Dublin Wall Textures: Stable Diffusion vs. Flux, Same Prompt, Different Results

Last Friday, after wrapping up a client meeting in a back alley off Temple Bar, I stopped in front of a wall. Someone had stencilled a bird's face onto half-peeled brick, and rain streaks clung to it like another layer. I immediately thought I could use this as a UI background texture. The problem is that what the client wants is "a grungy texture with a Dublin street feel," but using an actual photo raises copyright issues, and I don't have time to create one from scratch.

So I decided to run two AI image generation models side by side. Stable Diffusion SDXL and Flux. The approach was to feed the same five prompts into both and compare the results.

Five Prompts, One Criterion

The prompts described textures I'd actually seen on Dublin streets, things like "cracked red brick wall with faded wheat-paste poster remnants" or "rain-streaked concrete with moss creeping through expansion joints." All five included specific materials, weather marks, and colour keywords. I ran both models close to their default settings and matched the negative prompts as well.

The criterion was just one thing. When you lay this down in Figma, can you place typography on top of it? I didn't need a pretty picture. I needed a texture to sit behind text.

Stable Diffusion: The Limits of a Familiar Friend

The SDXL results were within expectations. The brick textures were pretty decent. Details like cracked mortar lines and faded poster remnants came through as prompted. But there was a common issue I noticed in three out of five outputs. The textures were "too cleanly dirty." On a real Dublin alley wall, stains pile on top of stains until you can't tell where the original wall ends and the grime begins, but with SDXL the layers looked separated, like layer blending in Photoshop, if you know what I mean.

The moss prompt was particularly disappointing. Each strand of moss was rendered so crisply that it looked more like a botanical illustration than a background texture.

Flux: Unexpected Texture

I'll be honest, I was a bit surprised when I first saw the Flux results. Same brick prompt, but the surface irregularities felt much more natural. There was a sense of mortar applied unevenly, of brick edges worn smooth and round. Most importantly, the layers were blended together. The poster remnants looked like they had seeped into the brick texture, and that was exactly the feeling I'd seen in that back alley off Temple Bar.

That said, Flux wasn't perfect either. In the concrete crack prompt, the crack patterns came out overly symmetrical, and one image produced letter-like artifacts in the cracks. Post-processing is inevitable if you're using these for textures anyway, but text artifacts are an annoying problem.

So What Did I Use for the Client Project?

In the end, I chose Flux outputs as the base for four out of five images. When laid down in Figma with white text on top, the Flux versions didn't interfere with the text while still giving that tactile sense of "something's there." From SDXL, I only kept one brick prompt output, because its colour tones were warmer than Flux's and better suited a particular section.

The interesting thing is that neither could be used as a finished product. I took every output into Photoshop, cropped them, colour-corrected them, and processed them to be tileable. If you thought you could use AI-generated images straight out of the box, think again. It's a tool for quickly extracting raw material, not a vending machine for finished work.

Next time, I'm planning to run the same prompts through MidJourney as well. But right now it's more urgent to post-process the Flux outputs, so I'll stick with this combination for a while. These comparison notes are what I jotted down over a cup of coffee at a bakery in Stoneybatter on Saturday morning.