RTX Video Super Resolution Upscaler workflow for videos

RTX Video Super Resolution (4x, non-diffusion)

GPU-accelerated upscale using NVIDIA's Video Effects SDK (nvidia-vfx), wrapped by the RTX Video Super Resolution ComfyUI node. No diffusion, no frame interpolation, no denoise step - it is a dedicated super-resolution model that runs frame-by-frame on the GPU.

Requires an RTX GPU + the comfyui-rtx-simple custom node (installs the nvidia-vfx Python wheel). See the install notes shipped alongside this file for exact versions.

Node-by-node

1. Load Video - pick the source MP4.

2. Get Video Components - demuxes the video into frames (IMAGE batch), audio (AUDIO, untouched), fps (FLOAT), and bit depth (INT). fps and audio are piped straight through to the encoder later, unmodified - that's what preserves original frame rate/duration/audio.

3. Get Image Size - reads the width of the first decoded frame, used below to decide whether upscaling is even needed.

4. Primitive Int (1920) - the "already large enough" width threshold. Change this value if you want a different skip cutoff.

5. ImpactCompare (a > b) - outputs true if width > 1920px. 6/8. VRAM Debug (before / after) - passes frames through unchanged; logs free VRAM to the ComfyUI console at each point (empty_cache/gc_collect left off so it only reports, it doesn't force a cleanup mid-pipeline). This is how VRAM usage gets surfaced - there's no in-graph VRAM gauge node, but ComfyUI-Crystools (if enabled) also shows a live VRAM/RAM bar in the top toolbar during the whole run.

6. RTX Video Super Resolution - the actual upscale. scale = upscale factor (default 2.0x). quality = LOW/MEDIUM/HIGH/ULTRA (default ULTRA, prioritizes quality per the brief). Internally this node auto-batches frames to stay under a fixed ~16.7-megapixel-per-batch VRAM budget - see the install notes for why there's no separate tile-size widget on this node.

7. If/Else Switch - if the source was already >1920px wide, passes the original decoded frames straight through (upscale branch never executes - lazy-evaluated, so this really does skip the work, not just discard its result). Otherwise passes the upscaled frames.

8. Create Video - reassembles frames into a VIDEO using the ORIGINAL fps and ORIGINAL audio pulled from step 2.

9. Save Video - encodes to MP4/H.264, re-encode mode, CRF 17 (mid-point of the requested 16-18 visually-lossless range). Edit filename_prefix to set both the output subfolder and filename, e.g. upscaled/my_clip saves to output/upscaled/my_clip_00001.mp4.

What's intentionally NOT here

• No diffusion/KSampler anywhere - this is a pure CNN super-res pass.

• No frame interpolation - frame count in equals frame count out.

• No denoise/sharpen node - the RTX SR model's own internal processing is the only thing touching pixel content; nothing extra layered on top.

• No explicit color-space conversion node - ComfyUI's IMAGE tensors and video codec path use standard sRGB/Rec.709; there's no core node to force a different color space. See troubleshooting notes if your source uses HDR/Rec.2020.