Bringing Generative AI Into Nuke - Natively

Nuke compositors have never had more AI tools available to them — but almost none of them live where compositors actually work.

Today's generative AI workflow for a Nuke artist typically looks like this: render a frame, export it, switch to a browser-based tool or standalone app, upload the image, run the AI process, download the result, import it back into Nuke through a Read node, and repeat. Every round-trip breaks creative momentum. Every context switch adds friction that compounds across hundreds of iterations on a shot - and worse, the AI output usually arrives without alpha, without metadata, and without any way to chain into the next step of the comp.

Studios face an additional challenge: most AI models are trained on data of uncertain origin, creating legal exposure for commercial projects. When a tool lives outside the pipeline, there's no audit trail, no metadata, and no content moderation layer between the AI output and the final deliverable.

Bria AI Tools for Nuke is a set of 11 AI-powered nodes that appear in Nuke's standard menu bar (Bria → Create/...) and the Tab search ("bria"). All are built as Group nodes - so they save cleanly into .nk files, behave like every other Nuke Group, and chain together at full resolution inside the comp graph.

Install is one line. Drop this into ~/.nuke/init.py:

nuke.pluginAddPath("/path/to/bria-nuke/nuke")

Restart Nuke, open Bria → Dashboard, paste your API key, and start creating.

All processing is non-destructive. Results are saved to disk and loaded through internal Read nodes, keeping your original plates untouched.

And because Bria's models are trained exclusively on licensed data, studios can use these tools in commercial production without legal uncertainty.

Generate

Fibo Generate creates images from text prompts with control over aspect ratio, guidance, and seed. New in v0.2.0: toggle Generate Multiple Images and Fibo Generate fires up to four parallel async calls per click, each with a per-variant seed offset for real variety. Drag the Image Index slider to preview each variant; click Use This Image to lock in the one you want as the canonical result. Sidecar metadata records every variant.

For more precise results, Generate VGL produces structured Visual Generation Language (VGL) prompts that break down scene composition, lighting, style, and objects into discrete fields. Connect a Generate VGL node into Fibo Edit, Fibo Generate, or Fibo Edit Recipes for granular control that simple text prompts can't offer.

Edit - Prompt-Driven, Plus 100+ Recipes Built for DCC Artists

Fibo Edit applies prompt-based edits to existing images with adjustable guidance, seed, and an optional mask input.

On top of that, Fibo Edit Recipes is the DCC-unique standout — a categorized launcher of 100+ curated preset prompts built specifically for Houdini and Nuke artists who don't want to engineer prompts from scratch. 11 categories: Style, Weather, Seasons, Time of Day, Camera, Lighting, Compositing, Cleanup, AI Corrections, Object Edits, and Custom. Pick a category, pick a preset, click Edit.

Object-targeted presets (delete, replace, isolate, change color, change material) reveal 5 fixed Object slots — fill in the names of objects in your image and the prompt is rebuilt with constraint-injection ("STRICT LOCKS") wording so the model only edits what you specify. Compositing-targeted presets work the same way for named composited elements, with everything from harmonize lighting and match shadows to integrate elements and color harmonize.

Remove Backgrounds, Erase, Fill, Expand, Enhance, Upscale

RMBG removes the background from a plate with smart edge detection — the same model that powers Bria's standalone product, now one Tab-search away from your Read node. Erase removes content under a mask. GenFill paints generated content into a masked region with a prompt. Expand outpaints the canvas in any direction. Enhance improves image quality at 1MP, 2MP, or 4MP target resolutions with an optional region mask via internal Keymix. Upscale increases resolution by 2x or 4x.

All of these chain together. The output of one Bria node feeds directly into the next at full resolution, enabling multi-step workflows entirely within Nuke.

Targeted Masks and Spill (within Edit Recipes)

For cases where RMBG's foreground/background split isn't the right tool — when you need a specific named object isolated, an atmospheric element matted out, or just the green-screen spill cleaned off a comped subject — Edit Recipes ships three targeted compositing presets:

• Create Mask outputs a clean binary alpha matte over a named element ("the cat", "the lamp"), ready as a luminance key. Useful when several subjects are present and only one needs isolating.
• Create Soft Mask outputs a grayscale density matte for atmospherics — smoke, fire, dust, fog, mist — where solid objects in the same frame remain pure black and density follows brightness.
• Remove Greenscreen Spill targets named element(s) and removes green/blue-screen color contamination, fringing, and edge spill while preserving naturally-occurring greens (foliage, signage, wardrobe).

When either mask preset is active, a Premultiply with Mask toggle and Mask Edge Softness slider appear automatically. Premultiply ON routes the API-returned matte through an internal Blur → FilterErode → Copy(rgb=input, a=mask.red) → Premult chain, producing a clean cutout element with anti-aliased edges. Switch presets and the toggle clears automatically — no surprise cutouts on later runs.

Batch Across Frame Ranges

The Sequence Output node brings Bria into shot-based workflows. Connect it to the last node of your Bria chain, set a frame range, and click Render. Sequence Output walks upstream from itself, identifies every Bria node by knob presence, topologically sorts them with cycle detection, then per frame: cooks each Bria node in order and renders the final composite via a temporary Write node. Pre-flight validation catches missing prompts and missing mask inputs before frame zero, so one bad frame doesn't derail the rest of the batch. A Lock VGL toggle on Generate VGL nodes lets artists keep a structured prompt constant across all frames, or regenerate per frame.

This isn't a proof of concept. The integration ships as a standard nuke.pluginAddPath plugin, so it drops into existing studio environments alongside any other Nuke tool set. Group-first architecture means every node saves cleanly into .nk files; gizmo export is available as a secondary distribution path.

Every node includes granular content moderation toggles for input, output, and prompt filtering. Proxy support (http_proxy, https_proxy) works out of the box for studios behind corporate firewalls. Metadata sidecars create an audit trail for every AI-generated asset - and v0.2.0 adds View Generation Data and Copy Generation Data buttons to the Settings tab of every Bria node, so prompt audits and result reproduction don't require digging through the file browser.

Cross-platform support covers macOS, Windows, and Linux - TLS certificate auto-detection handles Debian/Ubuntu, RHEL/CentOS/Fedora, openSUSE, and Alpine; clipboard and file-open helpers cover Wayland, X11, and Windows.

The shared ~/.bria/bria.json config file works simultaneously with the companion Bria Houdini integration — set up once, use across both DCCs.

1. Download the latest release from github.com/Bria-AI/Nuke-tools and unzip anywhere.
2. Add one line to ~/.nuke/init.py: ``python nuke.pluginAddPath("/path/to/bria-nuke/nuke") ``
3. Get a Bria API key at platform.bria.ai.
4. Restart Nuke, open Bria → Dashboard, paste your key, and Tab-search "bria" to create your first node.

This release covers Nuke alongside the existing Houdini integration, both built on the same DCC-agnostic API client and shared bria_core package. We're working on additional compositing presets and cross-DCC workflow improvements.

We'd love to hear how you use these tools. Share your results, report issues, or contribute at github.com/Bria-AI/Nuke-tools.