⚙ Image Sequence Blend

Frame-by-frame animation has a signature flaw: every frame is generated independently, so moving objects strobe - each frame's noise and details jump around instead of flowing. Image Sequence Blend [Dream] is the post-processing pass that fixes the feel of that: it takes your rendered animation sequence and blends each frame with its neighbors, producing the soft, motion-blurred look that reads as "video" instead of "slideshow."

It's one of the sequence-processing nodes in the Dream Project Animation Nodes pack (alt-key-project), the 2023 Deforum-style animation toolkit. It operates on the pack's ANIMATION_SEQUENCE type - the thing Image Sequence Saver produces - before you hand the frames to a video encoder.

What it does

Inputs: sequence plus three tuning knobs:

• fade_in (FLOAT, 0.01–0.5, default 0.1) - how much of the previous frame's look bleeds in
• fade_out (FLOAT, 0.01–0.5, default 0.1) - how much of the next frame's look bleeds in
• iterations (INT, 1–10, default 1) - how many passes to run; each pass blends the already-blended frames

Output: a sequence, same length, same timeline, softened.

Mechanically each frame is recomputed as prev*fade_in + current + next*fade_out (roughly), so the knobs set the width of the temporal blur. Higher fade values = more blur and smoother motion but less sharpness. Multiple iterations stack the effect - great for an aggressive "film grain smoothing" feel, but don't expect to recover detail you blended away.

How to slot it in

It's deliberately the middle of a pipeline, never in parallel - the README warns the sequence nodes are "severely limited" and shouldn't be parallelized. The typical flow:

Image Sequence Saver → Image Sequence Blend → FFMPEG Video Encoder

Render your animation to frames, run the blend, then encode. If the result is too mushy, the pack's Image Sequence Tweening node (which interpolates in-between frames instead of blurring) is the alternate path.

The honest trade-off

This is a cheap, frame-count-preserving smoothing, not optical-flow interpolation. It won't fix genuinely broken motion - if two frames are wildly different, blending them just gives you a smear. Where it shines is killing the low-level flicker and noise dance between similar frames, which is the exact thing that makes Deforum-style output look amateurish. One light pass (0.1/0.1, 1 iteration) is a good starting point for most renders.

Installing

cd ComfyUI/custom_nodes
git clone https://github.com/alt-key-project/comfyui-dream-project.git

Or via ComfyUI Manager under "Dream Project Animation". No model downloads; deps are imageio, scipy, torchvision, pilgram, evalidate (with the pack's numpy<2.0 pin to watch). Find it under Dream → animation → postprocessing.