Feature description

Add metrics for assessing image quality, including deep learning metrics.

• [x] L1/MAE loss, torch, torch-image-metrics
  • Closed via https://github.com/tracel-ai/burn/pull/4341
• [x] L2/MSE loss, torch, torch-image-metrics
  • Closed via https://github.com/tracel-ai/burn/pull/4341
• [x] PSNR, Wikipedia, scikit-image, PyTorch-ignite, torch-image-metrics
  • Closed via https://github.com/tracel-ai/burn/pull/4379
• [x] SSIM, Wikipedia, scikit-image, PyTorch-ignite, torch-image-metrics
  • Closed via https://github.com/tracel-ai/burn/pull/4396
• [ ] MS-SSIM, lightning.ai, pytorch-msssim
• [x] Smooth L1 torch
  • Closed via https://github.com/tracel-ai/burn/pull/4547
• [ ] Gram Matrix Loss, torch tutorial, piq
• [ ] FID, lightning.ai, pytorch-fid, torch-image-metrics
• [x] LPIPS, lightning.ai, torch-image-metrics
  • Closed via https://github.com/tracel-ai/burn/pull/4403
• [ ] DISTS, lightning.ai
• [ ] A-FINE, IQA-PyTorch

Feature motivation

When training visual AI models the loss function may include various image metrics. Users would have to re-implement these image metrics unless readily available.

Examples

Take for example the impressive recent ml-sharp which uses at least PSNR, SSIM, LPIPS, DISTS.

<img width="766" height="223" alt="Image" src="https://github.com/user-attachments/assets/44a586cf-4d3d-4e3a-b1d2-13bef73ed33b" />