aeic-se-onnx

AEIC-SE — 512×512 image codec for LoRa mesh radio

An ONNX build of AEIC-SE (Zhang et al., CVPR 2026), an asymmetric extreme image codec. It compresses a 512×512 photograph to roughly 156 bytes and reconstructs it with a one-step diffusion decoder.

At that size an image fits in one or two MeshCore LoRa packets, which is what makes photo messaging possible on a mesh radio link that carries a few hundred bytes at a time.

What it does

Examples

A 512×512 photograph reconstructed from 135 bytes:

The same packet budget spent on a conventional codec. WebP cannot encode 512×512 at this size at all — the largest that fits is a 24×24 thumbnail:

And the failure mode. A small, distant subject at 214 bytes — the scene, the path and the light are right; the animal is not. The fox has become a sheep, and nothing about the output looks damaged:

Intended use

Sending photographs over bandwidth-constrained links — LoRa mesh, satellite messengers, HF packet, anything where a kilobyte is expensive. Built for MeshCore Open, where an image travels as one or two GRP_DATA packets on a channel.

Encoding is cheap enough for any phone. Decoding runs the diffusion model and wants a device with real memory headroom.

What's in the bundle

aeic_decoder_qdq_conv_pct.onnx(.data)   832 MB   image decoder, int8
aeic_entropy_side_fp32_op17.onnx         64 MB   entropy model, encode side
aeic_entropy_decode_fp32_op17.onnx       58 MB   entropy model, decode side
aeic_cdf_ft32.bin                       0.8 MB   rANS CDF tables

Both entropy graphs stay fp32 — the entropy path has to reproduce integers exactly on both ends or the range coder desynchronises. The image decoder is quantized to int8, which costs 0.25 dB against the original photograph.

Limitations

The decoder is generative. At 156 bytes it is not recovering the photograph; it is synthesising an image consistent with a 156-byte description. Output is sharp and natural-looking and can still be wrong — in testing a fox came back as a convincing sheep. Small or distant subjects are the usual casualty; a large, centred subject survives well.

Treat received images as impressionistic. They are not evidence, and nothing in the output signals which details were reconstructed rather than transmitted.

Fixed at 512×512. Smaller inputs collapse: the diffusion decoder needs a 64×64 latent, and 512×512 is exactly that.

Credits

AEIC — Tianyu Zhang, Dong Liu, Chang Wen Chen (CVPR 2026). Decoder built on SD-Turbo; pruned VAE decoder from AdcSR.

@InProceedings{Zhang_2026_CVPR,
  author    = {Zhang, Tianyu and Liu, Dong and Chen, Chang Wen},
  title     = {Ultra-Low Bitrate Perceptual Image Compression with Shallow Encoder},
  booktitle = {CVPR},
  year      = {2026},
  pages     = {12118-12128}
}