Could you release an FP32 vae?
Just to try to squeeze out every last bit of quality, an extra 5gb of storage space doesn't bother a lot of people!
go download that for the panacea, you will see NO difference because autoencoders are insensitive as long as they're not quantized (fp8)
Problem isn't the storage space, it's a diffusion VAE and running it in fp32 would be very slow, with no real benefit, even the original code uses fp16 for it.
If anyone still wants to try it for yourself anyway, you can find it here: https://huggingface.co/MiniMaxAI/MiniMax-H3/tree/main/FL2VA/video_vae/source
After downloading model.safetensors rename it to minimax_h3_video_vae_fp32.safetensors. SHA256 is exactly the same for Ref2VA, so no need to download it twice. As Kijai pointed out, it doesn't appear to be improving quality at all. I'm not sure if it's being downcasted anywhere as I didn't dive that deep.