I asked Claude because it fixed the issue for me. Might be helpful for anyone else encountering the same problem:

Bug: The Preview Override node decides whether it needs to "unpack" the sampler's latent before decoding it into a preview image. It only did this when latent_shapes had more than one entry — that check was meant to detect models with separate video and audio streams.

MiniMax H3 packs video and audio into a single combined stream rather than two separate ones, so it only reports one latent_shapes entry. That meant the unpack step was being skipped, and the node ended up handing the raw, still-packed latent straight to the tiny VAE decoder. The decoder happily "decoded" it anyway, just from the wrong data layout — which is why you got a plausible-looking image that never changed: it wasn't reading the actual denoising progress, just some fairly static structural noise in the packed layout.

Fix: Changed the check so it unpacks whenever there's any latent_shapes data at all, not just when there's more than one entry. That covers single-stream packed models like H3 as well as the multi-stream ones it already handled.

One line, for the issue:

python

before

unpack_x0 = is_ltx or (latent_shapes is not None and len(latent_shapes) > 1)

after

unpack_x0 = is_ltx or bool(latent_shapes)

That's weird, the latent_shapes for H3 is always dual, I don't know how it could be anything else, that change shouldn't do anything... and obviously it works for me and everyone else I know that has tested it.

It works for me too without any code changes. I needed to do Reload Node for the new tiny_vae field to appear. First I was connecting VAELoader KJ with taeh3.safetensors (tiny_vae field was not visible yet), which was obviously wrong.
Just a note for newcomers.

Thank you Kijai!