🐳 WanVideo视频集成采样器(简单)——Github:@luguoli

If the full WanVideoIntegratedKSampler is the all-in-one box, this is the stripped chassis. Same author, same pack, same dual-stage sampling core - and literally nothing else. No built-in prompt box, no start/end frame encoding, no memory cleanup, no SageAttention, no sound. The README describes it as "only the dual-model sampling integration for advanced custom use," and that's the whole story.

When you'd want it

You only reach for this node if you already have a working Wan 2.2 graph and the full version's conveniences get in your way. If you're assembling your own conditioning - a custom CLIP encode, a ControlNet or VACE setup feeding the latent, a prompt structure the built-in text box can't express - the Simple node lets you keep all that plumbing and replace just the pair of KSamplers with one box. It's the drop-in-for-the-samplers version, not the drop-in-for-the-whole-workflow version.

How it works

The mechanism is identical to the full node's core: the high-noise model denoises the latent through the first steps_high_noise steps, then the low-noise model takes over and finishes, with noise disabled and full denoise forced on the second pass. The switch happens exactly at the step split you enter. One simplification worth knowing: it feeds the same positive and negative conditioning to both stages - there's no per-expert conditioning here. If you want the classic speed-LoRA setup (fast LoRA on low noise only), that's applied at the model level on your model_low_noise before it reaches this node, so it still works.

The inputs

All twelve are required, and it's the shortest list in the pack:

• model_high_noise / model_low_noise - your Wan 2.1/2.2 expert pair from the normal loaders.
• steps_high_noise / steps_low_noise + cfg_high_noise / cfg_low_noise - same semantics as the full node. Full-quality runs want ~20–30 steps total with CFG 3.5+ on high noise; speed-distilled LoRAs run 3–4 per pass at CFG 1.0.
• noise_seed, sampler_name, scheduler - standard. euler + beta is the usual fast pick.
• positive / negative (CONDITIONING) - you bring these pre-encoded from a CLIP Text Encode or your custom setup.
• latent - you supply the latent too, and this is the one that bites people.

Output: a single Latent. Wire it into a VAE Decode (or whatever consumes your latent next). No frames, no last-frame shortcut.

Install

Same pack as the full node - install "ComfyUI-Wan-Video-Integrated-KSampler" once via ComfyUI Manager, or:

cd ComfyUI/custom_nodes
git clone https://github.com/luguoli/ComfyUI-Wan-Video-Integrated-KSampler.git
pip install -r requirements.txt

Then restart ComfyUI. You get both nodes from the one install.

Common issues

Because you construct the latent yourself, that's where the failures live. Wan 2.2's latent has to come from the right source - an official Wan empty latent or a proper VAE encode of your reference frame with the correct channel count and temporal layout. Feed it a latent shaped for another model and both stages will happily denoise it into garbage. Also remember the shared sampler raises if both step values are zero, and CFG semantics are Wan's, not SD's - a distilled workflow at CFG 5 will look burned, and a full-quality one at CFG 1 will drift from the prompt.

If you're new to Wan 2.2 entirely, start with the full node instead - it holds your hand through frame prep and decode. The Simple version is for the person who already knows exactly what they're doing and just wants the two samplers gone.