Compare SD and HD

AI enhancement is a separate step after video generation. It keeps the original video and adds an enhanced version to the same take, rather than creating another library card.

Choose an AI upscaler

Bundled upscalers run through Vpipe or ComfyUI. Connect the service you want to use in Settings → Services; it can differ from the service that generated the original video.

In Settings → Generation → Video upscaling, select a workflow:

Upscaler	Approach	Requirements
Real-ESRGAN	Enhances frames independently	ComfyUI built-in model upscaling and video nodes
SeedVR2	Restores batches of frames with temporal context	ComfyUI video nodes and the SeedVR2 extension
FlashVSR v1.1	Restores a clip with temporal context	Vpipe 0.1.51+ on Apple Silicon
VOSR 2.0	Restores frames independently	Vpipe 0.1.51+ on Apple Silicon

For the native options, select the workflow and use its Prepare models action if models are missing, or install the declared models in Vpipe yourself. FlashVSR needs its source projection, Wan VAE and companion prompt tensor; VOSR also needs DINOv2. The starters fetch those dependencies through the ordinary queue. Update Vpipe first: older versions do not support these models. FlashVSR buffers the clip and can use substantial memory; VOSR can vary between frames.

For SeedVR2, install ComfyUI-SeedVR2_VideoUpscaler in the ComfyUI installation you connected, following its ComfyUI extension instructions, then restart ComfyUI.

Install the workflow's models in ComfyUI using the model setup guide, then choose Refresh workflows in Frok. Frok does not install separate upscaler runtimes, Python environments or NCNN binaries.

If required nodes are missing, follow the workflow's Needs attention action, install or update the required nodes in ComfyUI, then restart ComfyUI and refresh readiness in Frok. Frok does not install ComfyUI models or custom nodes. Wait for the selected workflow to be ready before upscaling.

Frok supplies the source video and uses the SeedVR2 workflow's device bindings to select an available MPS or CUDA GPU reported by the connected ComfyUI service.

Enhance one take

1. Open a finished 480p video.
2. Choose Upscale in the shared action row below the video prompt. This action is available for SD videos, whether generated from an image, text, or references.
3. Follow the upscaling job in Queue. It may take longer than expected on a small GPU.
4. When finished, use SD / HD over the video to compare versions.

The selected workflow enhances the frames and rebuilds the video with its source frame rate and audio. The improvement is model-dependent: enhancement may sharpen texture or restore detail, but can also introduce artifacts. Compare faces, fine patterns and temporal consistency before keeping the result.

Compare and export

The viewer keeps both video elements available and switches visibility, carrying playback position across. The HD version shares the SD take's URL and render number. Download offers both versions. The bundled workflows preserve the original frame rate, audio and duration.

Changing the selected upscaler allows another enhancement from the original SD video; it does not feed an already enhanced result back through a different model. The viewer shows the latest enhanced copy for that take.

If rendering fails, check the job log and the selected service’s model or node requirements. For memory errors, stop other GPU work or select a compatible workflow with lower memory requirements. There is no silent switch to a different enhancement model.

Run directly in ComfyUI

1. Import run.json from the upscale/seedvr2/ or upscale/realesrgan/ folder into ComfyUI, with its required nodes and models installed. meta.json registers the workflow with Frok; import only the run graph.
2. Upload your source video or copy it into that ComfyUI installation's input folder, then select it in LoadVideo, replacing the input.mp4 placeholder.
3. For SeedVR2, select your GPU in both the DiT and VAE model loaders. The bundled file selects mps for Apple Silicon; CUDA users must choose their CUDA GPU in both loaders. Keep offload_device set to none; the bundled workflow has no CPU offloading.

Real-ESRGAN's final ImageScale defaults to height 720 and width 0, preserving the source aspect ratio. When run through Frok, width and height come from the job's bindings. Keep the source audio and frame-rate connections to CreateVideo intact. See bundled workflow settings.

Prepare native upscalers manually

Update to Vpipe 0.1.51 or newer first. These are the same starters queued by Prepare models. Copy the appropriate pipeline into a file of your choice, then run it from the Vpipe workspace configured in Frok:

cd ~/vpipe
vpipe --launch /absolute/path/to/prepare.vpipeline

Use your configured workspace if it differs from ~/vpipe. After it finishes, choose Refresh workflows in Frok.

FlashVSR v1.1

{
  "id": "prepare-flashvsr",
  "stages": [
    {
      "id": "fetch-flashvsr",
      "type": "model-fetch",
      "config": {
        "model_path": "JunhaoZhuang/FlashVSR-v1.1",
        "hf_token": "",
        "base_path": "./models",
        "skip_existing_files": true,
        "overwrite_existing": false
      },
      "iports": []
    }
  ]
}

VOSR 2.0

{
  "id": "prepare-vosr-2",
  "stages": [
    {
      "id": "fetch-vosr-2",
      "type": "model-fetch",
      "iports": [],
      "config": {
        "model_path": "CSWRY/VOSR",
        "model_variant": "VOSR-2.0",
        "hf_token": "",
        "source": "",
        "base_path": "./models",
        "skip_existing_files": true,
        "overwrite_existing": false
      }
    },
    {
      "id": "fetch-dinov2",
      "type": "model-fetch",
      "iports": [
        {
          "src": "fetch-vosr-2",
          "oport": 0
        }
      ],
      "config": {
        "model_path": "facebook/dinov2-large",
        "hf_token": "",
        "source": "",
        "base_path": "./models",
        "skip_existing_files": true,
        "overwrite_existing": false
      }
    }
  ],
  "subpipelines": []
}