WanVideo Scheduler v2

This is the tidied-up version of WanVideoScheduler, and if you're building a Wan graph today it's the one to grab. Same job - decide the noise schedule Wan denoises along, separately from the sampler - but instead of spraying six separate outputs across your canvas, v2 packs everything into a single WANVIDEOSCHEDULER bundle. One noodle into the sampler instead of a fistful. That's the whole pitch, and honestly it's a good one.

The reason a "just make it cleaner" node exists at all is that Kijai's WanVideoSampler has accumulated a lot of inputs over the last year of Wan features, and wiring the schedule as one object rather than as loose steps / shift / scheduler / sigmas wires keeps the graph readable. Nothing about the actual math changed.

How it works

Wan is a flow-matching model, so it walks a continuous curve of sigma values from full noise to clean image rather than stepping on a fixed grid. The scheduler sets the spacing of that walk. This node computes the sigma curve from your settings and emits it as a bundle the sampler unpacks internally. You're configuring how the denoising is paced, not what it denoises.

The inputs and outputs that matter

The inputs are identical to the original:

• scheduler (default unipc) - 21 choices. unipc, euler, dpm++ and the /beta flavors are your general-purpose options; the specialty ones (flowmatch_causvid, flowmatch_distill, lcm, longcat_distill_euler, res_multistep) exist to match specific distilled/speed-LoRA setups. Match the scheduler to your model, don't just leave it on unipc if you're running CausVid.
• shift (default 5) - flow-matching timestep shift. Nudge it up to bias the schedule toward high-noise motion/composition work, down to favor fine detail.
• steps (default 30) - 30 is full quality; 4–8 with a speed LoRA.
• start_step / end_step - leave at 0 / -1 unless you're deliberately slicing for multi-stage sampling.

Optional sigmas lets you feed in a schedule from another node, and enhance_hf is a high-frequency toggle you can ignore at first.

The single output is scheduler (type WANVIDEOSCHEDULER) - plug it into the matching input on your WanVideo sampler and you're done.

How to install it

Comes with the WanVideoWrapper. 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. Remember the wrapper is just the nodes - you still need the actual Wan weights (diffusion model, VAE, T5 text encoder; Kijai's fp8 sets at huggingface.co/Kijai/WanVideo_comfy) in your models folders.

Common issues & troubleshooting

"v2 vs v1, which do I use?" v2. The only reason to stay on the original WanVideoScheduler is an old workflow that already wired its six separate outputs and you don't want to redo them.

The bundle won't connect. WANVIDEOSCHEDULER is a wrapper-specific type - it only fits the scheduler input on the WanVideoWrapper's own sampler, not a stock ComfyUI KSampler. If the noodle refuses to snap, you're aiming at the wrong sampler.

Scheduler doesn't match the model. Same trap as always: a distilled scheduler on a full run (or vice versa) looks wrong. Pick the scheduler that fits how you loaded and LoRA'd the model, and change shift before you go blaming the scheduler for soft output.