Minimax_H3-LatentUpscaler

MiniMax H3 is the open-weight video model Hailuo dropped in mid-2026 - text, image, video and audio in one latent, up to 15 seconds at 2K. It also has a quirk that will eat your afternoon the first time you touch it: its video VAE compresses space 16x, not the 8x every image model you've used does. If you grab ComfyUI's stock LatentUpscale, it cheerfully divides your target by 8 and your upscale comes out half the resolution you asked for. This node from the ComfyUI_Element_easy pack exists to not do that.

Mechanically it's thin, which is the point. You feed it the H3 latent and a target size in pixels; it divides width and height by 16 to get the true latent dimensions, then calls ComfyUI's own common_upscale - the exact same interpolation code the built-in node uses, just with the right factor. Set either width or height to 0 and it preserves the aspect ratio. It also understands H3's NestedTensor format, which mixes a 5D video tensor and a 4D audio tensor in one latent: the video gets rescaled, the audio passes through completely untouched.

Only a few inputs matter. samples is your H3 latent (grab it after the sampling steps, before decode - or use this mid-sampling for a two-pass upscale). width and height are pixel values stepped by 16 - this is the single most common beginner trap, people type in the latent dimensions they saw in a preview. upscale_method offers the standard five (nearest-exact, bilinear, area, bicubic, bislerp); for latents you usually want bislerp, same as you'd pick in any video latent upscale. crop is disabled or center, and you'll leave it disabled almost always. The single LATENT output feeds straight back into your sampler for a second pass, or into the VAE decode if you're done.

Install is the same as for every node in this pack - it's a grab bag of convenience nodes by supElement:

cd ComfyUI/custom_nodes
git clone https://github.com/supElement/ComfyUI_Element_easy

Or in ComfyUI Manager, search ComfyUI_Element_easy and install, then restart. No model downloads for this node, and while the pack's requirements.txt pulls in kornia and opencv-contrib-python, those exist for its image-blending sibling - this upscaler is pure torch plus ComfyUI internals.

Where people get burned: the aspect-ratio auto path (width or height = 0) computes from the video tensor, so if you feed it a latent with no 5D video component it raises "No video tensor (5D) found in latent" - that's your cue you grabbed the wrong output. And remember the audio tensor is a passenger here: this node will happily upscale video that had sound attached without touching the audio, which is exactly what you want, but don't expect it to resize audio to match. For the shape-mismatch headache that comes after upscaling, when the conditioning still describes the old resolution, that's what the _Adv sibling in the same pack is for.