ANR GPT Connection

Overview

ANR GPT Connection brings GPT into a node, signed in with your ChatGPT account through the official Codex app from OpenAI: no API key. The node has three tabs. Login connects your account. Image Gen sends a frame and your request and loads the generated image into the node. Control Nuke shows GPT the whole shot and lets it build nodes for you.

In Image Gen you chat as you would with ChatGPT: describe an edit, an object to add or something to remove, and reply to refine it. A matte on Mask marks where the change goes, Crop before sending sends only a region of the frame, and Reference1 / Reference2 add object or style images. The returned image is saved as a PNG in your Save folder, colour-matched to the plate if you like, and shown in place of the Source.

In Control Nuke, every frame of the chosen range is prepared for the GPT model, which can look at frames closely, create and animate nodes (grades, transforms, blurs, keys, merges, Expression, STMap and IDistort effects, editable Roto shapes) and preview its own result before it finishes. Every node it makes is an ordinary Nuke node, and Undo reverts it.

Inputs

1. Sourceoptional

   The footage to work on: the frame Image Gen sends and the result is placed over, and the shot Control Nuke prepares. Image Gen can also generate without it (untick Use source frame as reference).

2. Maskoptional

   An alpha marking, in white, where to edit, add or remove content. Used while Use alpha as placement guidance is on. It guides GPT only: it never cuts the returned image.

3. Reference1optional

   An optional picture of an object or a style for GPT to use, sent with each request (both tabs).

4. Reference2optional

   A second optional object or style picture, like Reference1.

Output. With Show on Result, the returned image, converted to the project's working colour space and fitted to the Source's frame; a cropped patch is placed back in its box over the untouched Source. With Show on Source, the input untouched. Control Nuke builds its own nodes in the graph instead.

Quick start

1. Install the official Codex app (or the Codex command-line tool) from OpenAI, then add ANR GPT Connection.
2. In the Login tab press Log in with ChatGPT and finish signing in in your browser. The tab then says Connected to ChatGPT.
3. Connect your footage to Source. To mark where the change goes, connect a matte to Mask and tick Use alpha as placement guidance; Preview mask shows what GPT receives.
4. In Image Gen, choose a Save folder for the generated images.
5. To edit only part of the frame, press Crop before sending and draw the box in the viewer.
6. Type your request in the message box at the bottom and press Send (or Ctrl+Enter). Reply in the same box to refine the result.
7. The image is saved as a PNG and shown in the node. Show switches between Result and Source; Import image to graph makes a Read of it.
8. To have GPT build nodes, open Control Nuke, set First and Last (0 and 0 for the whole shot), describe what you want and press Send.

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

• With Continue editing latest generated image ticked, follow-ups edit the last result; untick it to start from the Source again.
• New chat starts a fresh conversation in that tab; images already made and nodes already built stay.
• Untick Color match output to input when you ask GPT to grade or recolour the scene; leave it on to keep the plate's exposure and colour balance. The original PNG is always kept.
• For a cropped edit, Refine patch alignment lines the patch up with the unchanged background before colour matching. Turn it off for an intentional reframe; Refine latest patch runs it on the last patch without a new request.
• Use the same crop box for repeated edits of one patch: changing the box starts again from the original Source outside it.
• Double-click the Source or Latest generated image preview to see it larger; Refresh source preview captures the current frame.
• In Control Nuke, ask GPT to inspect the start, middle and end of a moving shot, and to preview masks and effects before it finishes.
• Roto shapes and keys that GPT draws are placed by eye, not tracked: check them over the shot.
• Stop cancels the request that is running.

Good to know

• Needs the official Codex app or command-line tool installed, a ChatGPT account and an internet connection. Your sign-in stays with the Codex app; the node never asks for a password or an API key.
• Your requests, the frames and the reference images are sent to OpenAI, and each request uses your Codex allowance.
• The conversations are saved with the Nuke script. Never type a password or a key into them.
• Image Gen needs a Save folder before it sends anything.
• Generated images are PNGs, with PNG's display range: inside a crop the result has no HDR values, while outside it the Source keeps its original pixels.
• A placement mask with nothing white in it (inside the crop, when cropping) stops the request.
• The crop box and the mask tint draw in the viewer only while the node's panel is open and the viewer overlays are on.
• Control Nuke takes 1 to 1000 frames. Camera reconstruction requests need a separate camera backend that ANR Tools does not include, and give an AI estimate of the camera at an arbitrary scale, not a measured matchmove.

Works with

ANR AI Upscaler - MatcherBrings a low-resolution image up to plate size with an AI upscaler, then lines it up with the original plate and matches its colour.Photoshop LinkSends the current frame to Photoshop and brings it back as a Read that updates every time you save in Photoshop.