WanVideo Long SkyReelsV3 A2V

This is the node the pack itself describes as enabling "Multi/InfiniteTalk long video generation" - and InfiniteTalk is a name worth knowing if you've ever wanted to generate a talking-head clip longer than Wan's native few-second window. It's specifically for A2V: audio-to-video, the talking/singing/lip-synced-performance category where the driving signal is an audio track and the output needs to stay coherent for far longer than a single 81-frame Wan generation allows.

How it works

The node's own description is unusually direct about the mechanism: it generates video in windows with overlapping frames, chaining sampling windows together rather than trying to produce one arbitrarily long clip in a single pass. That's the same fundamental pattern the wider Wan ecosystem uses everywhere long-form generation is attempted - chunk, overlap, stitch - just implemented specifically for the SkyReels V3 A2V pipeline and InfiniteTalk-style talking content. It explicitly is not meant to combine with context windows or most of this pack's other long-video features; it's its own self-contained windowing approach, so don't expect to layer it on top of, say, the Diffusion Forcing sampler's own chunking behavior.

frame_window_size, motion_frame, and drop_frames are the three parameters that control the windowing math directly: how many frames each window generates, how many frames of motion context carry from one window into the next, and how many frames at the boundary get dropped (typically to avoid duplicated or degraded content right at the seam between windows).

The inputs that matter

vae (WANVAE) - your Wan VAE, needed to encode reference material. width/height (defaults 832×480) - your target resolution. frame_window_size (default 81, matching Wan's native context) - how large each generation window is. motion_frame (default 5) - how many trailing frames of motion get carried forward as context into the next window. drop_frames (default 12) - frames trimmed at each window boundary to keep seams clean. tiled_vae (default off) - enable if you're VRAM-constrained on the VAE decode step for long clips. colormatch (default disabled, several algorithm options) - corrects color drift across windows, which is a real and documented problem in long-chain generation; the KB notes exactly this kind of saturation creep in other extension workflows, so turning on a colormatch mode here is worth trying if your output shifts tone over its length.

Optional: start_image (IMAGE) to anchor the first frame, reference_video (IMAGE) for additional guidance, and clip_embeds (WANVIDIMAGE_CLIPEMBEDS) if you're feeding CLIP-based conditioning alongside.

Two outputs: image_embeds (WANVIDIMAGE_EMBEDS) for your sampler, and output_path (STRING), useful if this node writes intermediate windows to disk as it chains them.

How to install it

Via ComfyUI Manager, search WanVideoWrapper and install, or manually:

cd ComfyUI/custom_nodes
git clone https://github.com/kijai/ComfyUI-WanVideoWrapper

then restart. You'll need the SkyReels V3 checkpoint plus audio-driving components (an InfiniteTalk-compatible setup) beyond the base Wan download - these are separate, specialized weights outside the standard T2V/I2V path.

Common issues & troubleshooting

Combining this with context windows or other long-video nodes in the same graph. The node's own description warns against exactly this - it's a self-contained windowing system, not meant to compose with the pack's other chunking mechanisms. Pick one long-form approach per workflow.

Visible color drift over a long generation. This is the same saturation-creep problem documented elsewhere in the Wan extension ecosystem - try enabling colormatch before assuming it's specific to this node.

Seams between windows. Start by increasing motion_frame for more continuity context, and check drop_frames isn't set so low that duplicated boundary content is leaking through.