supra-openai

Run it

One container. Weights download on first start and are cached in the volume.

Start the server

docker run -p 8000:8000 \
  -v supra-hf:/data/hf \
  ghcr.io/sirmmo/openai-supra:latest

Generate with the OpenAI SDK

from openai import OpenAI

client = OpenAI(
    base_url="http://localhost:8000/v1",
    api_key="not-needed",
)

result = client.images.generate(
    model="supra2-img",
    prompt="a red fox in fresh snow",
    extra_body={"seed": 42, "steps": 50},
)

Samples

Generated by this server on CPU, at the model's native 256 × 256.

A jellyfish glowing in dark water — “a sea jellyfish floating in the pitch-black ocean depths” · seed 0, 50 steps

A red fox sitting in snow at golden hour — “a red fox sitting in fresh snow, golden hour” · seed 42, 20 steps

A lit cabin among pine trees at night — “a cozy cabin in a pine forest at night, warm window light” · seed 3, 30 steps

A hot air balloon above hills at sunrise — “a hot air balloon over green hills at sunrise” · seed 0, 50 steps

A wooden sailboat on turquoise water — “a wooden sailboat on a calm turquoise sea” · seed 0, 50 steps

Built-in web form

The server's root serves a form for prompting without any client — same API underneath.

Screenshot of the built-in generation form showing two lighthouse images

API

POST /v1/images/generations, with OpenAI's request and error shapes.

Field	Supported values
`prompt`	Required. Flan-T5 sees the first 128 tokens.
`n`	1–10, generated as one batch.
`size`	`auto` / `256x256` native; other sizes are upscaled and center-cropped.
`quality`	`low` = 20 steps, `medium` = 35, otherwise 50.
`response_format`	`b64_json` (default) or `url`.
`output_format`	`png`, `jpeg`, `webp`.
`seed`, `guidance_scale`, `steps`	Extensions — send via the SDK's `extra_body`. A seed reproduces the same image on any device.

Also GET / (web form), /v1/models, /docs (OpenAPI) and /health. Set SUPRA_API_KEY to require Authorization: Bearer.

Good to know

256 × 256 is the real resolution. Larger sizes are resampled, not generated at higher detail.
Speed. 50 steps take roughly 10–15 s on CPU; a GPU build is a one-line change to the Docker build arg.
One request at a time. The sampler is serialized, so the server is best used as a personal or small-team endpoint.
Not supported. Streaming, image edits and variations — the model is text-to-image only.