffusion model, I was using the Automatic1111 frontend (well, okay, actually, I think it was a variant, but the first one I put much time into). I was using Stable Diffusion 1.5 or a derived model, can’t recall.

Automatic1111 looks more-or-less like what I think most of the cloud service systems look like. At its core, you plug in a text prompt, you click a button, you get an image.

In 2026, I pretty much only use ComfyUI, which has a directed-acyclic-graph-based system akin to what some image processing software or visual-programming systems use. IIRC, GIMP uses a DAG internally these days, though doesn’t directly expose it to users. searches Yeah, GEGL. And AFAICT, that’s generally where the momentum is — it’s not just me. That’s a lot more powerful for making changes early in a workflow and re-running it, though more difficult to learn. And from a user standpoint, it looks wildly different.

I’m not saying that there was no transferable knowledge there. You’re still working with a prompt somewhere in there. And the parameters that go to models are similar. But…you’re talking about what is effectively a new syntax for prompts. A wildly different set of limitations to work around with those prompts. Shifting from a UI that is basically an “input a string and click go” to a visual programming language.

And…I haven’t even listed some of the other of the changes. Uh, let’s see. Stable Diffusion 1.5 supports negative prompts; one can list prompt terms that should not be presented in an image; Flux does not. Flux has some sort of natural-language-processing black magic that frequently lets one use English-language descriptions to do things like “Divide an image into five separated panels. The first panel should contain…”. With SD 1.5, one would have to use some sort of plugin to subdivide an image and manually apply different prompts to different regions of an image.

Like, virtually the only thing in common between the two are sampler settings (and I use different sampler settings with each model) and the fact that both have a prompt, and there’s virtually no overlap in prompt syntax. And that was over a timespan of three years. Well, less than three years.

I can believe that you can teach someone to directly use a model. I can believe that you can teach someone to write software that uses a model. I can believe that they can produce useful output using that model. But where I have some very real problems is buying into the idea that in 2026, we’ve reached the point where the user skillset is going to be long-lasting enough that this is where one is going to get the best return. It’s like…I don’t know, training people to fly the Wright Flyer en masse:

https://en.wikipedia.org/wiki/Wright_Flyer

The Flyer design depended on wing-warping controlled by a hip cradle under the pilot, and a foreplane or “canard” for pitch control, features which would not scale and produced a hard-to-control aircraft.

You’d be right if you predicted in 1905 that heavier-than-air aircraft were going to be a very important thing in the future, but it was extremely early days then, and the control interface and the challenges that the pilot faces are probably to be very different in short order.