ANR AI Upscaler - Matcher

Overview

The AI Upscaler - Matcher brings a small image back to full size: a frame repainted in another program or by an AI image tool, a low-resolution render, an old element. It upscales the Low rez input with the AI model you choose. In Match + upscale mode it then lines the result up with the Original plate and matches its colour, so the edit drops back onto the plate.

AI model offers Real-ESRGAN x2plus and x4plus, and SeedVR2, a heavier generative model. AI output keeps the whole model result, or adds only the fine detail the model made on top of a plain resize, keeping the original colours.

Alignment corrects a small shift and size difference between the upscaled image and the Original. Colour match fits a gain and offset per channel, or a full colour matrix, on the pixels the two images share: what you changed in the edit stands out and is left out of the fit. Exclude keeps areas out of both.

Nothing runs while you scrub. Process works through the frames you choose, on this computer, and saves the results as image files in the Cache folder; the node then plays them back without running the model again.

Inputs

1. Low rezrequired

   The image to upscale: the low-resolution edit, render or element.

2. Originaloptional

   The full-resolution plate the Low rez was made from. Needed in Match + upscale mode: the result is aligned to it and matched to its colour, and with Output size 2x or 4x its shape sets the height.

3. Excludeoptional

   A matte of areas to leave out of the alignment and the colour match, such as a large part you changed. It must have the Original's format.

Output. The upscaled (and matched) image at the size set by Output size, played back from the cache. Before anything is processed, the node shows the Low rez; frames outside the processed range are black, and a single processed frame shows on every frame. View and write from the ANR_Output Reformat under the node: it carries the result's format.

Quick start

1. Add ANR AI Upscaler - Matcher. The menu also adds a Reformat named ANR_Output under it: view and write from that one.
2. Connect the low-resolution image to Low rez. To match it to the plate, connect the plate to Original and leave Mode on Match + upscale; without a plate, choose Upscale only.
3. Choose a Cache folder, or press Generate Autofill to make a new one next to the saved script.
4. Choose the AI model and the Output size: the project format by default, or 2x, 4x or a custom format.
5. Set Process to Current frame to try one frame, or to Frame range with First frame and Last frame.
6. Turn proxy mode off, press the Process button, then view ANR_Output: processing starts when the viewer asks for the frame. A progress window follows each stage and can cancel.
7. Check the result. Keep the node, which plays the cached files back, or press Import result for a Read of them.

The panel

The node's panel as Nuke shows it, with the values of a new node. Click any control to see what it does; tabs and sections work as in Nuke.

All controls

Tips

• Try one frame with Current frame before a long range: every frame goes through the model.
• If the model shifts colours or tones, choose Detail only (preserve colour) or lower AI strength.
• With AI strength at 0 the node skips the model: it resizes, aligns and matches only, and needs no GPU.
• If the colour match fails or is pulled by a large change, connect a matte of the changed area to Exclude.
• For a plate with values above 1, choose EXR (float / HDR) in Output files: PNG clips them.
• Generate Autofill makes a new folder on every click, so earlier results are kept. To put cache folders next to the footage instead of the script, use ANR > Cache Folders > Location in Nuke's menu bar.
• A Read made by Import result stays as it is when you process again: each run writes new files with their own name.

Good to know

• The AI models need an NVIDIA GPU. SeedVR2 needs one with 8 GB of memory or more, and 8 GB of free system memory, and takes longer than the Real-ESRGAN models.
• The models are a separate download. If they are not on this computer, the node offers them when you add it or choose SeedVR2, and Nuke stays usable while they download; then press Process again.
• A Cache folder is required: nothing is processed without one. It holds only the finished image files, and a failed or cancelled run keeps the previous result.
• Processing runs on this computer: nothing is uploaded. Changing a setting does not run the model again; press Process.
• Alignment corrects small differences only. If the Low rez is far off the Original in position, size or rotation, line it up roughly with a Transform first, or set Alignment to Off.
• Each frame is processed on its own, with no link between frames: check a sequence for flicker in the generated detail.
• AI detail is added only where the image is between 0 and 1; brighter or negative pixels keep a plain resize.
• Limits: 16 to 16384 pixels per side and at most 64 megapixels, up to 10000 frames per run, and Low rez and Original need the same pixel aspect.

Works with

ANR GPT ConnectionChat with GPT inside Nuke: edit or add to a frame with image generation, or have it build and animate nodes in your script.Photoshop LinkSends the current frame to Photoshop and brings it back as a Read that updates every time you save in Photoshop.