Skip to main content

Table 2 The architecture of D1/D2/D3 network. Annotations are the same as Table 1. The different layers of D1, D2, and D3 are listed separately

From: Stacked generative adversarial networks for image compositing

The discriminative model D1/D2/D3
Input D1: Image (512×512×3)/ D2,D3: Image (256×256×3)
Layer 1 D1,D2: Conv. (3, 64), st=2; LRelu;/
D3: Conv. (3, 32), st=2; LRelu;
Layer 2 D1,D2: Conv. (64, 128), st=2; IN; LRelu;/
D3: Conv. (32, 64), st=2; IN; LRelu;
Layer 3 D1,D2: Conv. (128, 256), st=2; IN; LRelu;/
D3: Conv. (64, 128), st=2; IN; LRelu;
Layer 4 D1,D2: Conv. (256, 512), st=1; IN; LRelu;/
D3: Conv. (128, 256), st=1; IN; LRelu;
Layer 5 D1,D2: Conv. (512, 1), st=1; Sigmoid;/
D3: Conv. (256, 1), st=1; Sigmoid;
Output D1: Real or Fake (62×62×1)/ D2,D3: Real or Fake (30×30×1)