Config of the Kandinsky6 SR causal video KVAE (Kandinsky6SRVAE).
vae/config.json comes in two layouts: an older one nesting the KVAE architecture under encoder_config / decoder_config dicts (the config stored the string "None" for their unused channel fields), and the current one with those same knobs flattened to the top level instead -- decoder_ch / decoder_ch_mult are the only decoder-specific overrides, everything else (num_res_blocks, z_channels, the temporal-compression / norm / padding knobs) is shared between encoder and decoder. update_model_arch copies whichever layout the checkpoint has; __post_init__ synthesizes encoder_config / decoder_config from the flat fields when the nested ones weren't populated, so Kandinsky6SRVAE.__init__ (and its _arch_kwargs validation) only ever has to handle the nested shape. scaling_factor multiplies the raw KVAE latent to get the latent space the SR DiT / latent upscaler work in.