WanVideo Looper

Wan 2.2 has a 5-second wall. Its native context is 81 frames at 16fps, and one generation falls apart much past that. The community's workaround has always been the same: generate a clip, take its last frame, feed it back in as the start image for the next clip, and repeat - by hand, 40 minutes per 81-frame render on a 4090. WanVideo Looper is that loop, collapsed into one node. You give it a list of prompts, one per segment, and it chains them together with overlap and optional color correction so the result reads as one continuous take instead of a cut-up sequence.

It's built around Wan 2.2's two-expert MoE architecture, so it expects the split this family is famous for: a high-noise model for motion and structure, a low-noise model for detail. You feed both, plus the CLIP and VAE, and the node handles the two-pass sampling internally. The model_switch_point timestep (0.9 for I2V, 0.875 for T2V) decides where the high-noise pass hands off, and sigma_shift (8.0 default) applies to both models.

The inputs that actually matter

The five required model/image inputs are straightforward: model_high, model_low, clip, vae, and start_image (the seed image the first loop animates from). positive_prompts and negative_prompt come from the WanVideo Looper Prompts helper - the looper expects positive_prompts to be a list, one prompt per segment, which is exactly what that node outputs.

After that, the knobs a beginner should touch:

• duration_sec - length of each segment in seconds (5 is recommended; each segment is duration × 16 + 1 frames, so 5s is exactly 81 frames, Wan's native window).
• frame_merge - how many overlapping frames stitch each segment into the next. The last frame_merge frames of a segment seed the next one's start image. Start at 1-2.
• steps, cfg_noise, sampler_name, scheduler - normal sampling settings, applied per segment. cfg_motion_noise is the Wan-specific one: a separate CFG for the first step only, which nudges initial motion; set it to 1.0 to disable.
• seed - a single seed used for every segment, so the whole sequence is reproducible.

Color matching is where this node earns its keep. color_match_method offers Disabled, Reinhard (GPU), or MKL (GPU), and color_match_reference_frame picks the anchor: Sequential matches each segment to the previous segment's last frame (the drift fixer), while First/Last/Reference matches the whole video to a single frame. color_match_strength blends it in. One caveat: the color math runs on the GPU, but the node still refuses to run it without the color-matcher library installed - it just logs a warning and silently disables the feature.

Outputs are images (the full concatenated batch of frames, ready for a video save node) and last_frame (a single frame, handy for keeping the chain going). There's also an optional clip_vision input to guide the start image, and model_clip_sequence for the Lora Sequencer sibling.

Install

ComfyUI Manager is the easy path: search "WanVideoLooper" and install. Or the manual route:

cd ComfyUI/custom_nodes
git clone https://github.com/SquirrelRat/WanVideoLooper
cd WanVideoLooper
pip install color-matcher

Restart ComfyUI, then Add Node > WanVideoLooper. The pack itself ships no model files - the heavy lifting is the Wan 2.2 checkpoint you already need, running on ComfyUI's native Wan support (or Kijai's WanVideoWrapper).

Troubleshooting

• Color match does nothing. Check the console for the color-matcher warning. The node's requirements.txt lists color-match, but the code imports color_matcher - those are not the same package, so let Manager install what it wants and then run pip install color-matcher anyway.
• Test with dry run first. enable_dry_run skips sampling and decoding and just logs the setup and per-segment model choices. On a node that renders 20+ minutes per segment, that's the difference between a 5-second debug cycle and waiting an hour to find a wrong wire.
• It's slow, and speed LoRAs are a trap. Full-quality Wan 2.2 at 1280x720 runs about 25 minutes per clip on a 5090; chain five segments and you've booked your evening. The community consensus is that lightx2v/CausVid speed LoRAs degrade motion and detail - if you must, apply them to the low-noise pass only, and keep cfg_noise at 1.0.
• The node returns a black frame. positive_prompts isn't a list - the looper validates it and bails with a dummy tensor. Wire the output of the Prompts node, not a raw string.
• Dimensions snap down to multiples of 16 regardless of what you type, so set them in multiples of 16 and stop worrying.

This is a niche utility that shows up in real "MoE video" ComfyUI environments (next to WanVideoWrapper and WanVideoAutoResize), and it does one thing properly: it makes long Wan video a set-and-wait operation instead of a babysitting job. If you're chaining clips by hand today, you'll feel it on the first run.