Temporal Blur Module (Bending, Experimental)

Every spatial operation in this pack blurs within a frame. Temporal Blur Module blurs across frames - a 1-D Gaussian running along the time axis - which does something no post-process quite reproduces: the model's own sense of motion gets smeared, so content lingers, trails and bleeds forward instead of cutting cleanly between moments.

It's one of four temporal modules in the pack's experimental video section, and it's the one whose effect is most visible at small settings.

How it works

The module runs a separable Gaussian along the frame dimension of a 5-D video tensor, with edge replication at the ends - so the first and last frames get padded with themselves rather than going dark. Kernel radius is roughly 3σ, and sigma is measured in latent frames. On WAN that means one unit is four video frames (the 3D VAE's temporal compression), so an 81-frame clip is 21 latent frames, and sigma: 2 has a reach of roughly six latent frames either side - about 24 video frames, most of the clip. It bites much harder than the number suggests.

It only acts on 5-D tensors. A 4-D input returns unchanged - there's no "one frame" fallback, and no warning, so a temporal blur that seems to do nothing almost always means you attached it somewhere that isn't a video model.

What it does to the image depends entirely where you apply it:

• On blocks (via DiT Block Bending), the video token stream gets time-smeared, so the model's internal motion representation loses sharpness. Motion gets slower and mushier; fast actions become a soft smear rather than a sequence of poses.
• On the attention map (via Attention Map Bending), each prompt token's attention is averaged with neighbouring frames, so what the prompt "means" at any given moment is partly borrowed from the frames around it. Subject boundaries stop moving crisply with the action.

Both are the same mechanism at different depths. The attention route is the one the pack's own temporal workflow takes: a Temporal Blur at sigma 1.5 on self_query attention over blocks 10-20, early steps, renormalize: none.

Inputs and output

One input: sigma, default 1.0, range 0 to 16, step 0.05, in latent frames. One output: BENDING_MODULE.

Sigma 0 is a no-op. If the sequence is shorter than two frames, it's a no-op too.

Wiring it

DiT Block Bending for the token-stream version and Attention Map Bending for the attention version, both on a WAN-family video model, with spatial left on so the tokens are on the (frames, height, width) grid. Use self_query, self_key or cross_text for attention - for self_key the values moved are the ones being smeared in time.

Worth pairing with Timestep Gated Bending: temporal smear applied only in the early steps changes how the model conceives the motion, while the late steps still render a clean image. Applied at every step, it tends to make the whole clip look like it was shot through syrup.

Install

ComfyUI Manager → search "ComfyUI Model Bending", or:

cd ComfyUI/custom_nodes
git clone https://github.com/abuzreq/ComfyUI-Model-Bending
pip install -r ComfyUI-Model-Bending/requirements.txt

kornia and scikit-learn, then restart ComfyUI, and check for [model-bending] lines in the console when something misbehaves.

Where people get burned

• Thinking in video frames. Sigma is in latent frames. Halve the number you were about to type, then halve it again if you're on a long clip.
• Expecting motion blur. This isn't shutter blur, where a moving edge streaks. It averages the model's representation across time, so a fast action tends to slow down rather than streak.
• Using it as a repair tool for a single bad frame. The blur has no idea which frame is bad, and one frame of ghost in 21 is worth exactly as much as its neighbours. If you're chasing a flash frame, a temporal blur can hide it, but it'll also soften the whole clip's motion.
• Bending at every step and every block. Same advice as the rest of this video section: a couple of blocks and a window of steps, at 720p, at least to start. Video inference is already the slowest thing you'll run today, and this makes it slower.
• Reading the tooltip as a promise. The author is explicit: these temporal modules were validated on tiny random-weight WAN models in the test suite, not on real weights - fair game for reporting what you see, not a supported feature in the usual sense.