AuroraIMG-6M API

A small, authenticated text-to-image API for direct-img. It runs AuroraIMG-6M, a 6.8M-parameter model, on CPU with onnxruntime. An image takes about 0.4s and the container uses about 150-200 MB of RAM.

The output is 64×64 and upscaled, so images are soft and painterly: right colors, setting and mood, often not the exact object. That suits an abstract fallback when no real image exists.

API

GET /generate?prompt=frog+wizard+casting+a+spell
Authorization: Bearer $AURORA_API_KEY

Param	Default	Range
prompt	required	1-200 characters
size	512	64-1024 (output is square, upscaled from 64×64)
format	webp	webp, png, jpeg
steps	12	1-50 (12 looks the same as 50, at a quarter of the time)
cfg	6	0-20 (guidance; ≤1 disables it)
seed	CRC32 of the lowercased prompt	0-4294967295

The same prompt and params always give the same image. Responses include X-Seed and X-Generation-Ms.

Errors are JSON {"error": "..."}: 400 bad params, 401 missing or wrong key, 503 when the queue is full (Retry-After: 2).

GET /health returns ok without auth.

Config

Env var	Default
AURORA_API_KEY	none	Required. Without it /generate rejects every request
PORT	8080
THREADS	1	onnxruntime threads per generation
CONCURRENCY	2	Generations at once; more requests queue
MAX_QUEUE	32	Queued requests beyond this get 503

Deploy

Pushing to main builds the image with GitHub Actions, pushes it to ghcr.io/direct-img/auroraimg-6m:main, then calls the Dokploy deploy webhook (repo secret DOKPLOY_WEBHOOK_URL). It runs as the AuroraIMG-6M app in the direct-img Dokploy project.

Run locally:

pip install -r requirements.txt
AURORA_API_KEY=dev python app.py
curl -H "Authorization: Bearer dev" "localhost:8080/generate?prompt=orange+cat" -o cat.webp

Model and license

model/ holds an ONNX conversion of AuroraAI-Research/AuroraIMG-6M at revision 3d66743, licensed under the Apache License 2.0 (copy in model/LICENSE). The weights are unchanged; they were converted from the original safetensors to ONNX, and the latent un-normalization and output scaling were folded into decoder.onnx. Output matches the original pipeline to within 1/255 per pixel.

scripts/export_onnx.py regenerates model/ from the original repo (needs torch, onnx, onnxscript, onnxruntime, safetensors).