Wan 2.2 Text-to-Video Seamless Loop Workflow for ComfyUI

Our image-to-video loop workflow animates a photo and loops back to it. But often there is no photo — you want a looping scene that exists only in your head: a neon cyberpunk alley in the rain, a magical portal, an abstract gradient background for a stream overlay. That's what this workflow does: describe the scene in plain text and it generates a video that loops back onto itself with no visible seam.

The loop problem is the same as with I2V: text-to-video models have no reason to end where they started, so the last frame never matches the first and the loop "pops". The fix is the same two-pass VACE technique — but with one extra trick, because there's no input image to anchor to.

1. Base generation (Wan 2.2 T2V): your prompt drives a dual-expert text-to-video pass — the high-noise expert lays out motion, the low-noise expert refines detail. Accelerated by SageAttention and fp16 accumulation patching.
2. First-frame pinning: since there's no input photo to return to, the workflow extracts the clip's own first frame and uses it as the static anchor plate. The loop will close onto this frame.
3. The bridge (Wan 2.1 VACE): the tail and head frames of the clip are stacked as VACE reference context, and a bridge segment is generated that flows from the clip's ending motion back onto the anchored first frame. Florence2 auto-captions the seam frame so the bridge prompt always matches what's actually on screen.
4. Optional mask (limit the motion): a built-in mask stage can composite the animation over a frozen copy of the first frame, so only the area you paint moves — perfect for cinemagraph-style loops. A single Use Mask toggle switches it off to animate the whole frame.
5. Pick the length: set the loop duration in seconds and the workflow scales the frame count automatically — 5s, 10s or longer from the same graph.
6. Upscale, polish & export: SeedVR2 upscale to 1600p, RIFE interpolation to 32 fps, then H.264 export via VHS Video Combine. The result plays forever with no visible loop point.

The workflow ships with a ready-to-run example prompt (a looping fireplace scene) — replace it with your own. Best results come from scenes with cyclic motion: fire, water, steam, rain, snow, flowing fabric, drifting fog, pulsing light. Add "static camera, no camera movement" to your prompt; camera drift is the hardest thing for any loop to close.

Custom node packs (all installable via ComfyUI Manager): ComfyUI-VideoHelperSuite, KJNodes, rgthree, ComfyUI-Easy-Use, Florence2, RIFE, LayerStyle, ComfyUI-GGUF.

Do I need an input image?

No — you only write a text prompt. The workflow generates the scene and closes the loop onto its own first frame.

How is this different from the image-to-video loop workflow?

The I2V version animates a photo you provide and returns to it. This version creates the entire scene from text. The VACE loop-closing pass is the same proven technique in both.

What prompts loop best?

Scenes with cyclic motion: fire, waterfalls, rain, steam, snowfall, flowing fabric, drifting fog, pulsing neon. Keep the camera static — add "static camera, no camera movement" to the prompt.

What resolution, frame rate and length does it output?

Up to 1600p at 32 fps (SeedVR2 upscale + RIFE interpolation), and the length is yours — set the loop duration in seconds and the workflow scales the frame count automatically for 5s, 10s or longer seamless loops.

Can I animate only part of the scene?

Yes — an optional mask composites the animation over a frozen copy of the generated first frame, so only the masked area moves and the rest stays still (great for cinemagraphs). A single Use Mask toggle turns it off to animate the entire frame.

How much VRAM do I need?

16 GB is comfortable (fp8 T2V experts + Q8 GGUF VACE). 12 GB works at reduced resolution.

Can I use the videos commercially?

Yes — your generated videos are yours, including commercial use. Only the workflow file itself may not be resold or redistributed.