01 · for Nuke artists
Nuke layout and shortcuts.
Panels, node names, knobs and shortcuts follow Nuke 17. Opens and saves .nk, groups and gizmos included.
megasupersoftgpu node compositorAlpha 0.1.0-alpha.5
Oink
GPU compositor for VFX. For film and TV. Opens Nuke scripts.
1
2
3
4
- 1 viewer, 32-bit float
- 2 properties, Nuke knobs
- 3 timeline and cache
- 4 node graph
Oink with an NJR plate loaded.
in the box
Feature list.
- 169 node types
- 13 node categories
- 32-bit float, scene-linear
- one GPU shader per node
- nothing clamped before the display
- OpenColorIO 2.5.2
- opens Nuke .nk scripts
- exports back to .nk
- copy and paste with Nuke
- Nuke 17 shortcuts
- import nuke in Python
- 309 nuke Python functions
- gizmos from NUKE_PATH
- OFX plugins in their own process
- 12 AI nodes
- 36 AI models on your GPU
- one-click AI roto
- depth and normals
- relighting
- motion vectors
- upscaling
- frame interpolation
- clean plates
- 9 deep nodes
- multi-part EXR
- multi-layer EXR
- DWAA and DWAB
- ProRes 4444 XQ
- DNxHR in MXF OP1a
- NVENC H.264, HEVC and AV1
- Cycles in the Viewer (Linux)
- USD stages (Linux)
- Alembic, FBX, OBJ, glTF
- Karma (needs Houdini)
- review app included
- EDL, OTIO, XML and AAF conform
- review notes in the Viewer
- command-line renderer
- Deadline 10 plugin
- OpenCue and Flamenco
- built-in network farm
- mixed-OS path mapping
- collect to one folder
- .oink comps are plain JSON
- MCP server for agents
- AI edits need approval
- studio offline mode
- signed updates
- Linux, Windows, macOS
new free alpha licence · two machines · 14 days offline
every node runs on the GPU.
A node compositor for film and TV that runs every node on the GPU. It opens Nuke .nk scripts and uses the panels, node names and shortcuts Nuke artists already know, with AI roto and mattes, Cycles rendering and a review app in the same download.
nodes
169node types in 13 categories, each one a GLSL shader on 32-bit float images.
ai models
36models for mattes, depth, flow, relighting and more. They run locally on your GPU.
python
309functions in the bundled nuke module, so existing studio scripts keep working.
02 · ai roto
roto from a click.
Click a subject and the matte is tracked through the shot. Frames that need a look are flagged.
03 · roto
roto by hand.
Bezier shapes drawn in the Viewer with Nuke's tools and shortcuts, keyed per frame, feathered and motion blurred.
04 · tracking
2D and camera tracking.
Ctrl+Alt+click adds a track, then track forward. Tracks export to CornerPin and Transform; the CameraTracker solves a camera and a point cloud.
05 · clean plates
markers and objects removed.
Tracking crosses are tracked, masked and filled from the wall around them. AICleanPlate fills a removed object from other frames of the shot.
06 · 3d
Cycles and USD in the node graph.
Live in the Viewer or as a CyclesRender node, with GPU denoising. Linux today, macOS and Windows later.
07 · review
review and notes.
A review app in the same download: play the cut, compare versions, draw notes on the frame. Conforms EDL, OTIO, XML and AAF; notes show up in the compositor's Viewer.
08 · engine
32-bit float.
Scene-linear from Read to Write. Only the display is clamped.
09 · farm
farm rendering.
Nuke-style flags, Deadline 10, OpenCue, Flamenco, or the built-in network farm.
10 · agents
ai assistant.
An assistant and an MCP server. Changes are proposed and the artist approves each one.
11 · studio
offline mode.
A policy file turns off everything that reaches the internet.
technical data
specifications
From the current build.
System
- Linux
- x86-64, glibc 2.34 or newer; tested on Rocky Linux 9
- Windows
- Windows 10 or 11, x64. Per-user installer (no admin rights) or a portable .zip
- macOS
- macOS 14 Sonoma or newer on Apple Silicon
- Graphics
- OpenGL 4.6 on Linux; DX12 or Metal on Windows and macOS
- NVIDIA
- Recommended: GPU Cycles (CUDA / OptiX) and the AI tools are built for it
- No GPU
- Renders on the CPU with the same pixels, slower (
--cpu) - Download
- Linux 582 MB, macOS 445 MB, Windows 451 MB, Windows .zip 455 MB
- AI packages
- Downloaded on first use of an AI node, then each model when it is first needed
Image pipeline
- Precision
- 32-bit float RGBA on the GPU, scene-linear; nothing clamped before the display
- Nodes
- 169 node types in 13 categories, 12 of them AI
- Colour
- OpenColorIO 2.5.2: the bundled config or your own through
$OCIO; OCIO transforms run as GPU shaders - Deep
- DeepRead, DeepToImage, DeepMerge, DeepHoldout, DeepTransform, DeepCrop, DeepRecolor, DeepFromImage, DeepReformat
- Engine
- One GLSL shader per node, C++ core for evaluation and per-frame caching
Read
- Images
- EXR, DPX, CIN, TIF, TIFF, PNG, JPG, JPEG, HDR, TGA, SGI, RGB, PSD, BMP
- Movies
- MOV, MP4, MXF, AVI, MKV, M4V, WEBM, GIF, WEBP, AVIF, through LGPL FFmpeg in the app's own process
- EXR
- Multi-part and multi-layer; layers and channels picked per Read
- 3D
- USD, USDA, USDC, USDZ, OBJ, ABC, FBX, PLY, STL, GLB, GLTF
Write
- EXR
- Half or float; none, Zip (1 or 16 lines), RLE, PIZ, PXR24, B44, B44A, DWAA, DWAB; multi-layer, ACES-compliant option, autocrop, metadata; the bounding box becomes the data window
- DPX
- 8, 10, 12 or 16 bit; fill, byte order and transfer options
- TIFF, PNG, JPEG
- TIFF 8 / 16 bit / float, PNG 8 / 16 bit, JPEG with quality and sub-sampling
- QuickTime, MP4, MKV
- ProRes 422 Proxy, LT, 422, HQ, 4444, 4444 XQ; DNxHD; DNxHR; H.264, HEVC and AV1 (software or NVENC); VP9; Motion JPEG; PNG; Animation
- MXF
- OP1a with DNxHR, DNxHD or ProRes
- Web
- WebM (VP9 with alpha, AV1), animated GIF, WebP and AVIF
- Colour out
- Linear to the output colour space on the GPU; movies default to Rec.709 / Gamma 2.4 like Nuke
3D and rendering
- Cycles
- Built in on Linux: live in the Viewer or as a CyclesRender node, OptiX on NVIDIA, GPU denoising
- USD
- Stages render through Cycles' USD (Hydra) delegate on Linux
- Karma
- Through your own Houdini install (husk)
- Blender
- Not needed. Blender 5.2 LTS, if you add it, brings EEVEE and
.blendfiles
AI
- Models
- 36 models for mattes, tracking, depth, normals, flow, upscaling, frame interpolation, clean plates, relighting and generation
- Where
- On your own GPU. Cloud AI providers are off unless an administrator allows them
- Memory
- From 0.5 GB to 22 GB of graphics memory, listed per model below
- Licences
- Each model's licence is shown and accepted per user. Non-commercial research models sit in an opt-in pack an admin can turn off
Scripting and Nuke
- Python
- Python 3.13, bundled;
import nukewith 309 functions - .nk
- Open, import, export, copy and paste: groups, gizmos, user knobs, expressions and Tcl
- Gizmos
.gizmofiles in~/.nuke,~/.opencomp/gizmosor onNUKE_PATHshow in the Tab menu- OFX
- Plugins run in their own process, so a crash stays in the plugin
- Agents
- An MCP server and an in-app AI Assistant; changes are proposed and approved by the artist
Pipeline and farm
- Command line
- Nuke-style flags (
-X,-F,-m,--cont) and exit codes; JSON progress for farm tools - Farm managers
- Deadline 10 plugin, OpenCue, Flamenco, Open Job Description (AWS Deadline Cloud progress lines)
- Network farm
- Built in: any machine with the app can be a LAN render worker
- Path mapping
- Open Job Description path-mapping rules for mixed Linux / Windows / macOS farms
- Collect
- Copy a comp and every file it uses into one folder or zip
Review
- Conform
- EDL, OTIO, XML and AAF onto a multi-track timeline
- Notes
- Compare versions, draw notes, and see them in the compositor's Viewer
- Prep Show
- Analyses every shot of a show in one pass (key shots, ranges) for planning
Studio controls
- Policy file
/etc/opencomp/policy.json(Linux),/Library/Application Support/opencomp/(macOS),%ProgramData%\opencomp\(Windows)- Offline
"offline": trueturns off every internet request; the app and its AI worker refuse connections at the socket level too- Usage reports
- On in alpha and beta (crash reports and anonymous counts, no footage, paths or names), off in stable releases; the policy file can force them off
- Licence
- Free alpha licence, two machines, 14 days offline between check-ins
- Updates
- Signed feed (Ed25519). Channels stable, beta and alpha; off, notify (default) or install automatically
- Bug reports
- Built in the app; logs and a redacted diagnostic bundle you can read before sending
Files
- Comps
.oink: plain JSON text, easy to diff and merge; older.ocompfiles still open- Logs
~/.cache/opencomp/logs(Linux),~/Library/Logs/opencomp(macOS),%LOCALAPPDATA%\opencomp\logs(Windows)- Libraries
- OpenImageIO 3.1.17, OpenColorIO 2.5.2, FFmpeg 8.1 (LGPL), Electron 28, numpy 2.5.3
all 36 AI models, with licence and graphics memory
| Model | Job | GPU memory | Licence | Pack |
|---|---|---|---|---|
| MapAnything | camera solve | 12 GB | Apache-2.0 | standard |
| Wan 2.1 VACE 1.3B | clean plate | 13 GB | Apache-2.0 | standard |
| Video Depth Anything Small | depth | – | Apache-2.0 | standard |
| MoGe-2 ViT-B | depth and normals | – | MIT | standard |
| MoGe-2 ViT-L | depth and normals | – | MIT | optional |
| Grounding DINO Base | find objects by text | – | Apache-2.0 | standard |
| FLUX.2 [dev] 32B fp8mixed | generation | 22 GB | FLUX [dev] Non-Commercial License | research pack |
| FLUX.2 [klein] 4B | generation | 10 GB | Apache-2.0 | standard |
| LTX-2.3 22B distilled | generation | 22 GB | LTX-2 Community License | optional |
| Qwen-Image-Edit 20B | generation | 8 GB | Apache-2.0 | standard |
| Wan 2.2 I2V A14B | generation | 20 GB | Apache-2.0 | standard |
| Wan 2.2 TI2V 5B | generation | 20 GB | Apache-2.0 | standard |
| Z-Image Turbo | generation | 15 GB | Apache-2.0 | standard |
| Cosmos DiffusionRenderer 7B | material passes | 18.7 GB | NVIDIA-Open-Model-License | optional |
| BiRefNet HR matting | matte | – | MIT | standard |
| Robust Video Matting | matte | – | GPL-3.0 | standard |
| ViTMatte base | matte | – | Apache-2.0 | standard |
| ViTMatte small | matte | – | Apache-2.0 | standard |
| SAM 3 | matte from text | 8 GB | SAM-License | optional |
| SAM 3.1 | matte from text | 8 GB | SAM-License | optional |
| SEA-RAFT M | motion vectors | – | BSD-3-Clause | standard |
| SEA-RAFT S | motion vectors | – | BSD-3-Clause | standard |
| MiniMax-Remover | object removal | 16 GB | CC-BY-NC-4.0 (weights); code: no licence (non-commercial) | research pack |
| TAPNext++ | point tracking | 4 GB | Apache-2.0 | standard |
| Cosmos DiffusionRenderer 7B | relighting | 18.7 GB | NVIDIA-Open-Model-License | optional |
| RIFE 4.26 | retime | – | MIT | standard |
| DINOv2 ViT-S/14 | shot similarity | 0.5 GB | Apache-2.0 | standard |
| EdgeTAM | tracked matte | – | Apache-2.0 | standard |
| EfficientTAM-S | tracked matte | 2 GB | Apache-2.0 | standard |
| EfficientTAM-TI | tracked matte | 2 GB | Apache-2.0 | standard |
| SAM 2.1 Hiera Large | tracked matte | – | Apache-2.0 | standard |
| Real-ESRGAN x2plus | upscale | – | BSD-3-Clause | standard |
| Real-ESRGAN x4plus | upscale | – | BSD-3-Clause | standard |
| SeedVR2 3B | upscale | – | Apache-2.0 | standard |
| Shot-trained upscaler | upscale | 2 GB | GPL-3.0 | standard |
| MatAnyone 2 | video matte | 5 GB | S-Lab-License-1.0 (non-commercial) | research pack |
command-line render flags
<script> | a .oink or .nk script (an .nk is imported first) |
-x | render. Accepted for Nuke compatibility; rendering is the default |
-F <range> | frames: 1-100, 1-100x5 (every 5th), 1,5,10. Repeat -F for more ranges. Default: the script's frame range |
-X <Write> | render only this Write node. Repeat it (or use Write1,Write2) for several. Default: every enabled Write, in render order |
-m <n> | number of threads |
--cont | keep going after a frame fails |
-f | full size (the default) |
-p | proxy: not supported yet (Write nodes have no proxy file path); exits with code 2 |
-V [level] | more output: 1 script loading and the checks, 2 everything, including the engine's own messages |
-q | quiet: only the progress lines and errors |
--view <name> | views to render; scripts have one view, main |
-i | accepted and ignored (there is no interactive licence) |
--gpu [N], --gpu-index N | render on GPU number N (also sets CUDA_VISIBLE_DEVICES=N) |
--cpu | render with no GPU, on Mesa's software Vulkan driver (lavapipe). Same pixels, slower |
--require-gpu | fail with exit code 3 if the machine has no working GPU and the engine fell back to a software renderer |
--renderer any|gpu|cpu | the same choice as one flag, for farm templates: gpu = --require-gpu, cpu = --cpu |
--json | also print machine-readable OCBG {...} event lines |
--no-verify | skip reading each written file back |
--openjd | also print Open Job Description progress lines, which AWS Deadline Cloud shows as task progress |
--path-map <file> | remap file paths in the script (Read, Write...) with Open Job Description path-mapping rules (JSON) |
--list-writes | print the script's Write nodes as JSON (name, enabled, order, output path, frame range, dependencies) and exit; for farm submitters |
--localize | read the local copies made by Localization (when they match their source by size and time) instead of the network files. Same as Nuke's --use-local-file-cache |
--worker | run this machine as a LAN render-farm worker (see LAN render farm) |
--collect <script> <dest> | copy the comp and every file it uses into a folder (like File > Collect Files). Options: --all-frames, --proxies, --ai-caches, --no-ocio, --zip |
--prep-show <edl, otio, folder or json> --out <dir> | analyse every shot of a show (see Prep Show). Options: --shots, --chunk i/n, --merge-only |
--diag-bundle [--out F.zip] | write a redacted diagnostic zip (see Logs and diagnostics) |
--mcp-setup [claude-code|claude-desktop] | connect Claude Code or Claude Desktop to the running app's MCP server |
--version | print the version |
available now
download
Version 0.1.0-alpha.5, 2026-10-15.
Needs a free alpha licence to run. Request one before you install.
Linuxx86-64 .tar.xz, glibc 2.34+free alpha licence required582 MB · 0.1.0-alpha.5macOSApple Silicon .dmgfree alpha licence required445 MB · 0.1.0-alpha.5Windowsx64 installer, per userfree alpha licence required451 MB · 0.1.0-alpha.5Windows .zipx64, portable, no installfree alpha licence required455 MB · 0.1.0-alpha.5
alpha programme
alpha testing.
The alpha is free. Each tester gets a licence for two machines, and it keeps working offline for 14 days between check-ins.
licence
Fill in the form; the licence comes to your email address.
updates
The app checks for new builds and offers to install them. Help > Check for Updates... checks right away.
bug reports
Help > Report a Bug... writes an email to [email protected] with a diagnostic bundle you can read before sending.
Alpha and beta builds send crash reports and anonymous usage counts so we can see what breaks; never footage, file paths or names. Turn it off in the app. Studios can force it off with the policy file. Stable releases have it off by default.
Is it free?
It is free to download during the alpha. You need a free alpha licence: request one here.
Is it ready for production?
Not yet. This is an alpha: expect bugs, missing pieces and the odd crash. Test it on real shots next to your current pipeline, and keep delivering with what you trust.
When is the beta?
There is no date to announce. The beta comes when clean installs work everywhere, the Nuke shortcut list is complete and the app has no stalls in everyday work.
Is it open source?
No. Oink is proprietary software by MegaSuperSoft. It includes open-source components (FFmpeg, OpenImageIO, OpenColorIO, Cycles, Electron and others); their licences and source are listed with every download.
Can I open my Nuke scripts?
Yes. .nk scripts open: nodes, knobs, expressions, groups and links are translated, and anything that cannot be is listed. You can also export back to .nk and copy and paste between the two.
How is .oink different from .nk?
.oink is Oink's own comp file: plain JSON text, so it diffs and merges cleanly in version control. .nk stays a first-class import and export format.
Do gizmos and OFX plugins work?
Gizmos load from ~/.nuke, ~/.opencomp/gizmos or NUKE_PATH and show in the Tab menu. OFX plugins run in a separate process, so a crashing plugin does not take your session with it.
Does it do deep compositing?
Yes, with 9 deep nodes: DeepRead, DeepToImage, DeepMerge, DeepHoldout, DeepTransform, DeepCrop, DeepRecolor, DeepFromImage, DeepReformat.
Which Python?
Python 3.13, bundled with the app. import nuke gives you 309 functions of Nuke's Python API, enough for most pipeline tools and menus.
Which graphics card do I need?
Any GPU with OpenGL 4.6 on Linux, or DX12 / Metal on Windows and macOS, runs the compositor. GPU Cycles and the AI tools are built for NVIDIA cards.
Will it run on a laptop without NVIDIA?
The compositor will. The AI tools and Cycles fall back to the CPU, which is much slower, and the largest AI models need more graphics memory than most laptops have.
Does it run on Intel Macs?
No. The macOS build is for Apple Silicon and needs macOS 14 or newer.
Which Linux distributions?
x86-64 with glibc 2.34 or newer. It is tested on Rocky Linux 9; other current distributions should work but are not tested yet.
Is Cycles available on Mac and Windows?
Not yet. Cycles and USD rendering ship in the Linux build today; macOS and Windows get them later.
Do I need Blender?
No. Blender 5.2 LTS is optional and only adds EEVEE renders and opening .blend files.
Can I use our studio's OCIO config?
Yes. Point $OCIO at your config, as you would for Nuke; without it, Oink uses its bundled config.
What leaves my machine?
Alpha and beta builds send crash reports and anonymous usage counts: never footage, file paths, comp or shot names. You can turn this off in the app; stable releases have it off. Otherwise the network is used only for things you start: licence check-ins, updates, model downloads and bug reports.
Can a studio run it offline?
Yes. A policy file with "offline": true blocks every internet request, usage reports included. The alpha licence needs a check-in at least every 14 days.
How many machines can I use?
An alpha licence covers two machines, for example a workstation and a laptop.
How do updates work?
The app checks a signed release list and asks before installing (studios can switch this off or make it automatic). Pick the stable, beta or alpha channel; Help > Check for Updates... checks right away.
How do I report a bug?
Help > Report a Bug... writes an email to [email protected] with a diagnostic bundle you can read first. File paths, names and tokens are removed from the bundle before you see it.
Where are the logs?
~/.cache/opencomp/logs on Linux, ~/Library/Logs/opencomp on macOS, %LOCALAPPDATA%\opencomp\logs on Windows. Help > Open Log Folder takes you there.
How do I uninstall it?
Windows: Settings > Apps or the Uninstall entry in the Start menu. macOS: drag the app to the Trash. Linux: run ./uninstall.sh in the installed folder. Your settings and caches are kept unless you delete them.
Why does Windows or macOS warn about it?
Alpha builds are not code-signed yet, so the system warns once. The first-launch steps in Downloads get past it.
Free during the alpha.