zhengchen1999
commented
11 months ago Regarding the differences between HI-Diff and DiffBIR:
We indicate in our paper that we reference the paradigm of DiffBIR, which combines Diffusion and Transformer. The distinction between HI-Diff and DiffIR lies in the hierarchical integration: multi-scale prior and cross-attention interaction, which is more apt for non-uniform realistic deblurring.
The prior is different. HI-Diff applies the multi-scale prior, while DiffIR uses the single-scale prior. The multi-scale prior adapts to different scale features in the encoder-decoder Transformer architecture for better integration.
The interaction is different. HI-Diff adopts cross-attention, while DiffIR uses prior as dynamic modulation parameters. Features pay different attention to the prior with cross-attention, which is more suitable for non-uniform deblurring.
Our method outperforms DiffIR with comparable parameters and FLOPs. This indicates that our approach, utilizing hierarchical integration, is more effective for realistic non-uniform deblurring.
| <!DOCTYPE html> Method | Params (M) | FLOPs (G) | PSNR (dB) | SSIM |
|---|---|---|---|---|
| DiffIR | 26.94 | 120.99 | 33.20 | 0.963 |
| HI-Diff (ours) | 28.49 | 142.62 | 33.33 | 0.964 |
| HI-Diff-2 (ours) | 23.99 | 125.47 | 33.28 | 0.964 |
PS: For [3], I am not very familiar with it, but applying this paradigm to other task is also a form of exploration.
Riras-Misaka
wangxinzhe0617
[1] Diffir: Efficient diffusion model for image restoration. ICCV, 2023.
[2] Hierarchical Integration Diffusion Model for Realistic Image Deblurring. NeurIPS, 2023.
[3] Reti-Diff: Illumination Degradation Image Restoration with Retinex-based Latent Diffusion Model. arXiv preprint arXiv:2311.11638, 2023.
[1,2,3] almost adopt the same network structure and training strategy, and the novelty is consistent.