MiniMax H3 RTX VSR 4K Post

MiniMax H3 renders are generated small - the Shotboard's proven 12GB route works at 960×544, roughly half a megapixel - and then you stare at a 1080p timeline and want pixels. That's exactly what this node is for. IAMCCS_MiniMaxH3RTX4KPost is the optional last step in the H3 delivery chain: it takes your generated frames and runs them through Nvidia's RTX Video Super Resolution (VSR) to a true 4K master, using the GPU's hardware path rather than another diffusion model.

It's a lazy node in the good sense. You don't wire it into the happy path and eat the cost every time; the Shotboard's delivery router only calls it when the plan actually asks for 4K. The node reads your shotplan out of the cine_linx contract, checks the upscale_settings block - target width/height default to 3840×2160, gated by ltx_4k_enabled - and if 4K isn't requested it just returns your frames untouched with a "RTX VSR 4K off" report. Set that gate and you get the real thing.

How it works

The mechanism is worth knowing because it explains the one thing people trip on. It computes the scale factor from your source to the 4K target and requires it to sit between 1× and 4× - that's the hard limit of Nvidia's VSR node it delegates to (RTXVideoSuperResolution, from the Comfy-Org/Nvidia_RTX_Nodes_ComfyUI pack). Then it does the smart part: instead of pushing the whole clip through as one giant float32 tensor, it unloads the diffusion model, clears VRAM, and feeds the video through in chunks of at most 8 frames. The 4K output is kept on CPU as float16, so a 241-frame 1080p clip doesn't blow your system RAM. Frames go in as RGB float32, which is what NvVFX expects.

Inputs and outputs

It's a deliberately small interface:

• images (IMAGE) - your decoded H3 frames, as a frame batch.
• cine_linx (IAMCCS_SUPERNODE_LINX) - the contract from the Shotboard that carries the upscale settings.

Outputs are images_4k and a report string that tells you what scale and quality actually ran. Wire images_4k into your video saver.

The catch

Two things to know before you build this into a workflow. First, it needs an RTX card - VSR runs on Nvidia's Video Effects API, not on the generic CUDA path, so this is an Nvidia-only node. Second, VSR is a realtime-ish scaler, not a generative upscaler. That's a feature on already-clean H3 output: the community reads it as more natural than diffusion upscalers on clean sources, and it's seconds rather than minutes. But it won't invent pores or fabric weave the way a SeedVR2-class model will. You're buying pixels, not new detail.

The common failure is a scale error: if your source is already close to 4K, or your target is set beyond the 1×–4× window, the node raises rather than silently resampling. The practical fix is to keep the Shotboard's upscale target at 3840×2160 and let H3 do its small-native render. And if you get "requires node RTXVideoSuperResolution," that's the Nvidia nodes pack missing - install it and restart ComfyUI.

Install the pack itself through ComfyUI Manager (search "IAMCCS"), or cd ComfyUI/custom_nodes && git clone https://github.com/IAMCCS/IAMCCS-nodes.git. The RTX VSR node it depends on is a separate add-on.