CogVideo Enhance-A-Video

This is one of those "cost you nothing, might help" nodes. Enhance-A-Video (from the NUS HPC-AI Lab) is a training-free tweak to how attention works during generation - it nudges the model toward more temporal coherence and sharper detail without retraining anything and without a separate model to download. You wire it into the sampler, and it does its thing during denoising. If your CogVideoX clips are coming out a touch soft or flickery, it's worth a try.

The key phrase is training-free. There's no checkpoint here, no LoRA, no extra weights - it's a modification applied at inference time, tuning the balance of the model's temporal attention so frames relate to each other a little better. That's why it's cheap to try: the downside is mostly just a bit of extra compute, not a heavy dependency.

The inputs that matter

There are only three, all required, and honestly you'll mostly touch one:

• weight (default 1) - how strongly the enhancement is applied. 1 is the sensible starting point. Crank it up for a stronger effect, dial it down if it starts to over-sharpen or introduce artifacts. This is the knob you'll actually adjust.
• start_percent (default 0) and end_percent (default 1) - the slice of the denoising schedule the effect is active over, from start to finish. The defaults apply it across the whole generation. You can restrict it to, say, the later steps if you only want it refining detail rather than shaping structure.

The single output is feta_args (a FETAARGS bundle - "FETA" for the Enhance-A-Video method's internals), and it plugs into the feta_args slot on CogVideo Sampler. That's its only destination. The sampler applies it while it works; there's no separate pass.

How to install it

It's part of the CogVideoX wrapper - no separate install for this node:

• ComfyUI Manager - search ComfyUI-CogVideoXWrapper, install, restart.
• Manual - cd ComfyUI/custom_nodes && git clone https://github.com/kijai/ComfyUI-CogVideoXWrapper, then pip install -r ComfyUI-CogVideoXWrapper/requirements.txt, restart.

Common issues & troubleshooting

You can't tell if it's doing anything. That's the honest reality with subtle inference-time tricks - the effect varies by prompt, model, and settings. The way to actually see it: generate the same seed twice, once with the feta_args connected and once without, and compare. Fixed seed is the only fair test.

It over-sharpens or adds weird texture. Turn weight down. At high values the enhancement can push too hard and introduce artifacts instead of cleaning them up. Back off toward 1, or restrict it to later denoising steps with start_percent/end_percent.

It's not a magic quality button. Enhance-A-Video helps at the margins; it won't turn a bad generation into a good one or rescue a weak prompt. Treat it as a finishing touch you stack on top of good base settings, not a fix for fundamental problems. It also plays fine alongside the caching and text-encode nodes - they all feed the sampler independently.