fast_h3_vae support, 10 times faster
it decodes the 2560×1440 clip in 1.6 s instead of 15 s.
fast_h3_vae_noncausal_24l.safetensors Fast H3 VAE decoder from StreamRender-H3 (non-causal, 24 layers, 1.62B parameters)
https://huggingface.co/Efficient-Large-Model/Open-H3-Regenerate-2K
Sounds great. let me try.
It just gave me error loading vae when i replace the vae.
Vae is invalid. Are you asking Kijai to support this Vae?
Support is merged to ComfyUI master as of 6 hours ago.
I don't know where the 10x number comes from, it's not faster and has worse quality than our int8 VAE, but when used with int8 itself it's twice as fast:
encode - decode test:
I compared it to minimax_h3_lynnreal_light_vae_int8_convrot.safetensors and found it shows noticable horizontal banding. It's definitely fast though.
Perhaps they exaggerated and overestimated