lpips

**LPIPS** is the **Learned Perceptual Image Patch Similarity metric that measures perceptual difference using deep feature activations instead of raw pixel error** - it is widely used for image restoration and generation quality evaluation. **What Is LPIPS?** - **Definition**: Feature-space distance metric computed between corresponding patches in two images. - **Perceptual Basis**: Uses pretrained network representations to approximate human visual similarity judgments. - **Comparison Mode**: Primarily full-reference metric requiring target and generated image pairs. - **Task Coverage**: Applied in super-resolution, deblurring, translation, and synthesis benchmarking. **Why LPIPS Matters** - **Perceptual Fidelity**: Better captures visual similarity than pixelwise metrics in many tasks. - **Training Guidance**: Can serve as optimization objective for perceptually plausible outputs. - **Benchmark Utility**: Helps compare models where multiple plausible reconstructions exist. - **Artifact Sensitivity**: Detects structural and texture differences overlooked by PSNR or MSE. - **Model Selection**: Supports choosing outputs that align with human quality preferences. **How It Is Used in Practice** - **Reference Pairing**: Evaluate LPIPS on well-aligned reference-generated image pairs. - **Metric Mix**: Use together with distortion and realism metrics for balanced assessment. - **Domain Calibration**: Validate correlation with human ratings on target application data. LPIPS is **a standard perceptual-distance metric in vision model evaluation** - LPIPS provides strong perceptual signal when used within a broader metric portfolio.