Changing an alleyway into an anime image

No Python knowledge required — just press two buttons to reach the image-generator GUI. I like the interface: it’s already sophisticated, with the less-relevant controls tucked away inside “advanced” pull-down menus.

For this experiment I tried to keep the generated anime cityscape close to the original, and I only partly succeeded. The main problem: the sun shines down vertically, and the shadow lines from the roofs were often interpreted as a narrow field of grass and plants, or simply drawn as pavement. When the shadows were drawn right, another problem tipped the balance — the house on the left came out with reddish-orange walls.

This is roughly what Stable Diffusion generates with the SAI anime style, without switching the whole model to the dedicated “anime” setting. Everything skews orange-red, but the shadows on the alleyway are neatly aligned with the goal.

Here’s the closest representation of the original: the end of the alley draws the eye and the shadow lines are perfectly mirrored — yet something feels off. The original looks like a warm summer day; this one reads like winter. So even this version doesn’t mirror the real-world photo.

Another strong generation — but the orange really doesn’t fit. That makes for a nice mixed image, visualising the cornerstones that generative AI has reached. AI is pretty good, but it clearly needs more human guidance: you and I judge these as aesthetically profound, yet sparking very different feelings.

With inpainting we can fix the orange area, but the picture’s wholeness decreases — compare the leaves on the left tree and you’ll see the inpainted region looks subtly different. Below are a few other generations that show the diversity the anime-style tool has learned.

Note: the original WordPress post embedded several comparison images that weren’t part of the export. They’ll be re-added here.