Wan 2.2 + VACE Seamless Loop Video Workflow for ComfyUI

Looping videos are everywhere — live wallpapers, stream overlays, product ads, lo-fi backgrounds, NFT animations, digital signage. They also sell: a clip that loops invisibly reads as "professional" in a way a normal AI clip never does.

The problem: image-to-video models like Wan 2.2 only condition on the first frame. The model knows where the video starts but has no idea (and no incentive) to land back there. The last frame always drifts — lighting shifts, the subject moves, the camera floats — so when the player jumps back to frame one you get a visible "pop".

Crossfading the ends in an editor doesn't fix it; it just smears the pop over 10 frames. The real fix is to make the model generate a transition that ends exactly on the first frame. That's what VACE (Video All-in-one Creation and Editing) makes possible — it accepts reference frames as hard constraints, so you can demand: "start from the clip's last frames, end on its first frames."

1. Pass 1 — base motion (Wan 2.2 I2V): a normal image-to-video generation using Wan 2.2's dual-expert setup — the high-noise expert blocks out motion, the low-noise expert refines detail. This produces your main clip.
2. Pass 2 — the bridge (Wan 2.1 VACE): the workflow extracts the tail frames of the clip and its opening frames, reverses and batches them as VACE reference context, then generates a short bridge segment that starts where the clip ended and lands precisely on frame one.
3. Composite: base clip + bridge are concatenated with the overlap trimmed and masked, so there is no duplicated frame and no seam.
4. Polish: RIFE interpolation doubles the frame rate to 32 fps and the result is muxed by VHS Video Combine. The loop point becomes mathematically invisible — the last frame's successor is the first frame.

The trick that makes the bridge look natural rather than "morphy" is feeding VACE several context frames from each end (not just two stills) — it learns the motion direction in and out of the seam, so momentum carries through the loop point.

This is the exact pipeline of the workflow you can download below — the same file we deliver on Fiverr:

The download includes a notes panel inside the workflow with the exact filenames and where each model goes (diffusion_models, loras, text_encoders, vae, clip_vision). Custom node packs used: ComfyUI-VideoHelperSuite, KJNodes, rgthree, ComfyUI-Easy-Use, Florence2, RIFE, LayerStyle — all installable via ComfyUI Manager.

What exactly do I get for $2.99?

The complete 70-node workflow JSON — the same file we deliver to Fiverr clients — with embedded setup notes. Import it into ComfyUI via drag & drop.

How much VRAM do I need?

16 GB is comfortable (VACE runs as Q8 GGUF, Wan 2.2 experts as fp8). 12 GB works at lower resolution/frame counts.

Why doesn't my normal Wan 2.2 video loop?

I2V only conditions on the first frame, so the final frame always drifts away from it. The VACE bridge pass is what forces the video to land back on frame one.

Do I need to write prompts?

No — Florence2 captions your image automatically and the workflow assembles the motion prompt. You can still override it.

Can I sell videos I make with it?

Yes, your generated videos are yours, including commercial use. Only reselling/redistributing the workflow file itself is not allowed.

I'd rather not set up ComfyUI — can you generate loops for me?

Yes — try free Wan 2.2 image-to-video on JollyAI (5 free generations daily, no signup), or order a custom loop from us directly.