512x512 Diffusion (Architecture fine-tuned)
Detailed description
A 512x512 unconditional ImageNet diffusion model, fine-tuned for 900.000 samples from the 512x512 unconditional ImageNet diffusion model. It was fine-tuned using 60.000 images of architecture of the AIDA dataset from Harvard x ArchDaily.
Config (as used in Disco Diffusion)
{ "attention_resolutions": '32, 16, 8', "class_cond": False, "diffusion_steps": 1000, "image_size": 512, "learn_sigma": True, "noise_schedule": "linear", "num_channels": 256, "num_head_channels": 64, "num_res_blocks": 2, "resblock_updown": True, "rescale_timesteps": True, "timestep_respacing": "250", "use_scale_shift_norm": True }