Nuke Upscale
StyleFrame Team7 min read
In Nuke, "upscale" can mean three things. The first is the built-in Upscale node, a pre-trained machine learning node that doubles the input format (1920x1080 becomes 3840x2160). The second is a Reformat node with a sharper filter, which is the low-effort baseline. The third is a plugin or a trained network, such as NNSuperResolution, REZup, or a CopyCat model, for shots where the stock result isn't enough.
Which one you reach for depends on the footage, the target resolution, and how much temporal stability the shot needs.
What the Nuke Upscale Node Does
The Upscale node takes a single Source input and returns an image at twice the resolution. There is no scale slider. It is always 2x, so for other target sizes you follow it with a Reformat or choose a different tool.
Its controls are minimal:
- Use GPU if available: on by default. Rendering runs on the local GPU when one is found, and the Local GPU field shows which card is in use. It reads "Not available" if Nuke is set to CPU in the Preferences, no suitable GPU exists, or the GPU can't create a processing context (for example, not enough free memory). Changing the GPU in the Preferences requires a Nuke restart.
- Tile Size: None, 1024, 2048, or 4096. With None, the full frame is processed in one pass, which can eat a lot of GPU memory on large images. Tiling lets you process big frames on a machine with limited VRAM.
If you plan to render from the command line, enable the GPU option and use the --gpu flag. Leaving "Use GPU if available" on in a script also means the script will use a GPU whenever it is opened on a machine that has one.
How to Upscale a Shot with the Upscale Node
- Read your plate and set the colorspace correctly. Upscaling a mis-tagged plate bakes the error into every downstream node.
- Add an Upscale node directly after the Read, or after your denoise if the footage is noisy. Noise gives a network false detail to chase.
- Set the GPU option and tile size. Start with None. If you hit memory errors on a UHD or larger frame, step down through 4096, 2048, and 1024.
- Compare against a baseline. Put a Reformat with a conventional filter beside it and flip between them in a Viewer wipe. Zoom to 200% or more on edges, hair, text, and fine texture.
- Check motion, not just stills. Play a range of frames. A result that looks sharp on frame 1 can shimmer in sequence.
- Finish with a Reformat if needed. For a non-2x target, resize after the Upscale rather than before.
The Upscale node isn't flashy. Many AI upscalers hallucinate, redrawing grass or faces to look sharper, which is fine for a casual enhancement but dangerous when fidelity matters. A conservative upscaler that makes the image cleaner without inventing content is usually easier to trust in a composite.
The Reformat Baseline: Cheap and Controllable
Before reaching for machine learning, know what Reformat gives you. Its filter menu defaults to cubic, and a cubic reformat to 4K mostly just gives you a softer image with more pixels.
Sharper filters such as sinc4, lanczos4, and lanczos6 return more edge detail. The tradeoff is ringing: on high-contrast edges, like bright windows against dark interiors, sharper filters can push values negative and produce dark halos or darkened highlights. If you see dark lines, sample the pixels and check for negative values.
The usual remedy is to change the color space of the image before the Reformat. Test this on your own plate and verify the highlights afterward.
Reformat is also a legitimate choice for CG. If you render at 80% to save time and plan to soften the result in the grade anyway, a lanczos Reformat back to full size can be enough.
Third-Party Upscaling Plugins for Nuke
When the stock node isn't enough, two plugins come up often.
NNSuperResolution (Pixelmania)
NNSuperResolution is a 2x or 4x upscaler that works directly in your comp. Its main distinction is two modes. Still mode is meant for photos, DMP patches, and textures, and it produces sharper results but may flicker on moving material. Sequence mode is aimed at plates, final composites, and CG renders, prioritizing temporal stability. Other notes:
- It processes all RGB channels, supports multichannel layers, and can handle RGBA CG renders in sequence mode.
- It supports high dynamic range and native overscan handling (though bounding boxes can't change from frame to frame in sequence mode).
- It is CUDA-accelerated, so it needs an NVIDIA card, and it stitches inference patches internally so high resolutions don't demand as much VRAM.
- It supports Nuke 11.3+ on Linux, Nuke 12.0+ on Windows, and Nuke Indie, and runs sandboxed with no internet connection.
You can download it to test on your own footage, and you can request a free trial license that runs without the added watermarks or noise for a limited time.
REZup (RE:Vision Effects)
REZup is a machine learning upscaler that runs as an OpenFX plugin, so it also works in a range of other hosts. It differs from the stock node in a few useful ways:
- It supports fractional scale factors, so you can use it for digital zoom, and it can enhance without changing dimensions.
- It is alpha-weighted and supports extended dynamic range, including values below 0 and above 1.0.
- It has separate controls for sharpness, contrast, and noise reduction, so it isn't a black box.
- Its models are trained on 2x, 3x, and 4x, and larger factors may need preprocessing such as denoising.
REZup also has a free trial that watermarks output until you license it.
Training Your Own Upscaler with CopyCat
For a single hero shot, such as punching into a plate to create a close-up the crew never shot, you can train a network on the footage itself. A shot-specific model can beat a general one because it learns how your plate's grain, lighting, and detail behave.
The trick is that you don't train CopyCat to upscale directly. You train it to turn a badly upscaled image into a good one. The setup runs roughly like this:
- Find a frame range that is close up and lit like your crop.
- Downscale that footage with a cubic Reformat to roughly the resolution of your punched-in crop.
- Add a Grade that changes nothing, simply to stop concatenation.
- Reformat back up to the original resolution with the filter set to impulse. This gives you the pixelated input, with the original high-res plate as ground truth.
- Use a FrameRange and FrameHold to pick four frames, then crop each one on the details you want preserved. Append them, and duplicate the same crops on the ground-truth side.
- Connect the pixelated version to the CopyCat input and the original to ground truth.
- Because you are starting from the Upscale node's initial weights, a low epoch count is enough (about 2000 in the walkthrough this method comes from). A crop size of 32 works because upscaling is a local effect that needs little global context, unlike something like human matting.
After training, apply the result to your punched-in crop with a Reformat feeding the trained network. Treat the epoch and crop numbers as starting points and adjust for your footage.
Quality Control: What to Check on Any Upscale
- Temporal stability. Super-resolution and inpainting-style methods are prone to frame-to-frame fluctuation. Always review in motion at full resolution.
- Invented detail. Zoom in on faces, text, and logos. If the network is redrawing them, reconsider.
- Edge halos and negative values. Especially with sharp Reformat filters.
- Noise and grain. Denoise before upscaling if the source is noisy, then re-grain at the target resolution so the result matches other elements in the comp.
- Alpha and overscan. Confirm that semi-transparent edges and bounding boxes survive. This is where tools with explicit alpha or overscan handling help.
Choosing the Right Approach
- Quick resize, minimal effort: Reformat with a sharper filter, and check for negative-value ringing.
- Standard 2x on a plate or render: the built-in Upscale node, tiled if VRAM is limited.
- 4x, stills versus sequences, or RGBA CG: NNSuperResolution if you have an NVIDIA GPU.
- Fractional scaling, alpha-weighted edges, or tweakable sharpness and noise: REZup.
- One important shot that needs the best possible result: a CopyCat model trained on that footage.
Whichever you choose, test on your hardest frames first, review in motion, and keep a conventional Reformat in the graph as a reference. Upscaling that is invisible in the final comp is what you're after.