Minimax H3 Latent Upscaler (2D)

If you only want to make a MiniMax H3 clip bigger and don't care about the fancy stuff, this is the node. The Minimax H3 Latent Upscaler (2D) is a LATENT → LATENT box: feed it an H3 latent, tell it how much to upscale, and it hands back a bigger latent that you decode once. No VAE encode/decode round-trip, no second sampling pass, no API key, nothing. It's the simplest node in the pack, and for a lot of the two-stage strategy that's exactly what you want.

Quick context if you're new to MiniMax H3: it's the 33B open-weights video model that landed in August 2026, generating 4–15 second clips at up to 2K/24fps with native audio, all in a 24-channel latent. Generating at 2K from scratch is slow and VRAM-hungry, so the standard play is to generate small and upscale. That's where this node sits: it's the cheap "more pixels" step in the middle of the pipeline.

How it works

Don't read "2D" as "per-frame." This is a learned neural upscaler - the same 24-channel architecture family as the 3D variant, 12 in/out blocks at 512 base channels - and it includes temporal convolution layers (every second block, kernel 5) so consecutive frames stay coherent. Spatial scaling uses bilinear feature interpolation inside the network, but the temporal layers are what keep it from looking like a smeary per-frame resize. It's a small model, trained on roughly 80,000 paired low-res/high-res latents, most of them video at 2×.

One honest caveat from the README: the learned upscale saves time, not VRAM. If your workflow runs a second H3 refinement pass afterwards, that pass executes the full transformer at the enlarged resolution - doubling latent W and H means roughly four times the video tokens per step. The upscale itself is cheap; the refinement after it may not be.

The inputs, all four of them

• latent - your H3 latent. Output LATENT goes to VAE Decode (or into a refinement pass, if that's your thing).
• model_name - dropdown of checkpoints in ComfyUI/models/latent_upscale_models/. Empty until you install one; the loader auto-detects the architecture, so you just pick the file.
• scale (default 2.0) - the spatial upscale factor, 1.0 to 4.0. The model was trained mostly at 2×; 2× is where it's happiest. Pushing 4× works but is past the sweet spot.
• device and precision - cuda/cpu and fp32/fp16/bf16. Note this node defaults to fp32 where the 3D variant defaults to fp16; fp16 is fine if you want the speed, and you can flip back if anything looks off.

That's it. No alignment, no megapixels, no aspect-ratio logic - the 3D node owns all of that. This one is the minimal version.

Install

Same as every node in this pack:

cd ComfyUI/custom_nodes
git clone https://github.com/xmarre/Comfyui_Minimax_h3_latent_Upscaler.git

Restart ComfyUI (or search "Minimax H3 Latent Upscaler" in ComfyUI Manager). The only dependency beyond stock ComfyUI is einops. The step people actually forget is the weights: download the checkpoint from LBH-123-AI/Minimax_h3_latent_Upscaler on HuggingFace and put it in ComfyUI/models/latent_upscale_models/. Until you do, model_name is just an empty dropdown.

Which node should you use?

Honestly? If your source is a video, reach for the 3D variant - it runs the full temporal sequence in one pass and gives you the alignment and sizing controls. This 2D node's niche is the quick, no-options upscale: a 1280×704 image-style latent, a test render, or a workflow where you want the minimal surface area. There's no wrong answer between the two for quality on short clips; the 2D one just has fewer dials to turn and defaults to fp32. Pick based on how much control you want, not on which is "better."

The name is the spec: it upscales H3 latents, in 2D, with weights. Wire it between your H3 latent and VAE Decode and you're done.