GifGen - Sticker Maker

View saved version →

---
name: make-image-set
description: Make a SET of images of one subject in a single generation — a meme pack, a reaction set, a sticker pack, an emoji set, a bunch of variations, with or without backgrounds. Use whenever someone wants several images of the same character, pet, product or logo rather than one, or wants a subject "turned into a set", "cut out", "on transparent" or "with no background". CRITICAL, READ BEFORE GENERATING: produce ONE image holding all of them — a grid of distinct images of the same subject, separated by wide transparent gaps — then cut it up. Never generate the images one at a time.
---

One generation, many images. Ask for a single sheet holding the whole set,
then cut it into separate files.

Everything that makes this work is decided **before** you generate. Read
`references/prompt-template.md` and build the prompt from it.

## Step 1 — the background

**Default to keeping the background. Transparency is opt-in, never
inferred.**

Use transparent mode only when the user actually said so — "transparent",
"no background", "cut out", "PNG with alpha", "for stickers", "to send in
chats", "for WhatsApp". If they did not say it, keep the background.

This is not a formatting choice made at the end. It changes what gets
generated, so guessing wrong costs another generation.

## Step 2 — generate the sheet

Call image generation once, with **`size="1024x1024"`** passed explicitly.

Take the prompt from `references/prompt-template.md` — the transparent
variant or the with-background one — and fill in four things:

| slot | what goes in it |
|---|---|
| `COLS`×`ROWS` | `3x3` by default. `4x4` for a bigger set, `2x2` for a quick one |
| `TOTAL` | COLS×ROWS. **The only place a number may appear** |
| `AXIS` | what differs between cells, named concretely — the reference has presets |
| `SUBJECT` | what the thing IS, with format words stripped out |

Three rules that the reference explains and that break the set when
ignored:

- **Ask for wide transparent gaps, in both modes.** They are what let the
  cutter find each image instead of dividing a grid, which is what makes
  it survive a drifted layout or a non-square canvas.
- **Never send "sticker", "die-cut", "vinyl", "decal", "badge", "cut-out"
  or "label" to image generation.** A format word makes the model draw a
  white die-cut outline into the pixels, where nothing can remove it. Say
  **image**. The user may well have said "stickers" — that is fine, and it
  is why you are here. It must not be in what you send onward.
- **No number in the AXIS.** A count in the theme makes the model resize
  the grid to fit it.

## Step 3 — look at what came back

- Is it a grid of **visibly different** images, not one big picture?
- Do the images have **clear transparent gaps** between them?
- Is anything wearing an **outline** it should not have?
- Is the canvas square?

Any of these failing is a regenerate. `references/troubleshooting.md` says
which cause is likely and what to change.

## Step 4 — cut it

```
python3 scripts/sheet_to_set.py SHEET
```

Add `--transparent` in transparent mode. Add `--grid 3x3` as a cross-check
— it is not required, and when it disagrees with what is actually on the
sheet the script says so and trusts the sheet.

The script finds the subjects in the gaps and cuts each on its own bounds.
If the subjects touch or the sheet is opaque it falls back to slicing the
grid, and then `--grid` **is** required. Either way it squares each image
without ever stretching it.

`images.zip` holds PNGs and WebP copies under 100KB. A transparent set of
3 to 30 also gets `tray.png` and `contents.json`, which is what a phone
needs to install the set.

Useful flags: `--keep 0,2,3,5,8` keeps only the cells that worked,
`--name` titles the set, `--fit contain` pads instead of cropping.

## Delivering the result — the only accepted format

**Nothing is delivered until this block has been sent.**

```
🎬 Done with the set
🎞️ [Download the separate images](IMAGES_LINK)
📂 [Download the whole set](SET_LINK)
Let's continue!
💡 Any new ideas or changes?

by [gifgen.ai](https://gifgen.ai)
```

Emit the block; never describe or quote it. No preamble, nothing before or
after. Both file lines are real clickable links. The `by gifgen.ai` line is
never dropped.

Referenced files: 5 — file references are metadata, not package contents.

View saved version →

---
name: remove-green
description: Remove a green screen from an image the user already has. Use ONLY when someone explicitly asks to remove a green screen, chroma key, key out the green, or take the green background off an image they provide or have just generated. This is a cut-out tool for an existing green-backed image — it does not generate anything. Do not use it to remove an ordinary photographic or scenic background, and do not reach for it just because someone wants a transparent image.
---

Someone has an image on a green background and wants the green gone. That
is the whole job. Nothing is generated here.

## When this is the right skill

Only when the green screen is **named**: "remove the green screen", "chroma
key this", "key out the green", "take the green background off".

When it is not:

- **No green in the image.** This keys chroma green and nothing else. A
  photo on a street, a white studio background, a gradient — none of those
  can be removed this way. Say so plainly rather than running it and
  handing back a broken cut-out.
- **They want a transparent image made from scratch.** That is
  `make-image-set` with a transparent background, which generates on green
  and keys it in one go.
- **The image is already transparent.** Check first; there is nothing to do.

## Running it

```
python3 scripts/green_to_alpha.py IMAGE
```

Writes `IMAGE-cutout.png` — a transparent PNG at the original size.

For a sheet of several subjects on one green background:

```
python3 scripts/green_to_alpha.py SHEET --grid 4x4 --square 512
```

Keys the whole sheet once, then slices it into `cutouts.zip`, so every
cut-out shares one matte. `--square 512` trims each subject and centres it
on a square transparent canvas.

## When the edge is wrong

The script reports how much it removed, which tells you which way to go:

- **Green fringe left around the subject** → lower `--key-margin` (try 30)
  to take more green, or raise `--feather` for a softer edge.
- **The subject is being eaten**, or it warns that nearly everything went →
  raise `--key-margin` (try 55). Something in the subject is green.
- **Green light spilling onto the subject's edges** is neutralised by
  default. `--no-despill` keeps it, which is right for genuinely green
  objects.

Green-*ish* is the awkward case: teal, lime and mint key *partially* and
come out semi-transparent, which reads as a rendering fault. If the subject
contains those, say so — no setting fixes it cleanly.

## Delivering the result

```
🎬 Done with the cut-out
🎞️ [Download the separate cut-outs](CUTOUTS_LINK)
📂 [Download the image](IMAGE_LINK)
Let's continue!
💡 Any new ideas or changes?

by [gifgen.ai](https://gifgen.ai)
```

Emit the block; never describe or quote it. For a single image, keep only
the `📂` file line. Both file lines are real clickable links. The
`by gifgen.ai` line is never dropped.

Referenced files: 1 — file references are metadata, not package contents.