Nifty Wan Video SLG

SLG stands for Skip Layer Guidance, but the technique this node wraps is really STG - Spatiotemporal Skip Guidance, from a CVPR 2025 paper (arxiv 2411.18664). Same family as NAG: guidance that isn't CFG. Where NAG pushes against a negative reference in attention space, STG measures what changes when you skip part of the model and steers with that difference. It works at CFG=1, which is the whole point - the fastest, most distilled Wan setups (LightX2V, few-step Lightning) run at CFG=1 and have no guidance budget left. STG gives them steering back, with no TeaCache required.

This is Nifty's WAN-native implementation, forked from KJNodes and reworked. You feed it your Wan model, it returns a patched model, and the sampler handles the rest.

How it works

The node patches a post_cfg_function onto the model. On every denoising step, it runs a second conditional pass through the model where your chosen blocks have been sabotaged - and the difference between the sabotaged output and the normal output becomes the correction, scaled and added to the result. It's guidance by subtraction: "here's what the model thinks without these blocks; steer toward the difference."

Two sabotage modes, and this is the main design decision you'll make:

• STG-A (default) - zeroes only the self-attention in the chosen blocks. Cross-attention and the feed-forward network still run. Less perturbation, lower flicker risk. Recommended for Wan at few steps.
• STG-R - the entire block becomes identity. Stronger correction, but noticeably more prone to flicker at 2–3 steps.

Inputs that matter

• model - the Wan model to patch. enabled - master switch; when False the model passes through untouched.
• blocks - comma-separated block indices, defaults to 9, 10, and supports ranges like 9-11. WAN 2.2 has 40 layers, and blocks 9/10 are the empirically recommended targets. The tooltip's advice for flicker is golden: start with 10 alone. Fewer blocks = safer.
• mode - STG-A or STG-R, above.
• scale - guidance strength, default 1.0. For LightX2V (2–3 HN steps), STG-A maxes out around 2.0–2.5 and STG-R around 1.5 before quality falls apart.
• start_percent / end_percent - the active range as a fraction of the full denoising schedule (HN + LN combined, so the percentages are absolute). Default is fully active (0.0→1.0). The formula to remember: end_percent = HN_steps / total_steps. A 4-step run split 2+2 means HN does half the work, so set 0.5 to confine STG to the high-noise pass where it has the most structural impact.

Optional extras: rescaling_scale (default 0) normalizes the correction's std against the normal prediction's std before scaling - set 0.3–0.5 if high scale makes it flicker. scale_end (default 0, disabled) linearly fades the scale from scale at the start to scale_end at the end, so you can run strong early and back off as noise clears.

Output is a single MODEL - the patched model, straight to your sampler.

Wiring it in

The workflow shape is: Wan model → Nifty Wan Video SLG → sampler at CFG=1. That's the entire integration. It composes with NAG too - different guidance families, and plenty of people run both - but start with one. SLG is the one that pairs with distilled/few-step models because it doesn't need a negative slot at all.

Installing it

Part of Nifty Nodes for ComfyUI (Stibo/comfyui-nifty-nodes). ComfyUI Manager → search "Nifty Nodes", or:

cd ComfyUI/custom_nodes
git clone https://github.com/Stibo/comfyui-nifty-nodes

Restart. Needs a current ComfyUI (v3 API, comfy_api.latest) and a Wan 2.1/2.2 model. No pip packages, no model downloads from the pack.

The honest troubleshooting bit

STG's main failure mode is flicker, and it's fixable in this order: fewer blocks (10 alone), lower scale, then rescaling_scale at 0.3–0.5, then STG-A if you were on STG-R. If it's still dancing, your end_percent is probably too high - confine it to the HN phase. That ladder handles nearly every "it flickers" report, and it's all in the tooltips if you forget.