WanAnimatePlus ContextOptions
Windowing settings for when your video won't fit in memory
- reference_latent
- context_options
Wan's native context is 81 frames, and its VAE compresses time 4x - which is the source of most of this ecosystem's weird arithmetic. Want 161 frames? The model can't attend to them all at once, and your GPU probably can't hold them. Context windowing is the answer: split the generation into overlapping chunks, sample each against the shared conditioning, and blend the seams. This node builds the settings object that tells the WanAnimatePlus Sampler how to do exactly that. Its single output, context_options, feeds the sampler's context_options input - and the same object also fits SamplerExtraArgs if you're bundling.
The numbers, in pixel frames. All three key inputs are pixel frames, and the tooltips hammer the same reminder: the latent space packs 4 frames into 1, so these values divide by four internally. context_frames (default 81) is how long each window is. context_stride (default 4) is how far the window advances. context_overlap (default 16) is how much adjacent windows share - that overlap is what gives the fusion pass material to blend, and it's the knob that trades artifacts against speed. Too little overlap and the seams show; too much and you're re-computing the same frames constantly.
The other settings. context_schedule offers three windowing strategies: uniform_standard, uniform_looped (for loop-friendly generation) and static_standard. freenoise (on by default) shuffles the noise between windows so chunked generations don't develop a repeating tic. fuse_method is where a lot of the seam quality actually lives - linear ramps weights at window edges only, pyramid uses triangular weights that peak in the middle, and for visible seams the pyramid usually wins at the cost of a bit of extra compute. verbose prints the window math so you can actually see what the scheduler decided. And there's reference_latent, which seeds non-first windows with an image latent - the tooltip flags it as mostly useful with MAGREF models, where every window needs a first-frame anchor.
The honest framing. This node is the manual version of what SCAIL-2's built-in loop and the community chunker nodes automate. The 81-frame ceiling hasn't gone away; the workflow burden is just moving into single nodes. For long Wan Animate clips, context windows are how you get there, and the classic failure mode is identity drift across chunk boundaries - no context_overlap setting fixes that, it's the model's limitation. Keep overlap at 16 to start, drop context_frames if you're OOMing rather than fighting it, and only touch context_schedule if you know what looped scheduling buys you.
Install via ComfyUI Manager (search "WanAnimatePlus") or git clone https://github.com/wuwukaka/ComfyUI-WanAnimatePlus into custom_nodes, restart, and keep the whole chain on WanAnimatePlus nodes - the README's one rule that saves you a headache.
Inputs (8)
| Name | Type | Default | Description |
|---|---|---|---|
| context_schedule | COMBO | 3 options: uniform_standard, uniform_looped, static_standard | |
| context_frames | INT | 812–1000 | Number of pixel frames in the context, NOTE: the latent space has 4 frames in 1 |
| context_stride | INT | 44–100 | Context stride as pixel frames, NOTE: the latent space has 4 frames in 1 |
| context_overlap | INT | 164–100 | Context overlap as pixel frames, NOTE: the latent space has 4 frames in 1 |
| freenoise | BOOLEAN | true | Shuffle the noise |
| verbose | BOOLEAN | false | Print debug output |
| fuse_methodopt | COMBO | linear | Window weight function: linear=ramps at edges only, pyramid=triangular weights peaking in middle |
| reference_latentopt | LATENT | Image to be used as init for I2V models for windows where first frame is not the actual first frame. Mostly useful with MAGREF model |
Outputs (1)
| Name | Type | Description |
|---|---|---|
| context_options | WANVIDCONTEXT | — |