Sonder Editor

Sonder Editor is the node people mean when they talk about "the ComfyUI timeline editor." Instead of babysitting one KSampler for a single clip, you stage the whole shot - clips, audio, guide frames, prompt sections - on a multi-lane timeline, select the range you want to generate, and hand that range to whatever workflow you've wired downstream. It's the same shape of idea as LTX Director, except it doesn't care which model you feed it; the author's own framing is blunt about the division of labor: Sonder doesn't generate anything, it stages the shot.

That's the part to internalize before anything else. This node is not a video model and it is not a sampler. It is a staging surface plus a project/asset manager wrapped in a single graph node, and the generation quality you get out is 100% your workflow. The bundled LTX 2.3 example workflow is the showcase because LTX is fast and complete, but a user in the launch thread has already run it with IC-LoRA motion transfer and a face LoRA on top of a normal LTX graph - Sonder just passes the frame data through.

How it works

Open the node and a fullscreen editor loads: import media into a project-scoped asset gallery, drag clips onto lanes, trim and split, drop diamond-shaped guide frames onto their own lane, cut prompt sections, set an in/out selection. Every project lives as local files under ComfyUI/output/sonder-projects/. When you queue a render, the editor snapshots the selected window and the node emits everything your downstream graph needs: composited frames, guide images with their frame indices and strengths, the prompt text, frame count, resolution, audio, and the mask start/end times that tell a temporal inpainting node where generation is allowed to touch.

The context-frame inputs (pre_context_frames, post_context_frames) are the secret sauce for continuity: frames before and after your selection are included in the render but not denoised, so a regen blends into what's already on the timeline. The mask_pre_offset/mask_post_offset pair then trims that context out of the denoise window. This is exactly the mechanism that makes "regenerate just this beat" possible without the cut jumping.

Inputs and outputs that matter

You only set four widgets when creating a project: project (pick an existing one or "+ Create New"), project_name, fps, and width/height. Everything else is editor UI state. The render_queue_active toggle is worth knowing - off means "render the live selection without consuming queued jobs."

The outputs you'll actually wire, in rough order of importance:

• project - the object that every other Sonder bridge and save node consumes.
• rendered_frames - the composited video (all visible clips layered with opacity).
• guide_images + guide_idx + guide_strengths - reference frames for LTX-style guiders.
• prompt, frame_count, fps, width, height, audio - the current selection's generation context.
• mask_start_time / mask_end_time - wire into a temporal mask node to keep context frames ungenerated.

Installing it

ComfyUI Manager, search Sonder Editor, install, restart. Or:

cd ComfyUI/custom_nodes
git clone https://github.com/SonderSaid/ComfyUI-Sonder-Editor.git
cd ComfyUI-Sonder-Editor
pip install -r requirements.txt

Its requirements are light (numpy, Pillow, opencv-python-headless, imageio, imageio-ffmpeg, scipy, aiohttp). Notably, torch is deliberately not in there - the pack won't clobber your CUDA build, and if you need torchaudio for audio you install a matching build yourself. A system ffmpeg on your PATH is recommended for the widest format support. Restart ComfyUI and the pack registers itself plus its web UI.

Common issues

• ffmpeg not found on export/decode - install ffmpeg system-wide and restart. The bundled imageio-ffmpeg binary is only a fallback.
• cv2 import errors - this pack uses opencv-python-headless; if another custom node pulled in full opencv-python, the two conflict. Keep exactly one OpenCV variant installed (headless).
• GPU "stopped working" after install - something reinstalled torch with a mismatched wheel. Don't let an auto-resolve do that; reinstall the torch build matching your CUDA.
• Missing nodes on an old ComfyUI - Sonder Switch, Cluster, and Metadata Collector Nodes 2.0 use the V3 node API and only load on recent builds. The rest of the pack still works.

One more thing: projects can embed your full workflow metadata when saving (see Sonder Save Video), so turn that off before sharing files if your graph is private - and don't expose a ComfyUI instance running this pack to untrusted networks. The editor is real software, early (v0.1.x), but the author is actively fixing what people hit. It's the node to reach for when you're tired of stitching one clip at a time.