WanVideo SLG

Skip Layer Guidance is one of those settings that costs almost nothing and quietly makes Wan clips look better. It landed for Wan back in early 2025 and the community verdict has been steady since: skipping a layer or two in the unconditional pass gives you better motion coherency - less of that ghosting and skipping where limbs smear or the image jitters. This node produces the SLG config; you feed it into the Wan sampler.

The node description is terse and accurate: it "skips uncond on the selected blocks." Translated: during the guidance step, the model normally computes an unconditional prediction across all its transformer blocks. SLG deliberately drops a chosen block (or blocks) from that unconditional pass, and the effect is a cleaner, more coherent result.

How it works

Wan uses classifier-free guidance - it runs a conditioned prediction (with your prompt) and an unconditioned one (without), then pushes away from the unconditioned toward the conditioned. SLG's insight, borrowed from the SD3-era trick, is that certain blocks in that unconditional pass contribute noise or incoherence more than useful signal. Skip them and the guidance gets cleaner. For Wan specifically, the community landed on skipping block 9 or 10 as the sweet spot for quality and motion - which is exactly why the node's blocks default is 10.

The cost is essentially zero. You're not adding compute; you're removing a little. That's what makes it a near-free win rather than a tradeoff.

The inputs and outputs that matter

• blocks (STRING, default "10") - which block(s) to skip in the unconditional pass. It's a string so you can list more than one (e.g. "9, 10"). Start with the default; block 9 or 10 is the community-proven choice for Wan. Skipping the wrong blocks can hurt, so don't go wild here.
• start_percent (default 0.1) / end_percent (default 1) - the slice of the denoising schedule where SLG is active. The default skips the very first sliver then applies throughout. Rarely needs changing.

Output: slg_args (SLGARGS) → the SLG input on WanVideoSampler.

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 - SLG is a sampling-time technique that works with any Wan model you've already loaded.

Common issues & troubleshooting

It does nothing / makes things worse when I change blocks. The default 10 (and its neighbor 9) are that value for a reason - they're the blocks the community found improve Wan. Skip an arbitrary block and you can just as easily degrade the output. If you experimented and it got worse, go back to 10.

SLG only applies to the unconditional pass - so at CFG 1 it can't help. If you're running a speed LoRA (lightx2v/CausVid) that pins CFG to 1.0, there is no meaningful unconditional pass for SLG to touch. SLG earns its keep on full-guidance runs (CFG 3.5+). On the fast CFG-1 path, don't expect it to do much.

Marginal effect on already-clean output. SLG is a polish, not a rescue. On a well-tuned full-quality render the gain is subtle; where it shows most is fixing ghosting and incoherent motion on shakier setups.

Node missing. Update WanVideoWrapper and restart.