multi-scale discriminator

**Multi-scale discriminator** is the **GAN discriminator design that evaluates generated images at multiple spatial resolutions to capture both global layout and local texture quality** - it improves critique coverage across different detail scales. **What Is Multi-scale discriminator?** - **Definition**: Discriminator framework using parallel or hierarchical branches on downsampled image versions. - **Global Branch Role**: Checks scene coherence, object placement, and structural consistency. - **Local Branch Role**: Focuses on fine textures, edges, and artifact detection. - **Architecture Variants**: Can share backbone features or use independent discriminators per scale. **Why Multi-scale discriminator Matters** - **Quality Balance**: Reduces tradeoff where models overfit either global shape or local detail. - **Artifact Detection**: Different scales catch different failure patterns during training. - **Stability**: Multi-scale signals can provide richer gradients to generator updates. - **Generalization**: Improves robustness across varying object sizes and scene compositions. - **Benchmark Gains**: Frequently improves perceptual quality in translation and synthesis tasks. **How It Is Used in Practice** - **Scale Selection**: Choose resolutions that reflect target output size and detail demands. - **Loss Weighting**: Balance discriminator contributions to avoid domination by one scale. - **Compute Planning**: Optimize branch design to control training overhead. Multi-scale discriminator is **an effective discriminator strategy for high-fidelity generation** - multi-scale feedback helps generators satisfy both global and local realism constraints.