WanVideo Sampler From Settings

Kijai says the quiet part out loud in this node's own description: the main WanVideo sampler "has become a messy monster." A year of Wan features - speed LoRAs, TeaCache, block swap, NAG, context windows, half a dozen embed types - all bolted onto one node, and it shows. WanVideoSamplerFromSettings is the cosmetic fix. It does the actual sampling, but instead of exposing every setting as its own input, it takes one pre-assembled sampler_inputs bundle and runs it. The description is refreshingly honest: "no other functionality than to look cleaner, useful for the live preview."

So this isn't a new sampling algorithm or a speed trick. It's the same denoising you'd get from the big sampler, wrapped so your canvas stays readable and the live preview isn't buried under forty widgets.

How it works

All the decisions - which model, how many steps, what scheduler, which cache and LoRA settings - get packed into a single SAMPLER_ARGS object upstream. This node unpacks that object and does the denoise. Think of it as the "run" button at the end of a settings pipeline: the thinking happened in the nodes feeding it, this one just executes and gives you latents back.

The inputs and outputs that matter

There's really only one input to point at:

• sampler_inputs (type SAMPLER_ARGS) - the bundle of everything the sampler needs. You don't hand-fill this; it comes from a companion settings node in the WanVideoWrapper that gathers the model, embeds, steps, scheduler, and the rest into one connection.

Two outputs:

• samples (LATENT) - the finished latents. Wire these into your VAE decode to get the actual video frames.
• denoised_samples (LATENT) - the denoised prediction, handy for previews, chaining a second pass, or grabbing the intermediate result.

That's the whole surface. The complexity lives upstream by design.

How to install it

Part of the WanVideoWrapper pack. 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. As with every node in this pack, the wrapper is the code and you still need the Wan weights themselves - the diffusion model, the Wan VAE, and a T5 text encoder - downloaded into your models folders. Kijai's fp8 packs at huggingface.co/Kijai/WanVideo_comfy are where most people get them.

Common issues & troubleshooting

Nothing connects to sampler_inputs. This node is useless on its own - it's the back half of a two-part split. If you don't have a node emitting SAMPLER_ARGS, you're missing the front half. Grab an example WanVideoWrapper workflow rather than trying to wire this from scratch; it's meant to be dropped into an existing graph, not built up freehand.

You expected knobs and there are none. That's the point. If you need to change steps, CFG, or the scheduler, you change them on the settings node upstream, not here. This node deliberately has one input.

Output looks identical to the big sampler. It should. Same math, same result - the only thing that changed is how much of your screen the node eats. If your generation is bad, the problem is in the settings you fed it, not in choosing this node over the monster.