OSD-JSCC-P

Progressive wireless image transmission with one-step diffusion. One model covers six rates. The transmitter encodes the image once into six nested symbol layers (2, 2, 4, 8, 8, 8 latent channels) and sends them over a simulated AWGN channel. A receiver that stops after the first k layers gets exactly a transmission at CR 768, 384, 192, 96, 64, 48 (CR 96 = one complex channel use per 96 source values). Each received prefix is decoded by the JSCC decoder, one SD-Turbo U-Net step and the VAE decoder.

0 20
Examples (click to transmit)
Layers received

Quality of each prefix

Model. Trained on the AWGN channel at 0–20 dB SNR. Both the encoder and the decoder know the SNR; only the receiver is told how many layers arrived.

Examples are CLIC2020 test images (not used in training) and the author's own photos.

Metrics are computed on the 8-bit outputs: PSNR, LPIPS (AlexNet) and DISTS from pyiqa; lower LPIPS / DISTS is better. Images wider or taller than 1536 px are downscaled first, and sides are padded to a multiple of 128 for the codec. The seed fixes the channel noise.

Conference version (single rate, CR 96): code · project page. SD-Turbo is used under the Stability AI Community License.