| name | gpt-image2 |
| description | Generate and edit images with Dreamfield's OpenAI-compatible gpt-image-2 image API. Use when the user asks Codex to create, test, or save an image from a text description, make wallpaper or illustration assets, perform image-to-image generation from reference images or clipboard images, use masks/inpainting, verify Dreamfield image generation, or troubleshoot image-generation responses such as empty data, b64_json, URL output, quota errors, or content policy errors. |
GPT Image 2
Use Dreamfield's OpenAI-compatible image endpoints:
POST https://www.dreamfield.top/v1/images/generations
POST https://www.dreamfield.top/v1/images/edits
Recommended environment:
- Provider: Dreamfield, https://www.dreamfield.top/
- Model:
gpt-image-2
- API key env var:
NEW_image2_API_KEY
- Author usage note: this skill has been tested with Dreamfield for personal use. It supports 4K output. For details, check Dreamfield for current pricing and model limits.
Use scripts/generate_image2.py for all calls. It handles text-to-image, image-to-image, clipboard images, input preprocessing, mask normalization/generation, manual mask drawing, mask overlay previews, conservative retry, response saving, multi-image output, target resizing, and summary.json indexing.
Workflow
- Convert the user's request to a concise English prompt. Include practical constraints such as
no text and no watermark.
- Use text-to-image when there is no usable local or clipboard image.
- Use image-to-image when the user provides
--image, asks to use a reference image, or asks for image-to-image.
- Use
--clipboard-image when the user copied an image to the OS clipboard.
- Use mask editing only for localized edits/inpainting; otherwise omit mask parameters. Use
--mask-draw when the user needs to manually paint or circle the edit area.
- Prefer
response_format=b64_json for deterministic local PNG saving.
- Show final local image paths and summary paths.
Common Commands
Text-to-image:
SKILL_DIR="${CODEX_HOME:-$HOME/.codex}/skills/gpt-image2"
python3 "$SKILL_DIR/scripts/generate_image2.py" \
--prompt "A cinematic 16:9 desktop wallpaper of golden lightning over green hills at sunset, no text, no watermark." \
--size 1792x1024
Image-to-image:
SKILL_DIR="${CODEX_HOME:-$HOME/.codex}/skills/gpt-image2"
python3 "$SKILL_DIR/scripts/generate_image2.py" \
--image "/absolute/path/to/reference.png" \
--prompt "Create a new photorealistic 16:9 wallpaper inspired by this reference image, no text, no watermark." \
--size 1792x1024
Masked edit:
SKILL_DIR="${CODEX_HOME:-$HOME/.codex}/skills/gpt-image2"
python3 "$SKILL_DIR/scripts/generate_image2.py" \
--image "/absolute/path/to/source.png" \
--mask-auto rect \
--mask-rect 120,80,640,360 \
--prompt "Replace only the masked area while preserving the rest of the image. No text, no watermark." \
--size 1792x1024
Manual masked edit:
SKILL_DIR="${CODEX_HOME:-$HOME/.codex}/skills/gpt-image2"
python3 "$SKILL_DIR/scripts/generate_image2.py" \
--image "/absolute/path/to/source.png" \
--mask-draw \
--prompt "Replace only the painted area while preserving the rest of the image. No text, no watermark." \
--size 1792x1024
Self-test:
SKILL_DIR="${CODEX_HOME:-$HOME/.codex}/skills/gpt-image2"
python3 "$SKILL_DIR/scripts/generate_image2.py" --self-test
References
- Full CLI parameters:
references/cli.md
- Mask semantics and examples:
references/mask.md
- Self-test, retry, and troubleshooting:
references/troubleshooting.md
Load the relevant reference file only when the user needs those details.