WanVideoSampler v2 Extra Args

The v2 sampler kept its own face clean by shoving all the optional, advanced behavior into one collector node. This is that node. It's a manifold: every extra thing you might bolt onto a Wan generation - caching, context windows, cross-frame enhancement, skip-layer guidance, RIFLEx position extension, and the whole family of talking-head and pose add-ons - plugs in here, gets bundled into a single WANVIDSAMPLEREXTRAARGS, and rides into the sampler's one extra_args input. Instead of ten inputs cluttering the sampler, you get one.

Nothing here is required - every input is optional. You wire in only the pieces your workflow actually uses and leave the rest empty.

How it works

It's an aggregator with no logic of its own; it collects whatever you connect and packages it for the sampler. That design is why the sampler stays readable and why you can build up a complex Wan graph incrementally - add a cache, test, add context windows, test - all through this one junction.

The inputs that matter (all optional)

• cache_args (CACHEARGS) - a caching node's output (EasyCache and friends). This is the big speed lever most people wire first; temporal caching skips redundant compute across similar frames.
• feta_args (FETAARGS) - from WanVideoEnhanceAVideo, for the temporal-coherence bump.
• context_options (WANVIDCONTEXT) - sliding-window context for going past the 81-frame native limit.
• slg_args (SLGARGS) - Skip Layer Guidance, a quality-tuning technique.
• riflex_freq_index (default 0) - RIFLEx frequency extension, which pushes the usable length before a clip "starts acting strangely" past the usual ~121-frame ceiling. 0 = off.
• rope_function (default / comfy / comfy_chunked) - the RoPE implementation; comfy is the default and comfy_chunked helps memory on long sequences.
• multitalk_embeds, fantasytalking_embeds, unianimate_poses, uni3c_embeds, loop_args, experimental_args - the specialized channels. Wire multitalk_embeds here from MultiTalkWav2VecEmbeds for lip-sync, unianimate_poses for pose-driven animation, and so on.

The single output, extra_args (WANVIDSAMPLEREXTRAARGS), connects to WanVideoSamplerv2's extra_args slot.

How to install it

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. The node downloads nothing itself, but the things you plug into it may - MultiTalk weights, cache backends, and so on live with their own nodes.

Common issues & troubleshooting

You wired an add-on and it does nothing. The number one cause: the add-on's args reached this node, but this node's extra_args output isn't connected to the sampler. It's a pass-through - if the final link to the sampler is missing, everything upstream is inert. Trace it end to end.

Two features conflict. Some combinations don't play nicely - an aggressive cache plus certain context options, or stacking multiple guidance tricks. When output looks wrong, disconnect extras one at a time to find the culprit rather than debugging the whole bundle at once. The modular design makes this easy; use it.

Your long clip still degrades past ~121 frames. RIFLEx (riflex_freq_index) and context options are what push the length ceiling, but they're mitigations, not miracles - the native context is still 81 frames and identity drift across windows is the real, unsolved limit. For genuinely long continuous shots, that's SVI's territory, not this node's.