What's this?

This workflow upscales and upscales and upscales your image into ultra-high resolution, then slices it into 512×512 tiles. Each tile is processed via image2image, resulting in a photomosaic-style output.

This is an old-school trick that's been around since the early days of Stable Diffusion. However, thanks to this brilliant idea — “Add pixel-space noise to improve your doodle-to-photo results” — we can now generate highly diverse outputs even with low denoise values.

Huge thanks to YentaMagenta for sharing the insight! ☺️

I’ve uploaded a sample of the generated output here. (⚠️ Please note: the file is 178MB.)

Workflow Summary

• 🟨 Upscale the input image by 4× using an upscale model, then crop it so the resolution is a multiple of 512. After that, upscale it again by 2×. (Of course, feel free to go even bigger if your PC can handle it 🤤)
• 🟥 Add noise — the only truly original part of this workflow — before feeding it into the image2image process.
• 🟩 The rest of the workflow is quite similar to Ultimate SD Upscale, but without overlap or ControlNet Tile, to preserve the distinct tile boundaries.
• 🟩 Add wildcards to apply a different prompt to each tile, enhancing variation and style diversity.