🐳 WanVideo视频集成采样器——Github:@luguoli

The name is doing a lot of work. A stock Wan 2.2 graph is two samplers by design - a high-noise expert that handles motion and composition, then a low-noise expert that refines detail - plus a CLIP text encoder, a VAE encode for your reference frame, a latent, a decode, and a save node holding it together with a small forest of wires. This node from @luguoli (same author as ComfyUI-Hive) collapses most of that into one box: hook up the model pair, CLIP, and VAE, type a prompt into the built-in text box, and frames come out the other side. Its own README calls it the "moms no longer have to worry about messy connections" node, and honestly that's the pitch.

Why you'd reach for it

Wan 2.2's two-pass architecture is why the community's speed-LoRA advice works - slap the fast LoRA on the low-noise pass only, keep high noise untouched, and you keep motion quality while cutting the slow part. Assembling that graph by hand is where beginners burn an afternoon, and this is the "I don't want to hand-wire it" answer. It also pulls in the chores people usually bolt on separately - prompt encoding, frame sizing, VRAM cleanup, batch generation, even a completion chime.

How it works

The sampling core is the standard dual-pass: the high-noise model denoises steps 0 through steps_high_noise, then the low-noise model picks up the latent and finishes the rest, with noise disabled and full denoise forced (that's straight from the pack's common_ksampler). The switch point is exactly the step split you type in. Around that core it clones both models and patches them: it flips on torch FP16 accumulation (needs PyTorch 2.7+), overrides attention with SageAttention when asked, applies an SD3-style sampling shift, and - for image-to-video - VAE-encodes your start/end frames and feeds them in as concat-latent conditioning, the same trick the official Wan I2V recipe uses. The motion_amplitude boost comes from PainterFLF2V and only fires when you've given it both a start and an end frame.

The inputs that matter

model_high_noise, model_low_noise, clip, and vae are required, and they come from the normal Wan loaders - ComfyUI loads Wan 2.1/2.2 natively, no wrapper needed. Everything else you mostly leave alone until you need it:

• length / width / height - 81 frames is the default and roughly the native context window (about 5 seconds). Width and height auto-round to multiples of 8, and the 720×1280 default is portrait - change it if that's not your intent.
• steps_high_noise / steps_low_noise - defaults are 4 and 4. Full-quality Wan 2.2 wants ~20–30 steps total without a speed LoRA; with CausVid or lightx2v you're back down to 3–4 per pass.
• cfg_high_noise / cfg_low_noise - default 1.0. If you're running a speed-distilled LoRA, keep it at 1. If you're doing full-quality no-LoRA runs, the community found 3.5+ on the high-noise stage gets noticeably better prompt adherence.
• start_image / end_image - wire a start image to switch from text-to-video to image-to-video. end_image requires a start_image or the node throws.
• Worth knowing: sage_attention (default "auto"), sd3_shift (default 5), the two memory-cleanup toggles, batch_size (1–10), and the seed.

Outputs: a full frame sequence (FrameImages, IMAGE - wire it to a save/preview node), a LastFrameImage (IMAGE - handy for chaining a follow-up clip off its ending), and an optional Latent.

Install

ComfyUI Manager → search "ComfyUI-Wan-Video-Integrated-KSampler" → Install. 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. No models ship with the pack - you bring the Wan pair yourself. The base requirements are light (torch, transformers, einops, tqdm); the heavy stuff is optional. SageAttention needs pip install sageattention (Windows needs the community Triton builds), the fp16-accumulation toggle needs PyTorch 2.7+, and the completion sound is Windows-only.

Common issues

• VRAM is the wall. Wan 14B wants 12–16 GB minimum and 24 GB for comfortable 720p; an 81-frame clip is 15–40 minutes on a 4090 without a speed LoRA. The levers are all in this node's panels: fp8/GGUF checkpoints, SageAttention, and enable_clean_gpu_memory between runs.
• SageAttention silently failing. If you see "sage attention failed" in the console, the package isn't installed - the node disables it and carries on. Set it to disabled and install sageattention before turning it back on.
• Block swap is gone. The update log says it was removed because newer ComfyUI versions block it. If a saved workflow references it, that's why it vanished.
• The negative prompt ships pre-filled with Chinese text. Clear it if you don't want it.

It's a convenience node, not a magic one - the native 81-frame window and the VRAM appetite are still yours to manage. But for churning out 2.2 clips without rebuilding the graph every time, it does exactly what it says.