Wan 2.2 Video Extender PRO

Wan 2.2's native context is 81 frames, and that's about five seconds at 16fps. If you've been hand-chaining chunks in ComfyUI - sampling 81 frames, feeding the tail back into the next pass, doing the frame math wrong and getting a clip that comes back two frames short - you know exactly who this node is for. WanVideoExtenderNative (listed as "Wan 2.2 Video Extender PRO") folds that whole extend-and-stitch loop into a single node, and unlike a lot of the single-node chunkers floating around, it hands you real story control while it's at it.

It's the workhorse of the Granddyser/wan-video-extender pack, from the same author who publishes free Wan fine-tunes on HuggingFace. The companion model he recommends is Phr00t's WAN2.2-14B-Rapid-AllInOne. "AllInOne" matters here, because this node is built on VACE conditioning and the embedded workflow guide says it flat out: "Feel free to try any other model, but VACE integration is required!" Feed it a plain Wan checkpoint and it gets conditioning it can't use.

How it works

Each loop builds a VACE control latent: a gray canvas at empty_frame_level (default 0.5, mid-gray), with the last overlap_frames frames of your previous clip stamped in at the front as context, and anything connected to image_loop_X pasted in as a hard cut. Turn on use_endframe_loop_X and it also takes the next loop's image and pins it to the end of the clip as a target to animate toward - a cheap way to force a shot to land somewhere specific.

Then it samples at denoise 1.0, decodes, and here's the memory trick: each segment is written to disk (a temp dir prefixed wan_segments_) instead of staying in RAM, and models are unloaded between loops. Only the overlap stays resident; the author claims roughly 150–200MB during processing, and the final video is assembled from disk at the end. That's what makes "5-second clips" stop being a ceiling. Ten loops at 81 frames is close to a minute of footage without the machine grinding to a halt.

The settings that matter

Wire up model, vae, and clip as usual, then the numbers a beginner actually touches:

• extension_loops (1–10): how many chunks to generate. Start at 2 for tests.
• generate_frames (default 81): frames per loop. Just leave it.
• overlap_frames (4–64, default 16): how much context carries into the next loop. More overlap = smoother, shorter, slower. The README's sweet spot is 16–24.

The tested combo for the Rapid model is steps 4, sampler euler_a, scheduler beta, CFG 1.0. Don't blindly copy that onto an undistilled checkpoint - CFG 1.0 at 4 steps is a speed-LoRA thing, not a Wan default. strength (default 1.0) is the VACE strength: how hard the input frames are enforced against the generated ones.

Input modes are the nice part: video in for V2V, a single image for I2V, or nothing for T2V. One gotcha - default_width/default_height only apply to T2V. With an image input the node uses the image's own resolution, so resize your stills before feeding them in. Connect a reference_image and VACE keeps faces and characters consistent across loops, which is the single biggest quality win for long takes.

For storytelling, prompt_loop_1..10 each take a String node, lora_loop_1..10 are dropdowns over your loras folder, and image_loop_1..10 do scene cuts. Outputs are full_video (IMAGE) and used_prompts (STRING), so you can see which prompt actually ran per loop.

Install

ComfyUI Manager - search "Wan Video Extender" - or the manual route:

cd ComfyUI/custom_nodes
git clone https://github.com/Granddyser/wan-video-extender
pip install -r requirements.txt

Then restart. The requirements pull in diffusers, peft, gguf (so quantized Wan checkpoints work), opencv, and friends. Small trap: the README literally writes pip install requirements.txt, missing the -r - easy copy-paste stumble that fails on the first token.

Grounded gripes

The node is a skyscraper on your canvas - ten loops of per-loop inputs adds up to a lot of widgets, and the author owns it: a "luxury high-rise of functionality rather than a sprawling slum of spaghetti wires." Fine, but leave the loops you aren't using alone. And remember identity drift across chunk boundaries is physics of the 81-frame context - the reference image and overlap are a mitigation, not a cure. It's the node I'd reach for when a short Wan clip needs to become a scene.