A visual comparison illustrating the impact of several unexpected 26 | transmission degradations: ① unseen channel fading, ② PAPR reduction, ③ 27 | with ISI (removed CP symbols), and ④ very low CSNR (0dB).
28 |
29 |
30 | ## RDP curves on [FFHQ](https://github.com/NVlabs/ffhq-dataset) testset
31 |
32 | 
33 |
34 | ## Generalization to unseen wireless conditions
35 |
36 | 
37 |
38 | ## Blind-DiffCom Achieves Pilot-Free Transmission
39 |
40 | 
41 |
42 | ## Requirements
43 |
44 | Clone the repo and create a conda environment (we use PyTorch 1.9, CUDA 11.1).
45 |
46 | TODO: check the dependencies
47 |
48 | ## Model Download
49 |
50 | We provide 3 pre-trained ADJSCC models in [this link](https://drive.google.com/drive/folders/1N0EzzxCv1wh6JeFr0g8vkmB0Qj23ozZJ?usp=sharing), please download them and put them in the `_djscc/ckpt` folder.
51 |
52 | The pre-trained Diffusion models are available at the following links, please download them and put them in the `model_zoo` folder.
53 |
54 | | Model | Download link |
55 | |-------------------------------------------------------------|:----------------------------------------------------------------------------------------------------------:|
56 | | 256x256_diffusion_uncond.pt(ILSVRC 2012 subset of ImageNet) | [download link](https://openaipublic.blob.core.windows.net/diffusion/jul-2021/256x256_diffusion_uncond.pt) |
57 | | ffhq_10m.pt | [download link](https://drive.google.com/drive/folders/1jElnRoFv7b31fG0v6pTSQkelbSX3xGZh?usp=sharing) |
58 |
59 | TODO: provide implementations of DiffCom based on \`SwinJSCC\` and \`NTSCC+\`.
60 |
61 | ## Inference Code
62 |
63 | ```bash
64 | python main_diffcom.py --opt ./configs/diffcom.yaml
65 | ```
66 |
67 | Please check the `diffcom.yaml` file for more details.
68 | The codebase now supports `diffcom`, `hifi_diffcom`, and `blind_diffcom`.
69 |
70 | You can change the hyperparameters in the YAML file to run the corresponding method.
71 |
72 | ## Acknowledgement
73 |
74 | The implementation is based on [DPS](https://github.com/DPS2022/diffusion-posterior-sampling), [PSLD](https://github.com/LituRout/PSLD).
75 | We thank the authors for sharing their code.
76 |
77 | ## Citation
78 |
79 | If you find this code useful for your research, please cite our paper
80 |
81 | ```
82 | @article{wang2024diffcom,
83 | title={DiffCom: Channel Received Signal is a Natural Condition to Guide Diffusion Posterior Sampling},
84 | author={Wang, Sixian and Dai, Jincheng and Tan, Kailin and Qin, Xiaoqi and Niu, Kai and Zhang, Ping},
85 | journal={arXiv preprint arXiv:2406.07390},
86 | year={2024}
87 | }
88 | ```
--------------------------------------------------------------------------------
/website/src/components/section1/Section1.js:
--------------------------------------------------------------------------------
1 | import React, {Fragment} from "react";
2 |
3 | const OverviewBlock = () => (
4 | 27 | End-to-end visual communication systems typically optimize a trade-off between channel 28 | bandwidth 29 | costs and signal-level distortion metrics. 30 | However, under challenging physical conditions, such a discriminative 31 | communication paradigm often results in unrealistic reconstructions with perceptible 32 | blurring and aliasing artifacts, despite the inclusion of perceptual or adversarial losses 33 | during training. This issue primarily stems from the receiver's limited knowledge about the 34 | underlying data 35 | manifold and the use of deterministic decoding mechanisms. 36 |
37 |38 | We propose DiffCom, a novel end-to-end generative communication paradigm that utilizes off-the-shelf 40 | generative priors from diffusion models for decoding , thereby improving perceptual 41 | quality 42 | without heavily relying on bandwidth costs and received signal quality. 43 | Unlike traditional systems that rely on deterministic decoders optimized solely for 44 | distortion 45 | metrics, our DiffCom leverages raw channel-received signal as a fine-grained condition to guide stochastic posterior 46 | sampling. Our approach ensures that reconstructions remain on the manifold of real data 47 | with a 48 | novel confirming constraint, enhancing the robustness and reliability of the generated 49 | outcomes. 50 | Furthermore, DiffCom incorporates a blind posterior 51 | sampling 52 | technique to address 53 | scenarios with unknown forward transmission characteristics. 54 |
55 |56 | Experimental results demonstrate that: 57 |
Slide the button for comparison
*/} 60 |Left: HiFi-DiffCom (with {method} Encoder); Right: {method}
61 |VTM + 5G LDPC
100 |Original Image
108 |182 | HiFi-DiffCom vs. 183 |
184 | {methods.map(t => ( 185 |Reconstruction
*/} 99 | {/*Original Image
111 |Estimated Channel Response Vector
117 |Ground Truth Channel Response Vector
123 |Measured on 20 samples. Mean {'$\\pm 1.0\\sigma$'} is displayed.
*/} 195 | {/*⎻⎻⎻ Drag time slider ⎻⎻→
211 |Progress of channel estimation-free transmission over a multipath fading channel with L = 8 paths.
212 |