Mon 11 August 2025

You don’t need a paid API or a farm of GPUs to make a demo look real. Let’s turn placeholders into useful pictures with a small local server and a couple of dead‑simple calls. See it in action here: Bob’s Widgets demo.

What we’re using

ImageAIServer — a local, privacy‑first image/vision server with an OpenAI‑compatible images endpoint.
Your app — React, Vue, Django, whatever. If it can call HTTP, it can make pictures.

Why local? - No rate limits or surprise bills. - Works offline; sensitive data stays on your machine. - CPU mode for any box, GPU mode for quality/speed when you have it.

Quick start (pick one)

Option A — CPU/Any Machine (ONNX)

:::bash pip install imageaiserver imageaiserver

open http://localhost:8001

:::

Option B — GPU/NVIDIA or Apple Silicon (PyTorch)

:::bash pip install imageaiserver[torch] imageaiserver

SDXL/FLUX models will be available

:::

Option C — Docker

:::bash docker run -p 8001:8001 imageaiserver:latest :::

Once it’s running, hit http://localhost:8001 in your browser to test a prompt. The server also exposes an images API at /v1/images/generations.

Minimal integration example

cURL

:::bash curl -X POST http://localhost:8001/v1/images/generations \ -H "Content-Type: application/json" \ -d '{ "prompt": "a serene mountain landscape", "model": "sd15-onnx" }' :::

Python

:::python import base64, json, requests

resp = requests.post( "http://localhost:8001/v1/images/generations", json={"prompt": "cute robot in a garden", "model": "sd15-onnx"} ) img_b64 = resp.json()["data"][0]["b64_json"] with open("robot.png", "wb") as f: f.write(base64.b64decode(img_b64)) :::

Frontend (TypeScript)

:::ts async function generateImage(prompt: string) { const res = await fetch("http://localhost:8001/v1/images/generations", { method: "POST", headers: { "Content-Type": "application/json" }, body: JSON.stringify({ prompt, model: "sd15-onnx" }) }); const { data } = await res.json(); return data:image/png;base64,${data[0].b64_json}; } :::

From placeholders to a real catalog

Here’s a tiny batch script that generates thumbnails for a mock product list. Swap the prompts for your domain and you’re done.

:::python import base64, json, os, requests

products = [ {"id": 1, "name": "Precision widget", "prompt": "macro photo of a brushed aluminum precision widget on a neutral background"}, {"id": 2, "name": "Outdoor widget", "prompt": "rugged outdoor widget on a weathered wood table, soft daylight"}, # ...add 50–500 as needed ]

os.makedirs("public/images", exist_ok=True)

for p in products: r = requests.post( "http://localhost:8001/v1/images/generations", json={"prompt": p["prompt"], "model": "sd15-onnx"} ) img_path = f"public/images/widget_{p['id']}.png" with open(img_path, "wb") as f: f.write(base64.b64decode(r.json()["data"][0]["b64_json"])) print("wrote", img_path) :::

Wire those paths into your seed data and your UI suddenly looks like a real store. For a live example of that idea, open the Bob’s Widgets demo.

Picking a model

If you’re on CPU, start with sd15-onnx (INT8, ~500 MB RAM). On a GPU, use sdxl for quality, sdxl-turbo for speed, or flux1-schnell if you have the VRAM.

Model	Memory (approx.)	Hardware	Quality	Speed
SD 1.5 ONNX INT8	~500 MB	CPU	Good	Fast
SDXL	~8 GB	GPU	Excellent	Medium
SDXL Turbo	~8 GB	GPU	Very good	Very fast
FLUX.1 Schnell	~12 GB	GPU	Excellent	Fast

Tip: you can switch models per‑request via the model field. Keep your code the same and adapt quality/speed on the fly.

Production notes

Determinism: if you need repeatable results, add a seed field.
File naming: hash prompts or store IDs to avoid overwriting.
Caching: serve generated images via your CDN once baked.
Access: keep the generation endpoint private; ship baked assets publicly.

What you get for doing this

Realistic demos that sell your UI/UX.
Faster product iteration (no waiting on stock images or approvals).
Zero marginal cost at prototype scale.

If you want to go deeper, the repository has install options (CPU/ONNX, GPU/Torch), a simple CLI (imageai-generate), and Docker examples. It’s all local‑first and friendly to script.

Repo: https://github.com/imgailab/imgai-server

PS: If you like seeing this idea outside of a blog post, that’s exactly what the Bob’s demo is meant to show—what “good enough to evaluate” looks like without spending money or leaking data.