---
base_model: stabilityai/sd-turbo
pipeline_tag: image-to-image
tags:
  - joint-source-channel-coding
  - wireless-image-transmission
  - semantic-communication
  - one-step-diffusion
---

# OSD-JSCC weights

Pretrained weights for **OSD-JSCC: Neural Joint Source–Channel Coding with
One-Step Diffusion**. Code, usage instructions and a Colab
demo are at **https://github.com/tangshunpu/OSD-JSCC**; the project page is
**https://tangshunpu.github.io/OSD_JSCC_Project/**.

| File | Model | Size |
| --- | --- | ---: |
| `osd_jscc_cr96_awgn.pt` | OSD-JSCC, AWGN channel, CR = 96, trained on 0–20 dB SNR | 290 MB |
| `swinjscc_witt_c16_awgn.pth` | SwinJSCC baseline used in the paper's comparisons (WITT architecture, C = 16, CR = 96), AWGN, trained on DIV2K at 0–20 dB | 126 MB |

`osd_jscc_cr96_awgn.pt` holds only the weights trained for OSD-JSCC: the JSCC
encoder and decoder, the latent adapter, the SNR embedding, the U-Net LoRA,
`conv_in`/`conv_out` and the fine-tuned VAE decoder convolutions. It is loaded
on top of [SD-Turbo](https://huggingface.co/stabilityai/sd-turbo) (U-Net, VAE
and scheduler), which is downloaded separately and is subject to the Stability
AI Community License.

`swinjscc_witt_c16_awgn.pth` contains the encoder and decoder of the WITT
network in [TF_SemCom](https://github.com/tangshunpu/TF_SemCom) at commit
`75a581d`. It was initialized from the official WITT C = 16 AWGN release and
fine-tuned on DIV2K with SNR drawn from 0–20 dB.

```python
from huggingface_hub import hf_hub_download

osd_ckpt = hf_hub_download("tangshunpu/OSD-JSCC", "osd_jscc_cr96_awgn.pt")
swin_ckpt = hf_hub_download("tangshunpu/OSD-JSCC", "swinjscc_witt_c16_awgn.pth")
```