This project is an adapter for direct OpenAI-compatible APIs, self-hosted gateways, and third-party providers with small field differences. It keeps endpoint paths, multipart image field names, extra headers, and extra body fields editable from the node UI.

Features

Text-to-image through an OpenAI-compatible JSON endpoint.
Image edit / image-to-image through an OpenAI-compatible multipart endpoint.
Configurable api_base, endpoint_path, model string, headers, and body fields.
Reference images via fixed slots or a chainable reference list.
Per-reference prompts appended into the main prompt.
Optional provider-specific per-reference prompt multipart field.
Reads OPENAI_API_KEY by default, with optional node-level key override.
Supports image responses as data[].b64_json or data[].url.

Nodes

Node	Purpose
`OpenAI Compatible Image Generate`	Sends a text-to-image request to `/images/generations` by default.
`OpenAI Compatible Image Edit With References`	Sends reference images to `/images/edits` by default.
`OpenAI Compatible Reference Image`	Adds one image and one reference prompt to a chainable reference list.

Legacy node IDs from the initial GPT Image 2 prototype are still registered so older workflows can load.

Installation

Clone or copy this repository into your ComfyUI custom nodes directory:

cd ComfyUI/custom_nodes
git clone https://github.com/wennatre/ComfyUI-OpenAI-Compatible-Image.git

Install dependencies in the same Python environment used by ComfyUI:

pip install -r ComfyUI-OpenAI-Compatible-Image/requirements.txt

Set your API key before starting ComfyUI:

export OPENAI_API_KEY="your_api_key"

On Windows PowerShell:

$env:OPENAI_API_KEY="your_api_key"

Restart ComfyUI after installation.

Text To Image

Default request:

POST {api_base}/images/generations
Content-Type: application/json
Authorization: Bearer {api_key}

Default body:

{
  "model": "gpt-image-2",
  "prompt": "...",
  "n": 1,
  "size": "1024x1024"
}

Use extra_headers_json for provider headers:

{"x-provider-key": "value"}

Use extra_body_json for provider body fields:

{"user": "comfyui"}

timeout_seconds defaults to 600. Reference-image edits can be slow on proxy providers; if a request times out, increase this value before retrying.

Reference Images

The edit node defaults to:

POST {api_base}/images/edits
Content-Type: multipart/form-data
Authorization: Bearer {api_key}

Reference image input modes:

Fixed slots: connect image_1 ... image_8.
List mode: chain OpenAI Compatible Reference Image nodes and connect the final references output.

reference_image_count controls how many references are sent. Set it to 0 to send all connected fixed-slot images or all list items.

Each reference can have a prompt. By default, reference prompts are appended to the main prompt:

Reference image 1: use this as the subject identity.
Reference image 2: use this as the lighting and style.

Official OpenAI image edit requests do not define a separate per-image prompt field. If your custom endpoint supports one, set reference_prompt_field, for example:

reference_prompt[]

The default multipart image field is:

image[]

If your endpoint expects repeated image fields instead, set:

image

Workflows

Example workflows are included:

workflows/openai_compatible_text_to_image_workflow.json
workflows/openai_compatible_reference_images_workflow.json

Import them from ComfyUI after installing the node.

Repository Layout

This repository follows a small, repeatable layout for future ComfyUI node projects:

ComfyUI-OpenAI-Compatible-Image/
├── __init__.py
├── README.md
├── README.zh-CN.md
├── requirements.txt
├── pyproject.toml
├── LICENSE
├── workflows/
│   ├── openai_compatible_text_to_image_workflow.json
│   └── openai_compatible_reference_images_workflow.json
└── docs/
    └── repo-layout.zh-CN.md

For future work, keep each installable ComfyUI custom node as its own repository. Keep cross-node, ready-to-use workflows in a separate workflow collection repository.

Compatibility Notes

The model field is intentionally free text. gpt-image-2 is the default because many custom endpoints use that string, but you can set any model your endpoint accepts.

The node expects an OpenAI-style response:

{
  "data": [
    {"b64_json": "..."}
  ]
}

It also accepts URL responses:

{
  "data": [
    {"url": "https://..."}
  ]
}

Security

Prefer OPENAI_API_KEY over pasting keys into workflows. Workflows can be shared accidentally, and node widget values may be saved in workflow JSON.

Troubleshooting

The second generation times out

Some OpenAI-compatible proxies keep HTTP connections open incorrectly. The nodes send Connection: close by default to avoid reusing a stale connection.

If it still times out:

Increase timeout_seconds to 900 or 1200.
Reduce n to 1.
Reduce the number or resolution of reference images.
Check whether your provider uses an async job API instead of returning data[].b64_json or data[].url directly.
If your provider has a custom polling endpoint, this node needs provider-specific async polling support.

Development

Basic local checks:

python -m py_compile __init__.py
python -m json.tool workflows/openai_compatible_text_to_image_workflow.json
python -m json.tool workflows/openai_compatible_reference_images_workflow.json

License

MIT