Using text-to-image to batch-generate multi-resolution sliced assets is right up a frontend dev's alley.

Using diffusion models to batch-generate app assets is pretty appealing for frontend devs. A lot of mobile teams spend a ton of time each sprint manually slicing assets for different screen densities—this pain point hits close to home. Exporting @1x @2x @3x sets one by one, designers and frontend going back and forth forever just to align things. I don’t really buy those numbers like “X iterations per second” or “X times faster than the previous gen”—too much marketing fluff, and I can’t verify the architecture jargon anyway. But the core idea is worth thinking about: plugging the generation model into CI/CD so asset variants get auto-generated instead of hand-crafted.

What I’m more worried about for real-world deployment is consistency—whether details line up across different sizes of the same icon, and the determinism of the output. The last thing frontend wants is for each build to produce something different. Has anyone actually hooked this into their pipeline and run it? How many pitfalls are there?

“Determinism” hits the nail on the head—can’t have the output drifting all over the place.

Those speed multiplier numbers are just for show, don’t take them seriously.

Multi-density slicing is definitely a pain in the ass.

“CI/CD integration sounds nice in theory, but consistency is the real killer.”

For icons, vectors are the way to go. Bitmap generation is too risky.

Pulls up a chair.