| # pixel resolution of the model | |
| resolution: 512 | |
| # train 4 channel VAE latents instead of RGB at 1/8 resolution | |
| vae_latents: true | |
| # number of channels at each level of the UNet | |
| channels: [128, 256, 384, 512] | |
| # which levels have attention | |
| attention: [0, 1, 2] | |