ComfyUI-Majoor-OmniCam
Majoor OmniCam is a camera layout and animation tool for ComfyUI. It lets you prepare camera motion without running image or video generation while editing. You can position cameras, create keyframes, preview motion, and record a simple reference video for compatible video models.
Nodes (3)
Video models do not agree on how to be told about motion. One wants camera extrinsics, another wants 2D trajectories, a third wants a reference video and a prompt. OmniCam separates the two halves of that problem: you describe the motion once, and a compiler translates it per model.
Extractor ─┐
├─> OMNICAM_MOTION_SCENE ─> Monitor ─> camera embedding
Director ──┘ + playblast trajectory JSON / TRACKS
reference video + prompt
A MotionScene is the canonical model-independent authoring document: cameras, objects, motion layers, cuts and an authoring timeline. Motion layers and cuts use time-based scene semantics; camera and object tracks in v1 still retain authoring-grid information (canvas size, authoring fps, duration) where required, and compile to any model resolution or frame rate. It is what travels between the nodes.
Install
Requires ComfyUI 0.31+ and Python 3.10–3.13. After installing, restart ComfyUI; the nodes appear in the node menu under Majoor › OmniCam.
ComfyUI Manager (recommended)
Open Manager → Custom Nodes Manager, search for Majoor OmniCam, and click Install. This pulls the published Comfy Registry release with its prebuilt frontend bundle.
Manual (git clone)
cd ComfyUI/custom_nodes
git clone https://github.com/MajoorWaldi/ComfyUI-Majoor-OmniCam.git
The built frontend bundle (web/, web-chunks/) is committed, so a plain clone
runs as-is — no npm install or build step. There are no required Python
packages beyond ComfyUI's own. The Extractor's camera-solver backends (DPVO,
pycolmap, OpenCV/SIFT) are all optional and installed separately — see the
User Guide.
The three nodes
All three nodes are marked experimental in ComfyUI. Camera authoring and the playblast are stable in practice; the Monitor profile set and the Director Motion Tracks surface may still change before a stable release.
OmniCam Director
A small shot-layout tool in a live 3D viewport. Animate cameras and scene references, draw motion layers over the frame, cut between cameras, and record a neutral proxy playblast. This is where a MotionScene is authored.
OmniCam Extractor
Recover a relative 6DoF camera track from one continuous reference shot and hand
it on as a solved MotionScene. Connect it to the Director's solved_scene input
to keep editing the recovered move, or take it straight to Monitor.
Solves run outside the prompt queue, so you are not queueing a workflow to see a trajectory. Preview uses native browser video first and falls back to server-decoded frames when a container will not decode in the browser.
OmniCam Monitor
The model compiler. Pick a target profile; Monitor resolves the timeline, compiles the MotionScene into that model's representation, and runs a preflight that reports what will and will not survive the translation.
Preflight is binding, not decorative: for every named model profile, a downstream
node that is missing or whose socket contract has changed blocks the run rather
than producing a payload with nowhere to go. external_reference_video is the
one exception, by design -- see below.
Profiles
| Profile | Semantic | Monitor output | Connect to |
|---|---|---|---|
| external_reference_video | reference_video | reference_video + final_prompt | any destination model's own reference-video input |
| wan_camera_native | camera_embedding | camera_embedding | WanCameraImageToVideo.camera_conditions |
| wan_move_native | screen_tracks | native_tracks | WanMoveTrackToVideo.tracks |
| wan_track_native | screen_tracks | tracks_json | WanTrackToVideo.tracks |
| wanvideo_ati | screen_tracks | tracks_json | WanVideoATITracks.tracks |
| ltx25_motion_track | screen_tracks | tracks_json | LTXVDrawTracks.tracks |
| h3_native | reference_video | reference_frames + final_prompt | MiniMaxH3ReferenceToVideo.ref_videos |
| h3_api | reference_video | reference_video + final_prompt | MinimaxHailuo03ReferenceNode.reference_video |
external_reference_video is the Monitor default and the odd one out: it names
no upstream node, imposes no frame grid or fps conversion, and never blocks on
a missing or unrecognized downstream. Use it for a model OmniCam has no named
profile for. Every other profile is strict on purpose -- it encodes one real
model's contract, and a payload that contract cannot satisfy is a bug worth
stopping the queue for.
Switching profile never changes the MotionScene. It does change which Monitor output carries the result, so connect the output this table lists for the profile you selected.
Start here
- Add OmniCam Director and compose a shot. Press
Iat each pose to key it. - Connect
motion_sceneandplayblast_videoto OmniCam Monitor. - Choose the profile your downstream model needs, and queue.
- Read the preflight, then connect the output named in the table above.
To start from footage instead, put OmniCam Extractor in front and wire its
motion_scene output to the Director's solved_scene input.
Complete runnable graphs are in examples/workflows/:
each is the official Comfy-Org template for that model with its motion source
replaced by OmniCam, so every model, LoRA and sampler setting is upstream's.
What OmniCam will refuse to do
These are preflight results, not bugs:
- A multi-shot edit on a single-camera profile is blocked. One camera embedding, or one projection basis, cannot describe an edit that cuts to a second camera. Reference-video profiles accept it — the playblast carries the cuts — and drop the single-camera prompt in favour of a neutral one.
- Trajectories the JSON track formats cannot carry are reported. A layer hidden on the first sample cannot be expressed and is dropped; one that disappears and returns is cut at the gap. Monitor names the affected layers instead of quietly encoding less than you authored.
- A missing or changed downstream node blocks the run, per profile, so a missing LTX install never blocks a Wan Camera compile.
Documentation
- Node Guide — inputs, outputs, profiles and workflow contracts.
- User Guide — authoring, playblasts, extraction, installation.
- In-app help — contextual help from each node.
- Shortcuts — viewport, timeline and editing controls.
- Technical Reference — runtime behaviour, DPVO, validation, development.
- Security — managed files, upload limits, request boundaries.
License
MIT. See LICENSE.