This is a LatentDiffusion model for frame prediction.

Updated autoencoder: Uses 16x downsampling factor (updated from 8x)

Original checkpoint: ./test_scheduled_samplingk2777doom222222_pretrainrnn22_nibik/model-step=032000.ckpt.state_dict_only

Autoencoder checkpoint: ../autoencoder/saved_kl4_bsz8_acc8_lr4.5e6_load_acc1_512_384_mar10_keyboard_1024_768_4.5e7_cont_1e5_forui_real_4.5e6_4.5e7_gan_p1.0/model-step=732000.ckpt