Early Access · macOS, Windows and Linux
Plates, CG and machine learning in one node graph. Triton is designed from the ground up for shots that mix photography, 3D and models: a real colour pipeline, local ML nodes, and cloud GPUs when you need them.
Hybrid VFX means a shot moves between photography, 3D and models all day. In Triton they are the same kind of node, in the same graph, on the same frames, in the same colour space.
Not a model runner
Browser-based node tools are great for trying models, but they stop at the image. There is no timeline, no scene-linear colour, no roto, tracking or 3D. Triton is a native app where a model is one node in a full shot pipeline.
Not a bolt-on
Compositing apps built decades ago add machine learning as isolated, single-purpose tools. In Triton, depth, normals, mattes and flow are ordinary channels that feed every other node.
Ground up
Graph, cache, colour management, 3D and the job queue were built together for work that moves between photography, CG and models on every frame.
A typical hybrid shot in Triton never leaves the graph. Photography, models and CG pass the same frames, in the same colour space, through the same cache.
The finishing tools a hybrid shot needs ship with the app, so the work ends where it started.
3D
An interactive 3D viewport lives in the same project as the 2D work, so plates, cameras, CG and splats share one graph.
Look
Matching is most of hybrid work. FilmSimulation runs a spectral model of 28 still and motion stocks from published datasheets, and grain, grading and light tools are built in.
Shot work
Roto, tracking, keying, warping and merging are native nodes with GPU previews while you drag, so a hybrid shot finishes where it started.
Your graph, project files, viewers and local ML nodes live on the artist machine. Heavy generation and reconstruction leave the workstation and run on dedicated GPUs: hosted by default, or in your private cloud when you need it.
The desktop app owns the graph: media I/O, merges and grades, local ML passes, 3D scene ops, animation, viewers and caches. Interactive work stays on your machine with your media and project structure.
Cloud GPU nodes send only their inputs to hosted workers. Stage inputs, run the job, stream progress, and pull results back into the same graph, paid from prepaid credits by the GPU-second. No local multi-GPU farm required.
Studios that need inference inside their own network can run a private deployment. Available on request — contact us to scope access, security, and capacity.
Directed acyclic graphs with typed ports, tab-search node creation, groups, undo, media drop, and a pull engine tuned for interactive playback. Plate, 3D, local ML and cloud GPU nodes sit side by side in the same tree.
I/O
Read EXR sequences & video · Write EXR · Write video · Multi-layer EXR
Image
Merge · Grade · OCIO · Blur · Keyers · Transform · Reformat · Retime
3D
Camera · ReadGeo · Lights · Materials · Project3D · ScanlineRender
Local ML
Depth · Normals · SmartRoto · Matting · Flow & retime · Inpaint · Upscale
Cloud GPU
Chord · Trellis 2 · SHARP · Hunyuan3D · more via API
Image, float, string, scene, and multi-layer EXR flow through explicit connections — fewer silent type bugs mid-shot.
One dock stacks an OpenGL OCIO 2D viewport and a 3D scene view — channel isolation, exposure, gamma, look-through camera, and on-viewer handles.
Keyframe curve editor and dope sheet for float parameters that drive the graph over time, with autokey-friendly property edits.
Tab create, groups, bookmarks, viewer pipes, rubber-band selection, clipboard, and drop media onto the canvas to spawn Reads.
Each hosted model is a cloud GPU job: open it in a web workspace or fire it from a desktop cloud node. Inputs stage on GPU volumes; results come back as artifacts in your local graph.
Estimate PBR material maps from a single image. Optional relit outputs for look-dev and MaterialX-friendly pipelines.
Textured mesh reconstruction from a single image — resolution, decimation, and texture size knobs for production exploration.
Generate 3D Gaussian splat reconstructions for dense, photoreal point-based views.
Image-to-mesh generation with seed, octree resolution, steps, and guidance controls.
Video camera and geometry intelligence for matchmove-adjacent workflows via the cloud job API.
Dense point tracking across video for stabilization, paint prep, and motion-aware effects.
Desktop and web share one trusted surface for cloud GPU nodes: authenticate, stage media on GPU volumes, enqueue jobs, stream progress and download artifacts, while your project and graph stay local.
In Preferences → Account, hit Sign in with browser. You approve the desktop in the browser; Triton then picks up a device Bearer from the live API. Only the token hash is stored server-side. That links billing and enables cloud AI nodes.
Short-lived upload handles stream media straight to GPU volumes — not through your app server as base64.
Queue on Postgres (pg-boss), poll status, or subscribe to server-sent progress for long GPU runs from desktop or web.
Shared Zod schemas drive validation and docs. Desktop Pydantic models stay aligned with the same request and response shapes.
Native app for macOS, Windows and Linux on the VFX Reference Platform CY2026 stack. Local graphs, ML nodes, viewers and media I/O work with no account. Sign in from the app when you want cloud GPU nodes, paid with prepaid credits.
Free to install and use offline. Sign in with browser links your account; cloud GPU nodes run on prepaid credits from $10.
This server is not configured to read GitHub Releases. Set GITHUB_RELEASES_TOKEN (a PAT with Contents: Read on the private repo).
Need a build? Contact us
Free for non-commercial work. Indie is one payment, not a subscription: a year of updates and you keep that version forever.
$0non-commercial
The whole app, for learning and personal projects.
$199$99one time
Founder price during Early Access. Goes to $199 at 1.0.
$399per seat, per year
For companies over $100k revenue. Opens after 1.0.
Short answers about hybrid VFX, where compute runs, and how models get into the graph.
Shots that mix photographed plates, CG and machine learning: a generated set extension relit to match the plate, a depth pass from a model driving fog, an ML matte refined with roto. Triton puts all of that in one node graph, on the same frames and in the same colour space.
Triton is not a compositor with AI features added later. It is a workspace built from the ground up for hybrid VFX: roto, tracking, keying and merging are in the box, alongside 3D, local ML nodes and cloud GPU nodes, all sharing one graph, one colour pipeline, one cache and one job queue.
Local ML nodes run on your machine and never upload frames. Cloud GPU nodes upload only the inputs you connect to them, and only when you run them.
Yes. Private cloud deployment is available on request. Contact us to scope security, capacity and access.
Sign in with browser from Preferences → Account. Triton opens the site, you approve the desktop, and the app receives a device token. The same typed HTTP API then stages media, enqueues jobs, streams progress (SSE) and downloads artifacts. OpenAPI: /api/v1/openapi.json.
macOS, Windows and Linux on the VFX Reference Platform CY2026 stack.
Early Access builds go to the waitlist first, with the $99 Founder license until 1.0.