MiniMax H3 R38 NVIDIA RTX Delivery

The R38 latent-upres pipeline has an optional final stage: instead of stopping at the latent upscale's output resolution, you can hand the frames to NVIDIA RTX Video Super Resolution - the same hardware upscaler inside the RTX Video / VSR SDK that the community keeps comparing against generative restorers - to get from the stage resolution up to the final delivery size. IAMCCS_MiniMaxH3LatentUpresRTXR38 is that leg.

If that sounds familiar, it's because this mirrors a widely-discussed 2026 pattern: people run a generative pass for structure and then an NVIDIA interpolation upscale for the last step, because the hardware path is fast, stable across frames, and - per the "NVIDIA results look more natural" camp in the upscale debates - doesn't hallucinate new detail the way generative restorers can. In the R38 design the latent upscaler does the generative work at half resolution and RTX VSR carries the result to full delivery resolution. Two upscalers, each doing the job it's good at.

How it works

• Reads the plan and delivery size off cine_linx, and the RTX quality from the upscale settings (ULTRA default, mapped to VSR modes: ULTRA → VSR Ultra, HIGH → VSR High, MEDIUM → VSR Medium, LOW → VSR Low).
• If frames are already at delivery resolution it passes them through untouched (no pointless hardware pass).
• Otherwise it runs each frame through the VideoSuperRes effect via IAMCCS's RTX runtime, resizing to the delivery width/height with center-crop fill, consuming each DLPack tensor view immediately so VRAM stays flat.

Inputs

• cine_linx - where the delivery size and RTX quality come from.
• images - the frames to upscale (typically the assembled master from the chunk checkpoint, at half delivery resolution).

Outputs

• images - the upscaled frames at delivery resolution.
• report - which VSR mode ran and the target size.

Install

The pack install (git clone https://github.com/IAMCCS/IAMCCS-nodes.git) plus the RTX Video Super Resolution runtime that IAMCCS's iamccs_rtx_vfx module imports. That runtime is NVIDIA hardware- and driver-dependent - RTX GPUs with the VSR SDK available. On non-NVIDIA or unsupported setups, expect the RTX path to fail; the workflow's native route is your fallback.

Gotchas

• This is a decide-by-plan node: it only does real work when the plan says RTX is part of the delivery. In the default (RTX off) route, don't expect it to upscale anything - the R38 router sends native/upscaled frames elsewhere.
• The rtx_quality you set in H3 Settings maps to VSR quality levels here; there's no separate per-node quality widget.
• Center-crop fill means source aspect must match delivery aspect or you'll lose edges. The latent stage is planned at the right aspect by the Control node, so keep those in lockstep.
• It's the "finish" leg: feed it pre-upscaled frames and it does a light/no-op pass; feed it half-res frames and it's doing the hardware heavy lift. Position matters.