Skip to content
janacryptoqueenPublic

About

Keeps your face and body yours in AI images. A Claude skill: one photo, a locked prompt, a canvas method for turned poses, and measured checks. Nano Banana, Gemini, Higgsfield. Free, MIT.

Topics

Resources

Contributing

Security policy

Stars

0 stars

Watchers

0 watching

Forks

Repository files navigation

Still You: AI images that are still you

Still You

AI images that are still you.

A free skill for Claude that keeps your face and your body yours when an AI tool makes a picture of you, and measures the result before you post it.

License: MIT Agent Skills Tests

Step by step tutorial  ·  The story behind it  ·  Install  ·  The proof


AI image tools love to improve you. They give you a longer face, a slimmer nose, smoother and pinker skin, and they even out your roots. The picture looks nice, and it is not you anymore.

I make a picture of myself almost every day for my posts, so I kept running into this. Still You is what I ended up with: how to prepare the photo, how to write the prompt, and a few small scripts that measure what comes back, so you stop guessing.

It does not generate images itself. It works with the tool you already use (Gemini, Nano Banana Pro, Higgsfield, others). Claude prepares your photo, writes the prompt, and then checks the result.

The proof

One photo of me, one scene, three ways. The dashed lines mark where my real mouth and chin are.

My photo, a plain prompt, and Still You

Same photo, same tool (Nano Banana Pro on Higgsfield, 4K), same scene (a warm wall with a big round glow), one run each. The pictures, the prompts and the landmarks are in examples/seated-portrait/, so you can run the check yourself.

Try What I did Eye to mouth Eye to chin Skin a* (forehead, chin) What it looked like
my real photo 9.7, 10.7
1 A plain prompt: "keep her looking exactly like she does in the photo" 4.1% off 6.0% off 19.6, 22.1 Fail. Not me. A narrower jaw and chin, a slimmer nose, redder lips, smoother orange skin, honey hair. The skin check said PINK CAST
2 The full lock on a plain crop (rules 1 to 4) 2.0% off 1.9% off 15.5, 16.9 Pass, closer. The numbers were fine. By eye the jaw and the nose were still a little slimmer than mine and the skin pinker
3 The lock, the canvas and a face close-up (rules 1 to 5) 2.5% off 0.3% off 16.3, 19.6 Pass, closest. Cheeks, jaw and nose stayed. The skin is still a little pinker and smoother than my photo (the line is +10), and the pendant on my necklace came out as an oval

Tries 2 and 3 have almost the same numbers. What separates them is the cheeks and the jaw, and only a side by side shows that. This is why the check has two halves, the numbers and your eyes.

Here is the one I kept, next to the photo it came from:

The photo and the finished picture

Install

Pick one.

1. Paste this into Claude (the easiest)

https://github.com/janacryptoqueen/still-you

Install this skill, then confirm /still-you works.

2. As a Claude Code plugin

/plugin marketplace add janacryptoqueen/still-you
/plugin install still-you@still-you

3. With the skills CLI

npx skills add janacryptoqueen/still-you

4. In claude.ai (no terminal)

Download still-you.zip from the latest release, open Customize, then Skills, and upload it.

5. By hand

git clone https://github.com/janacryptoqueen/still-you.git
cp -r still-you/skills/still-you ~/.claude/skills/

The checking scripts need Python 3.9 or newer with Pillow and NumPy: pip install pillow numpy.

Use it

Just ask, in your own words:

Make a thumbnail from this photo and keep my face exactly as it is.
/still-you  I need a 4:5 portrait of me in front of a big round pool of warm light.
Is this picture still me? (attach the photo and the picture)

Claude asks for the photo, what the picture is for and which tool you use, cuts the photo by code, writes the prompt, tells you how to run it in your tool, and then checks what comes back. If the check fails you get a retry, not a picture of a stranger.

How it works

The six steps

The method in seven rules:

  1. One photo. Several photos get averaged into a generic face.
  2. Cut it by code. Never ask the model to crop, zoom or recentre.
  3. Lock the person before the scene. Name what a retoucher would remove (roots, pores, asymmetry, tattoos) as things to keep.
  4. Never relight. Your light stays on you, colour stays on the wall.
  5. Turned or seated? Canvas first. Paste yourself where you want to be and ask only for a new background.
  6. Measure, do not eyeball.
  7. Never show a failed face. You should not be the one who catches it.

What is in the box

File What it does
skills/still-you/SKILL.md the method Claude follows
scripts/crop_photo.py cuts the exact frame (and a face close-up) from your full-size photo
scripts/photo_canvas.py puts your photo on a plain 4:5 canvas at the size and place you want
scripts/face_zoom.py zooms into a face with a labelled pixel grid so you read landmarks properly
scripts/face_check.py compares face proportions between the photo and the result, writes a proof sheet
scripts/skin_check.py measures whether the skin tone moved (pink or orange cast, relit)
scripts/generate_higgsfield.py optional, makes the picture on Higgsfield and prints the price first
assets/prompt-template.txt the locked prompt, with slots to fill
assets/prompt-template-canvas.txt the same lock, written as a background replacement
references/ every line of the lock and why it is there, the canvas method, the check, the tools, a photo pool, responsible use
examples/seated-portrait/ the photo, the three results, the prompts and the landmarks behind the numbers above, so you can re-run the check

The check, on one screen

These are real commands on the files in examples/seated-portrait/. Try 3 against the photo:

$ python3 skills/still-you/scripts/face_check.py examples/seated-portrait/source.jpg \
    "639.3,396.6 765.4,396.1 700.1,524.3 700.1,600.3" \
    examples/seated-portrait/try3-canvas.jpg \
    "520.2,723.3 603.4,721.6 563.3,804.8 562.7,856.6"

ratio          SOURCE   RESULT    DRIFT
eye_mouth       1.015    0.990     2.5%
eye_chin        1.617    1.612     0.3%

VERDICT: NUMBERS OK, under 5%. Now look at the proof sheet by eye: numbers can fail a picture but never pass one.

And the skin on try 1, the plain prompt:

$ python3 skills/still-you/scripts/skin_check.py examples/seated-portrait/source.jpg "666,286,727,322 675,557,725,583" \
    examples/seated-portrait/try1-plain.jpg "568,328,622,359 571,578,616,600"

patch            source L a* b*          result L a* b*           delta L a* b*
1            71.8    9.7   21.3      74.0   19.6   32.4       2.2    9.9   11.1
2            64.5   10.7   19.9      67.5   22.1   30.8       3.0   11.3   10.9

VERDICT: PINK CAST, a* rose by 10 or more. Retry with the skin line in the prompt (warm beige, never pink) and keep warm words out of the scene text.

The face numbers on try 1 were only borderline (4.1% and 6.0%). It was the skin check and the side by side that showed it was not me.

What I measured, and what I did not

I would rather you know the limits than find them.

  • This is one person, one set of photos and a handful of tries, mostly on Nano Banana Pro through Higgsfield. The example above is one photo and one run per try. It is not a benchmark, and your numbers will differ.
  • The numbers come from four landmarks that you (or Claude) read on a zoom grid. A reading carries about 3% noise. So the numbers can fail a picture, and they cannot pass one on their own. You still look.
  • On this photo the lock alone (try 2) already got the face numbers under 5%. The canvas gave the closest cheeks and jaw by eye, with about the same numbers. On an earlier, harder photo (seated, head turned further, not shown here) a plain crop made the face 14.0% too long on both ratios, and the canvas brought it to 2.5% and 3.3%. So the canvas is cheap insurance for turned and seated photos. It is not a step that always changes the numbers.
  • Even try 3 came out smoother and pinker than my photo (skin a* +6.5 and +8.9). The tool calls that close, because the line is +10. If your skin shifts more than that, the skin line in the prompt matters more than anything else.
  • The model made me about 13% smaller than the size I pasted on the canvas (the distance between my pupils is 286 px against the 329 px I handed it). My proportions stayed, the layout did not. Look at the layout.
  • Frontal, standing photos mostly passed with the plain crop in my earlier tests (1.7% to 4.2% drift on the ones that passed).
  • It is not a face recognition model. It does not tell you whether strangers would recognise you.
  • Image models change fast. The tool names and settings in references/generators.md may be out of date by the time you read this.

If you run it on other faces, skin tones, ages or tools and get different numbers, please show your results. That is how this gets better.

What it runs, sends and fetches

The scripts run on your machine, on your files, and the skill uploads nothing. The one optional script, generate_higgsfield.py, runs the higgsfield command you installed yourself and sends your prompt and images to Higgsfield. It never reads or prints your login. Your image tool receives your photos whichever tool you use, so read its terms. Nothing in this repo contacts any other server.

Responsible use

Use it for your own face, or for people who clearly said yes. It will not help with pictures that put a real person in a false situation or look intimate or deceptive, and it asks you to label AI images when a platform requires it. More in references/responsible-use.md.

FAQ

Does it make the image? No. It prepares the photo, writes the prompt, and checks the result. You make the picture in your own tool. The Higgsfield script is optional.

Why not just write "keep my face the same"? That is try 1. I wrote "keep her looking exactly like she does in the photo" and the face still came out 6.0% off, with orange skin. The lock names what to keep, feature by feature (try 2), and the canvas takes the placing and scaling away from the model (try 3).

Does it work for men, other ages, other skin tones? The prompts are written neutrally. I have only tested it on myself, so tell me what you find.

Do I need Claude Code? No. It works in any Claude with code execution. Without code you can paste the SKILL.md and do the check by eye, but you lose the scripts.

Can I use it with other image tools? If the tool takes a reference photo, yes. Results vary a lot between tools, so run the check on every picture.

What does it cost? The skill is free. Your image tool has its own price (a 4K picture was 4 credits on Higgsfield).

Roadmap

  • read the four landmarks automatically, so no hand reading
  • a Gemini API script next to the Higgsfield one
  • a gallery of results from other people, with their numbers
  • more tools tested

Pull requests are welcome, see CONTRIBUTING.md.

Made by

Made by Jana (@JanaCryptoQueen), cryptojana.ai. The tutorial shows it step by step, and the article tells how it came out of making pictures for my own posts. If it saved you a retake, a star helps other creators find it.

Not affiliated with Anthropic, Google or Higgsfield. Claude, Gemini, Nano Banana and Higgsfield belong to their owners.

License

MIT, see LICENSE. Take it, change it, ship it.

About

Keeps your face and body yours in AI images. A Claude skill: one photo, a locked prompt, a canvas method for turned poses, and measured checks. Nano Banana, Gemini, Higgsfield. Free, MIT.

Topics

Resources

Contributing

Security policy

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages