gooseworks-ai/goose-skills / skills/video/create-image-fal
create-image-fal
VerifiedGenerate or edit an image via any FAL image model (nano-banana edit, gpt-image, flux, ...), routed through the GooseWorks proxy, so the call is charged to your GooseWorks credits. image_urls must be public URLs; --image puts a local file on the fal CDN first. The recipe names the model + prompt. Use for keyframes, flat-cover transforms, product hero edits. (For the OpenAI gpt-image family specifically, create-image-gpt-image-fal also exists.)
| File | Kind | Size |
|---|---|---|
| SKILL.md | skill md | 8.0 KB |
| references/avatar-generation.md | markdown | 9.0 KB |
| scripts/brand_pack_lint.py | script | 4.4 KB |
| scripts/gen_image.py | script | 6.6 KB |
| scripts/image_quality.py | script | 1.5 KB |
| scripts/media_proxy.py | script | 56.8 KB |
| skill.meta.json | config | 177 B |
| tests/smoke-test.md | markdown | 280 B |
| tests/test_brand_pack_lint.py | script | 4.8 KB |
| tests/test_image_quality.py | script | 1.2 KB |
SKILL.md
View on GitHub---
name: create-image-fal
description: Generate or edit an image via any FAL image model (nano-banana edit, gpt-image, flux, ...), routed through the GooseWorks proxy, so the call is charged to your GooseWorks credits. image_urls must be public URLs; --image puts a local file on the fal CDN first. The recipe names the model + prompt. Use for keyframes, flat-cover transforms, product hero edits. (For the OpenAI gpt-image family specifically, create-image-gpt-image-fal also exists.)
status: active
---
# create-image-fal
**GPT Image 2 quality follows the customer's tier.** `gen_image.py` always sends an explicit quality: `--tier standard` (the default) is `medium`, `--tier best` is `high`; a payload's own `quality`, or `--quality`, wins. Above the tier is a one-shot upgrade, only after a failed check at the cheaper setting or when the customer asks (see create-creator-takes-h3's quality tier). `--query '{"shot_scene": 3, "upgrade_reason": ...}'` sends the call's GooseWorks query fields.
**Never hand-write a person prompt.** Any image with a person in it is made by create-creator-takes-h3's `scripts/make_character.py` (GPT Image 2, phone capture, deep focus, no retouching), unless the recipe prescribes its own creator still; read [the avatar-generation guide](references/avatar-generation.md) first. Nano Banana is for cartoons only, never for real people or real products, and never use Flux for a face.
Generate or edit an image via any FAL image model (nano-banana edit, gpt-image, flux, ...), routed through the GooseWorks proxy, so the call is charged to your GooseWorks credits. image_urls must be public URLs; --image puts a local file on the fal CDN first. The recipe names the model + prompt. Use for keyframes, flat-cover transforms, product hero edits. (For the OpenAI gpt-image family specifically, create-image-gpt-image-fal also exists.)
## Run
gen_image.py --model fal-ai/nano-banana/edit --payload '{...}' --out keyframe.png — charged to your GooseWorks credits; returns the *.fal.media URL.
## The brand's own product
**Never ask for a blank pack of the advertised product.** "No brand text in a prompt" bans *typing*
label text; it never means removing the product's real label. A test video asked for a pouch with
"its unprinted back facing the camera" and "no printed words", sent no product photo, and the pack
in the hands read as a different product.
- **Refused words.** When the object is the brand's product, a prompt with "unprinted", "blank"
(pack, pouch, box, label...), "unlabeled", "no printed words" (or "without a label") or "back facing
the camera" is refused. Pass `--brand-product static` or `--brand-product held` and
`gen_image.py` runs `scripts/brand_pack_lint.py` before any upload or paid call.
- **Static or flat:** on a counter or shelf, the real photo usually goes on afterwards with an overlay.
- **Held, opened, poured or used on screen:** `openai/gpt-image-2/edit` with the real product photo
as image 1 (`--product-image pack-front.jpg`, put first), its front facing the camera. The script
adds that line to the prompt. Never type the label's words into the prompt.
- **Check the label at 2x against the photo.** If it garbles, reframe so the hands cover the label
area while the shape and colours still match, or cut to the real photo during the action. A blank,
unprinted, back-facing or unlabeled pack of the advertised product is a P0 defect.
- **Each visible state** (closed pack, open pack, contents, cooked food) comes from a real brand
photo sent as the edit reference, never from prose. No photo of a state: don't show it.
```bash
gen_image.py --model openai/gpt-image-2/edit --brand-product held --product-image pack-front.jpg \
--image presenter.png --payload '{"prompt": "She tips the open pouch into the pot of boiling water"}' \
--out pour.png
```
## Contract
- Paid calls route through the GooseWorks proxies (charged to your GooseWorks credits) via the
bundled `media_proxy.py` — never a provider SDK's default host.
- The template recipe (DB) supplies the model + params; this capability is generic.
## Edit an image (Nano Banana edit)
`fal-ai/nano-banana/edit` is the default editor for a prompt-driven edit: a restage or scene swap,
a subject swap in the same pose, a product placed into a scene. It keeps subject geometry, the room
and the light. Use create-image-gpt-image-fal instead when lettering is the point of the edit
(logos, wordmarks, signage) or for a transparent cutout. Small text inside the source can come back
mangled with either model; check it.
```bash
gen_image.py --model fal-ai/nano-banana/edit --image product.png --image https://.../scene.jpg \
--payload '{"prompt": "Put the bottle on the counter; keep everything else unchanged", "aspect_ratio": "9:16"}' \
--out edit.png
```
- `--image` is repeatable and keeps its order; a public url passes through, a local file is put on
the fal CDN through the GooseWorks storage proxy (free). Urls already in the payload's
`image_urls` come first.
- Payload: `prompt` (required); `aspect_ratio` `auto` (the input's), `21:9`, `16:9`, `3:2`, `4:3`,
`5:4`, `1:1`, `4:5`, `3:4`, `2:3` or `9:16`; `output_format` `png`, `jpeg` or `webp`; `seed`. The
script saves the first image, so leave `num_images` at 1.
- There is no mask. For a region-locked edit, say "keep everything else unchanged" in the prompt.
- Keep the long edge of each input at about 2048 px or less.
- Check the result: it decodes, its size matches the asked aspect, and no watermark, garbled text
or extra fingers appear.
## Creator references
Whenever the image will contain a person (creator, presenter, interviewee or any face in
focus), with or without a template recipe, read the bundled
[avatar-generation guide](references/avatar-generation.md) before any paid call. Unless the
selected recipe prescribes its own creator still, make the person with
create-creator-takes-h3's scripts/make_character.py, which calls this skill. Never hand-write
a person prompt here and never use a Flux route (fal-ai/flux/dev, fal-ai/flux/schnell) for a face.
When the selected video route forbids person references (for example the street-interview
format), make no person still at all.
If a required guide cannot be fetched or opened, stop and name the missing file.
Use the current project's creator choices and approved model binding; fetching
this guide does not authorize another image, a retry or a different render engine.
## Model notes
How the image models behave, measured on shipped projects. Each note lives here once; recipes
point here instead of repeating it.
- **Nano Banana: a reference anchors identity and style, not layout.** It keeps a product's
colourway, materials and proportions, and takes framing, background and pose from the prompt.
For a photo-faithful product, composite the real photo instead.
- **Nano Banana does not lock a face.** With one reference or four real frames, it matched the room
and the clothes and returned a lookalike. Never use it to re-pose or re-light a person already in
published work, and compare faces side by side at the same scale before anything downstream spends
on the result.
- **Nano Banana redraws wordmarks** with wrong glyphs. For product lettering use gpt-image-2 edit
(create-image-gpt-image-fal) or composite the real packshot.
- **Nano Banana paths** are `fal-ai/nano-banana` and `fal-ai/nano-banana/edit`. The engine label
`nano_banana_2` is not a fal path, and the proxy refuses it as one.
- **Seedream 4.5 fills the canvas with a same-shaped subject:** a wide monitor in a 9:16 canvas came
back turned portrait three times in four. Generate at the subject's own aspect and crop.
- **Seedream 4.5 draws the device you name.** "Shot on a phone camera" put a phone and a hand in the
frame despite the negatives. Describe the artefacts (close range, wide lens, mild barrel
distortion, sensor noise), never the device.
- **No image model holds a caption safe zone** from framing words. Generate loose, then crop or pad
to the zone and check it by drawing the band on the frame.
- **Hard constraints:** see the note in create-video-fal; it applies to image prompts too.
Related skills
| Skill | Stars | Updated | Works with | Status | Install |
|---|---|---|---|---|---|
| reverse-engineer-anythingmorluto/reaReverse engineer native, managed, Electron/JavaScript, packaged, firmware, and browser targets with REA, and build agent environments (AgentEnv bundles) of web applications from a live instance or a capture, and of Linux desktop apps with a task from a request. Use shipped-artifact or requested runtime evidence to explain features, compare versions, decompile code, or guide a reconstruction. Skip REA for ordinary source-repository architecture analysis. | Stars: 472k | Updated | Claude CodeCodex | Verified | Install |
| diagnosing-superpowersobra/superpowersUse when a superpowers session went wrong and your human partner wants to know why — repeated work, ignored plans, stumbles, poor results, a skill that didn't fire, "it took too long", "why is it so expensive", "what is it doing" — or wants to build a bug report for the superpowers maintainers, for the current session or a past one identified by id or path, on any harness. | Stars: 323k | Updated | Claude CodeCodex | Verified | Install |
| dispatching-parallel-agentsobra/superpowersUse when facing 2+ independent tasks that can be worked on without shared state or sequential dependencies | Stars: 323k | Updated | Claude CodeCodex | Verified | Install |
| executing-plansobra/superpowersUse when executing an implementation plan in the current session as the implementer yourself — your human partner chose inline execution, or no subagent tool is available | Stars: 323k | Updated | Claude CodeCodex | Verified | Install |