Difforum · Echo Trails

Two inputs, one output, and a look you either want or you don't. Echo Trails lays the past of a clip over its present, like leaving a camera shutter open: bright areas smear into where they came from, and motion gets a tail.

What it does

It walks the batch forward, accumulating a decayed copy of previous frames and mixing it into the current one. decay (0.6) is how long trails last - 0.9 is genuinely long exposure, where a light source draws a continuous streak for seconds. mix (0.5) is how much of that trail is laid over the frame, so you can have a subtle ghost or a full smear.

Both at default gives you a soft dreamy drag on anything bright. decay 0.9, mix 0.95 gives you something closer to a light-painting photograph, and it will also blow out highlights on any clip with a strong light source, which is either the point or the problem depending on your clip.

The whole thing is deterministic frame maths - no sampling, no optical flow, no GPU work worth mentioning. It runs at effectively instant speed compared to anything upstream of it.

Where it goes in the graph

After the render, after the video model, before the save. It works on any batch of frames, so it doesn't care whether they came from the Feedback Sampler, H3, LTX or a file you loaded. That's why it's in the Post group alongside Loop, Symmetry and Flow Stabilize rather than the render group - it's a treatment, not a generator.

Two pairings worth knowing:

• With a Shake or Handheld camera move. The camera's jitter makes the trail wander, which is what makes it read as a physical long exposure rather than a static overlay. A locked-off still shot with heavy trails looks like an accident; a shaky one looks like a photograph.
• With Look Mix or Detail Guard afterward. Trails soften the image (you're averaging frames, after all), so a small unsharp pass after it puts the structure back without killing the effect.

What it isn't

It is not a motion blur. Motion blur is directional and follows the actual velocity of each pixel in one frame; this is temporal accumulation and it makes trails behind everything, including things that aren't moving. On a mostly-static clip the difference is invisible. On a busy one, Echo Trails produces a thicket rather than a smear.

It's also not the same as the cadence crossfade inside the Feedback Sampler. That one exists to hide the missing frames between diffusion passes; this one is a look.

Install

ComfyUI Manager → search "Difforum", or:

cd ComfyUI/custom_nodes && git clone https://github.com/chillithebillis/Difforum.git difforum

Restart ComfyUI. numpy and torch - the base install is enough for this node.

Where people get burned

Memory, mildly. The input batch is already in RAM; the accumulation buffer is the same size, so on a very long 2K clip in a crowded graph you've added one more full clip's worth of tensors. On a 240-frame 1080p batch that's noticeable but not fatal. On a 2,000-frame one, render in chunks.

And the highlight crush. Any clip with a sun, a lamp or a bright practical will have a hard-edged blob after a couple of seconds at decay 0.9. Try it, look at the last frame, and decide whether you want the trails or the highlights - you generally can't have both without a compression step in between.