MiniMax H3 · Native → RTX Final · Streaming

If you have an NVIDIA GPU with RTX Video Super Resolution available, this is the R42 route that turns your native H3 render into a properly upscaled 1080p/4K master. IAMCCS_MiniMaxH3UniversalRTXFinalR42 - Native → RTX Final · Streaming - takes the already-checkpointed native frames and audio, runs each frame through RTX VSR, and encodes straight to the final mp4. No intermediate upscaled image batch, ever.

The design constraint hiding in there is the interesting part. Look at the node's inputs and you'll notice it takes no model, no conditioning, no latent - only the checkpointed native_frames, native_audio, and the render bookkeeping. That's deliberate, and the code says why: with no denoise inputs, "ComfyUI cannot evaluate a second H3 denoise pass for this delivery path." The RTX route is a guarantee that H3 gets sampled exactly once, and everything after is pure post. If you want to re-sample, that's a different route.

The streaming part is how it stays cheap on memory. It opens a pipe to ffmpeg, then feeds frames one at a time through NVIDIA's RTX VSR runtime (nvvfx.VideoSuperRes): each native frame is cropped to the target aspect, sent to the upscaler, center-cropped to the delivery size, and written directly into the encoder pipe before the next frame loads. A 2K batch of tensors never materializes in RAM or VRAM, and it writes progress logs every 24 frames so a long render doesn't feel like a hang. Quality comes from the shot plan's settings (LOW/MEDIUM/HIGH/ULTRA, HIGH default), and the native audio is locked through - it writes a wav to the pipe and muxes it with the video.

Inputs and outputs

Required: cine_linx (route + delivery settings), native_frames and native_audio (from the checkpoint), resolved_render_id, native_saved_report, current_segment/total_segments, fps, join_trim_frames, and queue_next_segment (default on). Outputs: video_path and report. Segments land under output/IAMCCS/MiniMaxH3/RTX_FINAL/<run>/, and when the last segment finishes it concatenates everything into a final_film.mp4 and enqueues the next run if asked.

The honest take

This is the highest-fidelity, highest-dependency route in R42. It needs the NVIDIA RTX VFX runtime installed into your ComfyUI Python (the code raises a pointed error pointing at the RTX install guide if it can't import nvvfx), plus ffmpeg reachable on the system. And it refuses to overwrite - existing segment files are a hard error. If you're on an older GPU without RTX VSR, this route simply isn't available; pick the native or fast branch instead. If you do have the hardware, this is the "final polish" node - one clean native H3 sample, then hardware upscaling that holds up far better than naive bilinear resize at delivery size.

Install

The pack is installed once:

cd ComfyUI/custom_nodes
git clone https://github.com/IAMCCS/IAMCCS-nodes

or ComfyUI Manager (search "IAMCCS"), then restart - and additionally install the NVIDIA RTX VFX runtime for your ComfyUI Python and make sure ffmpeg is available. You also need the H3 stack: current ComfyUI (≥ 3.x), MiniMax H3 weights (~42 GB) and its video/audio VAEs. The H3 licence excludes the US, EU, UK and Korea, so check that before you start rendering.