WanVideo Experimental Args

This is the junk drawer, and I mean that affectionately. The node's own description is literally "Experimental stuff" - it's where kijai parks sampling techniques that are worth trying but aren't stable enough to promote into the main sampler. Everything defaults to off, so dropping it into a graph changes nothing until you flip a switch. That's the right way to use it: leave it neutral, then turn on one thing at a time and see if your specific workflow likes it.

Don't cargo-cult this node. These are guidance and score-manipulation tricks; some help on some content and quietly hurt on other content, and stacking five of them at once means you'll never know which one did what.

How it works

It bundles a set of independent sampling-time modifications into one EXPERIMENTALARGS object that the sampler reads. Each toggle patches a different part of the guidance or scoring:

• cfg_zero_star (+ use_zero_init, zero_star_steps) - CFG-Zero*, a tweak that zeros out or stabilizes guidance on the earliest steps, where the model's prediction is noisiest and standard CFG can push things in a bad direction. zero_star_steps sets how many early steps get the treatment.
• use_fresca (+ fresca_scale_low, fresca_scale_high, fresca_freq_cutoff) - FreSca, which applies guidance differently across frequency bands: low frequencies (broad structure) and high frequencies (fine detail) get scaled separately, split at fresca_freq_cutoff. Lets you push detail without over-cooking composition, or vice versa.
• temporal_score_rescaling (+ tsr_k, tsr_sigma) - rescales the score across the temporal dimension, aimed at motion stability over the clip.
• use_tcfg, raag_alpha, bidirectional_sampling, video_attention_split_steps - further experimental guidance/attention variants, all off or neutral by default.

The output is exp_args (EXPERIMENTALARGS) → into the sampler.

The inputs that matter

Honestly, most of them don't, most of the time. The two I'd actually reach for:

• cfg_zero_star - flip it on if your early-step composition or color looks off; it's one of the more broadly-liked of these.
• use_fresca with the scale/cutoff trio - if you want a detail or sharpness nudge without changing steps or sampler.

Everything else: turn on singly, compare against a baseline render with the same seed, keep it only if you can see the improvement.

How to install it

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

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

then restart. Nothing to download.

Common issues & troubleshooting

You turned on everything and results got worse. Predictable. These interact, and several change the guidance math in overlapping ways. Reset to all-off, change one toggle, fix your seed, compare. Treat it like A/B testing, not seasoning to taste.

It's "experimental" for a reason. Features here can change or break between pack updates - that's the deal with the junk drawer. If a graph you saved suddenly errors on this node after an update, an input probably got renamed or removed; delete and re-add the node and reconnect. Don't build a production pipeline that depends on a specific experimental toggle behaving forever.

No visible effect. Some of these only bite in specific regimes - a particular CFG range, step count, or content type. If a toggle does nothing on your workflow, that's a legitimate outcome; it isn't a universal quality button.