Nodes/Comfyui-MMH3-UltimateUpscale/MMH3 Latent Upscale Params
ComfyUI Node

MMH3 Latent Upscale Params

No model, no download, just interpolation

By bbaudio-2025·Created 3 days ago·Updated 2 days ago· 88
MMH3 Latent Upscale Params
    • latent_upscale_param
    methodbilinear
    width1280
    height704

    The name oversells it a little. MMH3 Latent Upscale Params isn't a model - it's a bundle of settings that tells the MMH3 Ultimate Upscale node to resize the video latent with plain interpolation. Think ComfyUI's built-in Upscale Latent node, but smart enough to leave the audio half of H3's nested latent alone. If you don't have the H3 3D upscaler checkpoints, or you don't want to load a second model, this is the path you take.

    Why it's the right first choice

    Here's the thing about upscaling: there are two jobs hiding in the word, and only one of them needs an AI model (the KB makes this point well). Making pixels bigger is mechanical - interpolation. Inventing plausible detail that was never there is generation. The model-free route does the first job for free, in milliseconds, and then lets the re-sampling pass in MMH3 Ultimate Upscale do the second job anyway. Since the main node re-samples every chunk and tile through the diffusion model regardless, a cheap resize up front is often all the "upscale" you needed - the diffusion pass adds the detail. It's the latent-video equivalent of a 2x ESGAN-less hires fix.

    What you actually set

    Only three inputs, and they're all obvious:

    • method - the interpolation: nearest-exact, bilinear, area, bicubic. Bilinear is the default and a sane starting point; bicubic is a touch crisper if you want it.
    • width and height - the target frame size in pixels, both snapped to a multiple of 32.

    The one gotcha is printed right in the tooltip: the target size must match the conditioning's generation size - the resolution the clip was conditioned at after upscaling. If those disagree, the re-sample gets conditioning that doesn't line up with the latent, and you get warped or blurry frames. Get this wrong and no amount of method-swapping fixes it.

    The single output, latent_upscale_param, is a H3_UPSCALE_PARAM bundle that plugs into the latent_upscale_param input on MMH3 Ultimate Upscale. That's all it does - this node is a settings carrier, not a sampler.

    Install and notes

    It ships in the Comfyui-MMH3-UltimateUpscale pack, so it's the same one install: search Comfyui-MMH3-UltimateUpscale in ComfyUI Manager, or clone the repo into ComfyUI/custom_nodes and restart. No model download required - that's the entire point of this variant.

    One honest caveat: the pack is new, vibe-coded per its own README, and has zero track record in the wild yet. But this particular node is about as safe as a custom node gets - it's a thin wrapper around interpolation parameters. If the pack's behavior ever looks off, method/width/height are the only knobs you can turn, so there's not much to misconfigure beyond that size-matching rule.

    Categorymodel/latent/minimax

    Inputs (3)

    NameTypeDefaultDescription
    methodCOMBObilinearInterpolation used to resize the video latent's spatial HxW (same as Upscale Latent).
    widthINT128064–4096Target overall pixel width of the upscaled frame (snapped to a multiple of 32). Must match the conditioning's generation size.
    heightINT70464–4096Target overall pixel height of the upscaled frame (snapped to a multiple of 32). Must match the conditioning's generation size.

    Outputs (1)

    NameTypeDescription
    latent_upscale_paramH3_UPSCALE_PARAMModel-free upscale settings consumed by 'MMH3 Ultimate Upscale'.