ComfyUI OpenAI/FAL-Compatible Images API

A single-container, multi-protocol image-generation backend backed by ComfyUI. The container starts ComfyUI on the first generation request, keeps it alive while you use it, and cold-stops it after an idle timeout so GPU memory is freed when nothing is generating.

Supported protocols:

• OpenAI-compatible — POST /v1/images/generations, GET /v1/models

• FAL.ai-compatible — POST /fal/run/{model_id}, POST /fal/queue/{model_id}, GET /fal/queue/{model_id}/requests/{request_id}/status

• Wrapper native — POST /generate, GET /health, GET /status, POST /stop

Generated images are persisted under ComfyUI/output and served at /outputs/{filename}, so response_format: url and FAL-style responses return real URLs.

Endpoints

OpenAI-compatible

POST /v1/images/generations accepts the standard OpenAI request shape:

{
  "prompt": "a photo of an astronaut riding a horse on the moon",
  "n": 1,
  "size": "512x512",
  "response_format": "b64_json"
}

Supported parameters:

• prompt (required)

• n (1–4, default 1)

• size — "WIDTHxHEIGHT", e.g. "512x512" (default), "1024x1024"

• response_format — "b64_json" (default) or "url"

• model — accepted but currently informational

• quality, style, user — accepted but ignored

• ComfyUI overrides: negative_prompt, width, height, steps, cfg, seed

FAL.ai-compatible

POST /fal/run/{model_id} accepts a FAL-style payload:

{
  "prompt": "a serene mountain landscape with cherry blossoms",
  "image_size": "landscape_16_9",
  "num_images": 1,
  "seed": 42
}

Supported fields:

• prompt (required)

• negative_prompt

• image_size — e.g. square_hd, landscape_16_9, portrait_16_9

• aspect_ratio — e.g. 1:1, 16:9, 9:16

• width / height (override size/aspect)

• num_inference_steps → mapped to ComfyUI steps

• guidance_scale → mapped to ComfyUI cfg

• seed

• num_images (1–4)

• image_url / reference_image_urls — rejected with a clear error (editing is not supported)

Response shape mirrors FAL:

{
  "images": [
    {
      "url": "http://192.168.2.51:8002/outputs/abc123.png",
      "width": 1536,
      "height": 1024,
      "content_type": "image/png"
    }
  ],
  "prompt": "a serene mountain landscape with cherry blossoms",
  "seed": 42,
  "has_nsfw_concepts": [false]
}

Related MCP server: Universe Imagine MCP

Running the container

cd wrapper
# place v1-5-pruned-emaonly.safetensors in this directory first
docker compose up -d --build

The API is exposed on port 8002 (ComfyUI direct is on 8190).

Environment variables

Using with Hermes

Hermes supports FAL.ai as a backend. Point Hermes at this container by making fal.run / queue.fal.run resolve to http://192.168.2.51:8002. The easiest way is to add a local DNS or proxy rule, or set the FAL client base URL if Hermes/FAL's SDK exposes one.

Set FAL_KEY to any non-empty dummy value in Hermes config, pick any FAL model id (e.g. fal-ai/flux-2/klein/9b), and generation requests will be handled by your local ComfyUI backend.

Example requests

OpenAI:

curl -X POST http://127.0.0.1:8002/v1/images/generations \
  -H "Content-Type: application/json" \
  -d '{
    "prompt": "a photo of an astronaut riding a horse on the moon",
    "size": "512x512",
    "response_format": "b64_json"
  }'

FAL:

curl -X POST http://127.0.0.1:8002/fal/run/fal-ai/flux-2/klein/9b \
  -H "Content-Type: application/json" \
  -d '{
    "prompt": "a serene mountain landscape with cherry blossoms",
    "image_size": "landscape_16_9"
  }'

Cold-start behavior

The first request after the wrapper has been idle will trigger a ComfyUI startup. The API server waits for ComfyUI to start and generate the image, so the client may observe a longer response time on the first call. Subsequent calls are fast until the idle timeout elapses and ComfyUI is stopped again.