WanVideoEnhanceAVideo

The honest one-line pitch: you wire a Wan diffusion model into this node, and the clips it produces get visibly more motion. It's the same FeTA-style "enhance a video" trick that Kijai's WanVideoWrapper made famous in the Wan/Hunyuan scene - a small patch to how the model computes temporal attention - repackaged as a standalone node. If you've got a Wan 2.1 or 2.2 I2V setup where the output feels stiff, like the model is reusing the same pose frame after frame, this is the cheapest fix in the pack.

What it actually does

It doesn't touch pixels or prompt. It modifies the model - you put it between your model loader and your sampler. Inside, it clones the model and replaces the forward pass of every self-attention block in the Wan transformer with a modified version. That modified pass computes one extra thing: the average strength of off-diagonal temporal attention - how much each frame pays attention to frames that aren't itself - and multiplies the attention output by that scalar. Frames that all ignore each other get punished; frames that genuinely relate across time get boosted. That's the whole trick, and it's why it fixes "everything moves like a still photo" without touching your prompt or sampler settings.

The node is flagged EXPERIMENTAL = True in the source, which in ComfyUI-speak means "works, but the author isn't promising it'll survive API changes." Fine for hobby use; just don't build a business on it.

The inputs that matter

Only four, and they're all required:

• model - your Wan diffusion model, straight from a loader.
• weight - strength of the effect, 0 to 10, default 2. At 0 the node is a pure passthrough, which is a useful way to A/B test without rewiring.
• length - number of frames in the video you're about to generate. This matters more than it looks: the node uses it to figure out the temporal layout of the attention matrix. Set it to your actual clip length or the enhancement is computed against the wrong frame count.
• first_end_frame_shift - frame shift offset, default 0. This is the same frame-shift accounting that the pack's encode and decode nodes use; if you're only using this node standalone, leave it at 0.

Output is a single model, which you feed into your sampler of choice.

Install and gotchas

It ships in the przewodo/ComfyUI-Przewodo-Utils pack. Easiest path is ComfyUI Manager - search "ComfyUI-Przewodo-Utils" and install. Or:

cd ComfyUI/custom_nodes
git clone https://github.com/przewodo/ComfyUI-Przewodo-Utils.git

Then restart ComfyUI. The pack's pyproject pulls in companion packs (teacache, comfyui-kjnodes, comfyui-gguf, wanblockswap, ComfyUI-Rife-Tensorrt), so the Manager install can take a minute. No extra model downloads for this node itself - the Wan checkpoint and VAE are the usual ones you already have.

Two traps worth knowing. First, this node patches the Wan blocks structure, so feed it a Wan model and not an SD/SDXL one - it will just fail or do nothing useful. Second, and this is the one people trip on: if you're using the pack's own big WanImageToVideoAdvancedSampler, it already runs this exact enhance internally with video_enhance_enabled=True. Add this node on top and you're stacking the patch twice, which reads as jitter and overcooked motion. Use the standalone node with a plain sampler, or use the built-in toggle - not both.

Start at weight 2 and nudge up if motion still reads stiff; past 4-5 you'll start trading away stability for movement.