Wan Smart Blend

Here's the problem WanSmartBlend solves, stated as the moment it shows up in your workflow: you want a longer Wan video than one clean generation gives you, so you feed the text encoder multiple prompts separated by | - "a red panda balancing on a bamboo stem | the camera pulls back to reveal the waterfall | a bird lands on the panda's head". On its own, Wan treats each prompt as its own segment and the cut between them lands like a jump cut. WanSmartBlend sits between the text encoder and the sampler and makes that cut a transition by interpolating the text embeddings across a band of frames instead of switching instantly.

It's part of the small ComfyUI-WanSeamlessFlow pack ("wanvideo - seamless flow") by fblissjr, which is built on top of Kijai's ComfyUI-WanVideoWrapper - the pack doesn't bring its own models, it works on the WANVIDEOTEXTEMBEDS objects the wrapper already produces. That's the first thing to know: install the wrapper first, this pack is a layer on top.

How it actually works

The node takes your bundle of per-prompt embeddings and does three things, in order:

1. Harmonizes shapes. If your prompts encoded to different token lengths, it pads or truncates them to a common shape (normalization: pad_truncate) so they can be interpolated at all.
2. Reorders the prompts so the semantic jumps are as small as possible. With optimize_order on it computes each embedding's centroid and does a greedy nearest-neighbor sort - prompt 3 goes after prompt 1 if they're more alike than prompt 2 is. If your |-list isn't already a sensible story order, this fixes it for you.
3. Tags the embedding dict with blend_width, blend_method, and the other parameters, so downstream sampling knows where and how hard to blend.

One honest caveat the README states plainly: this pack "needs modifications, for now, to Kijai's wanvideo wrapper" for the sampler to actually consume those blend tags during windowed generation. The node does the embedding-side preparation - shape matching, ordering, tagging - and the actual per-frame interpolation happens in the wrapper's context-window logic. If you're on a stock, unpatched wrapper, treat the output as well-ordered, ready-to-blend embeddings rather than expecting the blending to be fully wired.

The inputs that matter

• blend_width (default 8, 0–32) - width of the transition zone in frames. With 257 frames and 3 prompts you get ~85.6 frames per prompt, and the README recommends 8–16 frames of blend. Start at 8, widen if transitions feel abrupt.
• blend_method (default smooth) - the interpolation curve. linear is the flat version; smooth (smoothstep) gives a natural ease; ease_in/ease_out/sine/circ/bounce are increasingly aggressive accents. bounce is a bit of a novelty - don't use it on anything you care about.
• optimize_order (default true) - the nearest-neighbor reorder described above.
• normalization (default pad_truncate) - how mismatched embedding shapes get fixed. Leave it unless you know you want none.

Output is a single text_embeds (WANVIDEOTEXTEMBEDS) that wires straight into WanVideoSampler - and, if you want to check your work before sampling, into WanEmbeddingPrevizCanvas from the same pack.

Install

ComfyUI Manager, search for wanvideo - seamless flow (or "WanSeamlessFlow"), or:

cd ComfyUI/custom_nodes
git clone https://github.com/fblissjr/ComfyUI-WanSeamlessFlow

Restart ComfyUI. No requirements.txt, no model downloads - the pack is pure Python on top of the wrapper, and all the Wan models (T5, DiT, VAE) come from your existing WanVideoWrapper setup. An example workflow ships in example_workflows/wanvideo_seamlessflow_long_T2V_example_01.json.

Where people get burned

• Single prompt = no-op. With only one embedding in the bundle it logs "skipping optimization processing" and just passes the dict through. Check your WanVideoTextEncode actually has multiple |-separated prompts.
• Shape mismatch that survives harmonization disables blending entirely - it forces blend_width to 0 and logs it, so your video will have hard cuts and you'll need to check the console to know why.
• It's a young pack with zero community footprint. This is a "works on my machine, published for others" utility - read the logs, keep a backup of a working baseline workflow, and expect the rough edges.

For long-form Wan work where you're already chaining windows, this is the node that keeps the seams from showing. Just don't skip the previz step on the way out.