WanVideo Enhance-A-Video

Enhance-A-Video is one of those "why not" nodes - it's a lightweight technique from the NUS HPC-AI Lab that nudges a video diffusion model toward better temporal coherence and detail, and it's cheap enough to leave on. This is the WanVideoWrapper wrapper around it. You don't feed it a model or an image; you just set a strength and a schedule, and it produces a FETAARGS bundle that rides into your sampler. The effect is a modest bump in how consistent and crisp the motion looks, without the big compute cost of, say, a second sampler pass.

It's not a magic fix and it won't rescue a bad prompt, but it's a low-risk knob that's worth having in the graph.

How it works

The technique (the "FETA" in the output type - Enhance-A-Video's cross-frame attention enhancement) tweaks how the model attends across frames during denoising, boosting the temporal signal so frames relate to each other more coherently. Because it's applied as an argument to the sampler rather than a model patch, it composes cleanly with the rest of a WanVideoWrapper graph. You control its strength and which slice of the denoising schedule it's active over.

The inputs and outputs that matter

There's no model or image input here - it's pure config:

• weight (default 2) - the strength of the enhancement. The default of 2 is a real starting point; push it up for a stronger effect, but watch for over-sharpening or motion artifacts if you crank it.
• start_percent / end_percent (defaults 0 and 1) - the portion of the denoising schedule it runs over. Full range by default. Early steps set up structure, so if you want the enhancement to only touch detail refinement you can start it partway in.

The single output, feta_args (FETAARGS), goes into WanVideoSamplerExtraArgs (its feta_args slot), which then feeds the sampler.

How to install it

ComfyUI Manager - search ComfyUI-WanVideoWrapper, install, restart. Manual:

cd ComfyUI/custom_nodes
git clone https://github.com/kijai/ComfyUI-WanVideoWrapper
pip install -r ComfyUI-WanVideoWrapper/requirements.txt

then restart. Nothing to download - it's an algorithmic tweak, not a model.

Common issues & troubleshooting

You added it and see no difference. The effect is subtle by design, and it only reaches the generation if feta_args is actually wired through WanVideoSamplerExtraArgs into the sampler's extra_args. A dangling FETAARGS output does nothing. Confirm the chain is connected end to end.

Cranking weight made it worse. Over-driving the enhancement can introduce shimmer or over-sharpened edges that read as artifacts. Back off toward the default of 2 - this is a seasoning, not a main ingredient.

It fights your other accelerators. Enhance-A-Video changes attention behavior, so if you're also running an aggressive cache or an exotic attention backend and something looks off, isolate them one at a time. Start with just Enhance-A-Video on a clean graph, confirm it helps, then layer the rest back in.