Surreal Pop Collage: an agent skill that turns one photo into a desaturated pop collage

The problem: photo-to-collage prompts drift without a rule set

Anyone who has asked an image model for a surreal collage knows the failure mode. You get a gradient sky, three impossible objects, a whale, and a dolphin. The result reads as noise because nothing in it is strange relative to anything else. This skill exists to constrain that. It is not an image generator and it is not a Photoshop action. It is a written rule set plus an agent entry point that makes a generator produce one specific look: the photo desaturated to black and white as a reality anchor, the background replaced by large flat matte color shapes, and a single impossible giant element that has to be traceable back to the source photo.

The audience is narrow and worth naming. First, people running an agent coding tool such as Kimi Code, Claude Code or Codex who want the skill to fire when they say something like "turn this photo into a surreal collage". Second, people who never install anything and just want the prompt recipe to paste into ChatGPT, Midjourney or Jimeng. The README explicitly supports both paths, which is unusual for an agent skill and probably the reason it is packaged as a SKILL.md plus an agents directory rather than as a library.

How the four-part prompt is assembled

The mechanism is a fixed prompt skeleton with four variable slots. The README calls it the four-part structure: the recognizable but desaturated subject, the flat background color shapes, the single giant element, and a group of small elements in graduated sizes following an arc. The skeleton also pins composition (vertical 3:4), surface treatment (flat matte colors, no gradients), and two negative constraints (no text, no watermark).

The interesting part is where the variables come from. Before writing the prompt, the user writes three lines: what the subject is, which two colors dominate, and where the scene is. The background shapes are then derived from those dominant colors by one of three moves the README lists: purification, complement, or emotional inversion. The giant element must also come from the scene, chosen by one of three routes: enlarge a small object that is already in the frame, pick the semantically most distant object, or invert scale. Cultural landmarks are handled separately: the giant comes from the landmark's own reference, but it has to be translated into a visible object or action first. The README's example is a mountain-and-canal landmark where the reference becomes a lantern rising into the moon.

This is a derivation chain, not a style tag. The color shapes and the giant both have an audit trail back to the photo, which is what the rule "elements come from the image, colors come from the image" means in practice. It also means the skill cannot work on a photo you have not looked at, because the three input lines require a human read of the frame.

Installing the skill into an agent skills directory

The README gives one install path, and it is a file copy rather than a package install. There is no npm, pip or Homebrew step documented. You copy SKILL.md and the agents directory into your agent tool's skills folder. The README's example target is a Kimi Code skills path:

cp -r SKILL.md agents ~/.kimi-code/skills/surreal-pop-collage/

After the copy, the README states the agent will call the skill automatically when you ask it to turn a photo into a surreal collage. Substitute your own tool's skills directory if you are not on Kimi Code; the README names Claude Code and Codex as other supported tools but does not give their paths.

For the no-install path, the README says you can hand-write the prompt in any generator. The starting skeleton it gives is this:

surreal pop collage, vertical 3:4,
keep the [主体] clearly recognizable but desaturated to black and white,
the background replaced by huge flat matte color shapes: [色形与颜色],
one impossible giant element: [巨物],
[小元素群] in graduated sizes following an arc,
a few white hand-drawn graffiti strokes,
flat matte colors, no gradients, no text, no watermark

The bracketed slots are the ones you fill from your three input lines. Read the output against the quality gate in SKILL.md rather than against your expectations: the visible check is whether the subject is still recognizable after desaturation and whether there is exactly one impossible element in the frame. The README does not document what the agent does when the gate fails, so treat a bad first render as a prompt edit, not a retry.

The one-giant rule is the whole design, and also its ceiling

Every other rule in the skill serves the single-giant constraint. Flat matte colors with no gradients exist so the desaturated photo reads as a different material from the background. No text exists because the README states image models garble Chinese characters, so titles are meant to be typeset afterward in code. Elements must be sourced from the photo because a default red sun or blue sky would be a second surreal intrusion, and the README bans those defaults by name.

That discipline is also the limit. A rule set built around exactly one anomaly is a poor fit for anyone who wants layered, dense, many-object collage work in the style of classical cut-and-paste surrealism, where the density is the point. It is equally wrong for anyone who wants the photo left intact. Desaturation is not optional here; it is the reality anchor the whole contrast depends on. And because the giant must be traceable to the frame, photos with almost no usable small objects, a plain portrait against a blank wall, for instance, give the skill very little to work with. The README acknowledges the sourcing requirement but does not offer a fallback for thin source images.

The repository itself is small: LICENSE, README.md, SKILL.md, an agents directory and an examples directory. The examples directory currently contains only a README file, so the three worked examples the main README promises (landscape, city, portrait) are described rather than shown in the file listing. The last push to the default branch was on 2026-08-08.

Compared with doing it by hand in Photoshop

The obvious alternative is manual collage: cut the subject out, desaturate it, paint flat shapes behind it, drop in a scaled-up object, and add hand-drawn strokes. That workflow gives you exact control over edges, color matching and typography, and it is the standard route for anyone searching for how to make a surreal collage in Photoshop. The difference in approach is where the decisions live. In Photoshop the decisions are made with the lasso and the color picker, one image at a time. Here they are made in text, before generation, and the generator interprets them.

The trade is real in both directions. A text recipe is fast to iterate and easy to keep consistent across a series, which is why the four-part skeleton is written the way it is. It also cannot guarantee a clean silhouette, and it cannot place a title, since the recipe forbids text in the image. Manual work wins on the last ten percent of finish. The recipe wins on the first eighty percent of volume.

Licence, upgrade cost and what the repository actually ships

The licence is MIT, stated in the README and present as a LICENSE file at the repository root. MIT is permissive: reuse and modification are allowed with the licence and copyright notice retained. That is the extent of what can be said here; the repository does not discuss attribution for generated images, and this is not legal advice.

Upgrade cost is close to zero in the technical sense and non-trivial in the practical sense. There is no package version to bump and no dependency graph, so updating means re-copying SKILL.md and the agents directory over your installed copy. The cost is that any local edits you made to those files are overwritten, and the README does not document a merge or override mechanism. If you plan to adapt the rules to your own subject matter, keep your edits in a separate file rather than inside the copied SKILL.md.

The maintenance picture is thin rather than worrying. No releases have been retrieved, the examples directory is not yet populated, and the README itself says more examples are being added. Treat the prompt skeleton as the stable part and the examples as unfinished.

Editorial conclusion

Adopt it if you already run an agent tool that loads skills from a directory and you want a repeatable structure for surreal collage prompts rather than improvising each time. Do not adopt it if you need a rendering pipeline: the repository ships prompt rules, not code, and the actual image comes from whatever generator you point at the prompt. Before committing, open SKILL.md and check whether the decision priority order and the correction table cover the photo types you shoot, since the README only shows the 30-second version and the examples directory currently lists a single README file.

Frequently asked questions

Official sources

1. 2998980-hue/surreal-pop-collage on GitHub
2. Issues
3. License: MIT
4. README