Onion Skinning (Motion Ghosts)

Onion skinning is the animator's old trick of drawing the previous frames underneath the current one so you can see where a character is going. This node brings it to your rendered video: it takes a batch of frames and composites translucent "ghost" copies of nearby frames onto each one, so motion reads as a visible trail. Think of those sports-photo streaks, or the blue/orange motion trails on cycling videos - that's the look.

In this pack it exists to show off the reconstruction: render a mesh sequence from a new angle and onion-skin it, and the ghost trail makes the motion legible in a way a single frame can't. The pack's own onion_skinning_demo.json workflow does exactly this, and MeshSequenceRenderer even has a baked-in onion-skin toggle. This standalone node is the same effect, with every knob exposed, and you can run it on any image batch - it doesn't care whether the frames came from a 3D render, a video loader, or an animation.

How it works

For each frame, the effect pulls frames from frame_interval steps behind (and, optionally, ahead) - so ghost_count of 3 at frame_interval 5 means ghosts at 5, 10, and 15 frames back. Each ghost gets an opacity, optionally fades with distance, and can be color-treated, then it's composited under the current frame.

The three dials that define the look:

• opacity_falloff - linear fades ghosts out the further back they are, exponential fades them fast, constant keeps them all equally strong. Linear is the default and the most natural.
• ghost_color_mode - original keeps true colors, tinted washes ghosts toward your chosen tint, silhouette fills them with solid color. Tinted is the classic blue-past/orange-future look.
• blend_mode - over (normal alpha compositing), add (bright, glowy, great on dark backgrounds), screen (brightens without blowing out whites).

Defaults: 3 ghosts, every 5 frames, 0.3 opacity, linear falloff, tinted, blue for past and orange for future, include_future off. The defaults are genuinely good for a first pass.

Inputs and output

• images - the only required input: your rendered frames as an [N, H, W, C] batch. Output from MeshRenderer or MeshSequenceRenderer, or any video loader.
• Everything else is optional and defaults to the recipe above: ghost_count (1–10), frame_interval (1–30), ghost_opacity (0.05–0.8), past_tint / future_tint (hex colors), include_future, and the three enum dials.

The output onion_skinned is the same-length image batch, ready for VHS_VideoCombine or SaveImage. Note this node is pure tensor math - no CUDA rasterization, no model. It runs fast and on CPU if that's what your frames are on.

Installing

Ships with the pack; the standard setup applies:

cd ComfyUI/custom_nodes
git clone https://github.com/styletransfer/ComfyUI-TRELLIS2_Motion
cd ComfyUI-TRELLIS2_Motion
python install.py --full
pip install git+https://github.com/microsoft/TRELLIS.git
pip install git+https://github.com/NVlabs/nvdiffrast.git

Restart ComfyUI; it's under TRELLIS2/Effects. (Strictly, for this node alone you only need torch - the heavy TRELLIS and nvdiffrast installs are for the rest of the pack.)

Common issues

• Ghosts too faint to see. At 0.3 opacity on a busy background they'll disappear. Raise ghost_opacity or switch blend_mode to add - additive ghosts are hard to miss.
• No ghosts at the start of the clip. The first few frames don't have enough history behind them (nothing at index −5 exists), so early frames look plain. That's expected behavior, not a bug - and it's exactly what include_future is for if you want a trail on both sides.
• Ghosts everywhere, looks like a mess. Drop ghost_count to 2 and raise frame_interval to 10. With a fast-moving subject you want fewer, further-apart ghosts, not a dense smear.