/create-image-gpt-image-fal

name: create-image-gpt-image-fal
description: Generate a single photoreal or designed image with OpenAI gpt-image via fal.ai. Supports gpt-image-1 (default, fixed sizes — the FAL fallback for Higgsfield's `gpt_image_2`) and gpt-image-2 (`openai/gpt-image-2`, custom output sizes up to 3840px). Routes to text-to-image or the edit variant depending on whether a reference image is provided. Use for photoreal character anchors, scene keyframes, and designed sheets (e.g. storyboards) where precise layout and legible text matter.

create-image-gpt-image-fal

Purpose

Generate one image via fal.ai's OpenAI gpt-image endpoints. Two model families are supported through a single `--model` flag:

• **`gpt-image-1`** (default) — `fal-ai/gpt-image-1`. The FAL fallback for Higgsfield's `gpt_image_2`. Fixed output sizes only. Used by:
• `video-orchestrator/lock-character` Phase 0 (anchor portrait) and Phase 1 (angle keyframes via `/edit`)
• `video-orchestrator/create-clips` Phase 1 for photoreal scenes
• the orchestrator's `generate_with_fallback.py` router on Higgsfield failure
• **`gpt-image-2`** — `openai/gpt-image-2`. The newer model; accepts **custom output sizes** (any multiple of 16, up to 3840px) and renders dense text/layouts well. Used for designed sheets such as ad storyboards.

The default stays `gpt-image-1` so existing callers and the lock-character anchor-parity contract are unaffected. Opt into the newer model with `--model gpt-image-2`.

Inputs

Required:

• `--prompt` — text prompt. A verbatim character descriptor block goes here for character work.
• `--output` — local PNG destination.

Optional:

• `--model` — `gpt-image-1` (default) or `gpt-image-2`.
• `--aspect-ratio` — `9:16` (default), `16:9`, `1:1`, `2:3`, `3:2`. gpt-image-2 also accepts `3:4`, `4:3`, `4:5`. Used when `--image-size` is not given.
• `--image-size` — explicit `WIDTHxHEIGHT` (e.g. `1728x2304`). **gpt-image-2 only** — values are rounded to multiples of 16 and capped at 3840px. On `gpt-image-1` a custom size is ignored with a warning and the aspect-ratio mapping is used instead.
• `--quality` — `low | medium | high` (default `medium`; use `high` for finals).
• `--ref-image` / `--ref-url` — a **PUBLIC image URL** for the `/edit` variant. **Repeatable** — pass it twice to send multiple refs (e.g. identity + style). The proxy does **not** upload local files, so a **local path is rejected** — host the image first (the MCP `media_upload`, or any public URL) and pass that URL. When present, routes to the model's `/edit` variant so the model can match the references. Order matters: pass identity (character) first, then style refs.
• `--with-logs` — stream fal queue logs.

Credentials (proxy-routed — NOT a raw FAL key):

• The bundled `scripts/media_proxy.py` routes every call through the GooseWorks **fal-proxy**, so every call **is charged to your GooseWorks credits**. It reads `~/.gooseworks/credentials.json` (`api_base`, `api_key`, `agent_id`), written by the GooseWorks CLI. Do **not** set `FAL_API_KEY`: an agent (`cal_`) token is not a FAL key and 401s against fal directly.
• Set `GW_PROJECT_ID=<ad project id>` in the env so the generation's spend attributes to that ad project (per-project cost shows in the app).

Credentials at run time

A cloud sandbox injects `GW_MEDIA_PROXY_TOKEN`; a terminal uses the CLI's credentials file. With neither, the bundled helper relays each paid call through the agent instead (see media-proxy), so the script never asks for a key or a sign-in.

Workflow

# Text-to-image, default model (gpt-image-1)
python3 skills/video/create-image-gpt-image-fal/scripts/generate.py \
  --prompt "..." \
  --output /path/to/anchor.png \
  --aspect-ratio 9:16 \
  --quality medium

# Edit-from-reference (anchor -> angle). --ref-image must be a PUBLIC URL,
# NOT a local path (the proxy does not upload local files):
python3 .../generate.py \
  --prompt "..." \
  --output /path/to/angle-3q-left.png \
  --ref-image "https://.../anchor.png" \
  --aspect-ratio 9:16

# gpt-image-2 with a custom output size (e.g. a designed storyboard sheet)
python3 .../generate.py \
  --prompt "..." \
  --output /path/to/storyboard.png \
  --model gpt-image-2 \
  --image-size 1728x2304 \
  --quality high

The script: 1. Loads the agent credentials from `~/.gooseworks/credentials.json` via the bundled `media_proxy.py` (proxy-routed; charged to your GooseWorks credits). 2. Resolves the model family (`--model`) and output size (`--image-size` if given and supported, else the aspect-ratio mapping). 3. If one or more `--ref-image` / `--ref-url` flags are set, passes them as `image_urls=[url1, url2, ...]` (they must already be PUBLIC URLs) and routes to the model's `/edit` variant. Otherwise routes to the `/text-to-image` variant. 4. Submits through the GooseWorks **fal-proxy** and polls the queue to completion — host-swapping the `queue.fal.run` status/response URLs to the proxy base (see `media_proxy.py`); never polls `queue.fal.run` directly. 5. Downloads the first result image to `--output`. 6. Writes `<output>.meta.json` with `gateway: "fal-proxy"`, model id, `model_family`, request, and cost.

Output

• `<output_path>` — PNG (≥ 1 KB).
• `<output_path>.meta.json` — request + result metadata + cost, including `model_family` (`gpt-image-1` or `gpt-image-2`).

Quality Checks

• Output file exists and is > 1 KB.
• For character anchors: visually inspect against the descriptor block (hair, shirt color, age).
• `meta.json` includes `gateway: "fal-proxy"`, the resolved `model` id, `model_family`, `image_size`, and `quality`.
• For gpt-image-2 custom sizes: confirm the output dimensions match the requested `WIDTHxHEIGHT`.
• **No readable text in the prompt that should appear in the image.** AI image models mangle short brand text, URLs, code tokens, captions, and wordmarks even with explicit prompting. Examples observed: `"ffmpeg"` → `"ffmmg"`; `"claritude"` → `"clarytude"`; `"therapists"` → `"therap