Nodes/FunUI/FunUI Upscale Video
ComfyUI Node

FunUI Upscale Video

Upscale the Finished MP4 and Keep the Sound, No Graph Rebuild

By digital-garbage·Created about a year ago·Updated about 10 hours ago· 21
FunUI Upscale Video
      ◄filename►
      ◄subfolder►
      ◄typeoutput►
      ◄upscale_model▾►
      ◄rtx_options{}►

      Most upscaling in ComfyUI happens inside the graph, which means re-running the whole pipeline or rebuilding it around an upscale model. This node starts from the other end: a video file that already exists - one of yours, a render, something in output/ - and a chosen upscaler. It streams the file through frame by frame and writes a new MP4 with the original fps and sound.

      It's the "the clip is done, now make it bigger" node, and there should be more of them.

      What it actually does

      It resolves the file inside ComfyUI's output (or temp) folder, then picks an engine:

      • A model from ComfyUI/models/upscale_models - loaded through ComfyUI's own UpscaleModelLoader and run per frame with ImageUpscaleWithModel.
      • NVIDIA RTX Super Resolution - if you've turned on the nvidia_vsr add-on module, this option appears in the same dropdown and routes to NVIDIA's own video super-resolution instead of a neural upscale model.

      Then it streams: decode frames out of the source with ffmpeg, upscale in chunks, and pipe raw frames into an encoder that writes libx264 at crf 16 with +faststart. If the source has an audio track, it maps that in and re-encodes as AAC at 192 kbps - which is how the sound survives. Progress is interruptible, and on any failure the half-written output is deleted rather than left looking like a result.

      Output lands in a subfolder called funui_upscaled, named <yourfile>_<model>.mp4, with _1, _2 appended rather than overwriting anything.

      Inputs and outputs

      Required:

      • filename - the file, as it exists in the output (or temp) folder.
      • subfolder - where it lives in there, if anywhere. Empty for the root.
      • type - output or temp.
      • upscale_model - a combo. Model files from upscale_models (use .safetensors), or "NVIDIA RTX Super Resolution" when that add-on is enabled.

      Optional:

      • rtx_options - JSON for the RTX engine only, and only read when RTX is the chosen engine. Defaults are quality: ULTRA, resize: scale, scale: 2.0, strength: 1.0, tenbit: false, plus width/height when you set resize to size. The option names mirror NVIDIA's own RTX node so an existing setup carries over, and a value out of range is refused by name rather than quietly clamped.

      There are no outputs. It's a terminal node: it writes the file and shows it in the output panel.

      Installing it

      cd ComfyUI/custom_nodes
      git clone https://github.com/digital-garbage/FunUI
      cd FunUI && pip install -r requirements.txt
      

      ComfyUI Manager: search FunUI. Restart. (Pack page says olivv-cs/ComfyUI-FunPack; the repo is digital-garbage/FunUI. Renamed project, same nodes.)

      For the model path you need an upscale model in ComfyUI/models/upscale_models. For the RTX path, switch on the NVIDIA RTX Super Resolution add-on in Settings ▸ Modules: FunUI installs nvidia-vfx on the next ComfyUI start, and it's NVIDIA's package under NVIDIA's licence, so it wants an RTX card and driver. ffmpeg needs to be on the machine either way - Settings ▸ Ready to generate? checks that and the node folders for you.

      Where people get burned

      Per-frame models don't know about time. An image upscaler applied frame by frame has no idea what the previous frame looked like, so expect a little detail shimmer on motion, and more on anything with fine texture. It's inherent to the method, not a bug in the node - RTX Super Resolution and video-native models exist precisely because of it.

      4x on a long clip is a coffee break. Scale is quadratic in pixels: 2× is four times the work, 4× is sixteen. There's no resume; interruptions clean up and you start over.

      The file has to be where you say it is. filename + subfolder + type resolve inside ComfyUI's folders. A path from elsewhere on disk won't be found, and the error is the blunt "is not in the output folder (anymore)".

      JSON that isn't JSON. rtx_options is parsed strictly. An empty string means defaults; a trailing comma or single quotes gets you a refusal that names the problem. Also: won't apply while the add-on is off - the RTX option isn't even in the dropdown then.

      Expecting the sound to improve. Audio is a straight re-encode at 192 kbps AAC. Nothing here touches it, and nothing here syncs it - if the source audio was already out of step, it stays that way.

      Reaching for this when you wanted a second pass. Latent-space work between sampling passes is a different job, and doing it here means upscaling pixels the model can no longer improve. If the clip is soft because it was generated too small, resampling the latent is the better lever; this node makes a finished clip bigger.

      CategoryFunUI/Output

      Inputs (5)

      NameTypeDefaultDescription
      filenameSTRING—
      subfolderSTRING—
      typeCOMBOoutput2 options: output, temp
      upscale_modelCOMBOFrom ComfyUI/models/upscale_models (use .safetensors), or NVIDIA RTX Super Resolution.
      rtx_optionsoptSTRING{}NVIDIA RTX Super Resolution options as JSON: quality, resize, scale, width, height, strength, tenbit.

      Outputs (0)

      No outputs