The first-and-last-frame trick is a smart way to make AI video loop perfectly

A text prompt goes in, a looping animated WebP with real transparency comes out. That is the entire proposition of this Claude plugin, and the mechanism behind it is genuinely clever.

The core trick is how it closes the loop. Most attempts at looping video either cut abruptly or fade awkwardly. Here, the same generated still is fed to the video model as both the first and the last frame. The model has to travel from that image back to itself, which means the motion naturally returns to its starting point. No seam, no crossfade hack.

Transparency is handled with equal care. Rather than running a generic background-removal pass on whatever colours the model happened to produce, the tool picks the background colour itself before generation. That means when it solves for the original foreground colours during matting, it is working from a known quantity rather than guessing. Soft edges stay soft. The halo problem that plagues naive alpha extraction is avoided by construction.

The toolchain is GPT Image for the still, Seedance for the video, and a matting model that downloads on first run at roughly 180MB. Everything routes through a single OpenRouter key, which is a practical convenience — one credential, one billing surface. The only external dependency you need to install yourself is ffmpeg.

The two-stage approval flow is worth noting. The tool generates the still, shows it, and stops. You approve before it spends anything on video generation. Then it proposes the motion and stops again. For a workflow that involves two model calls of meaningfully different cost, that is the right default behaviour.

What you get at the end is an animated WebP with genuine alpha channel — not a GIF, not a video with a white box behind it. That matters for UI work, where dropping an icon onto a dark background or a gradient should just work.

The scope is deliberately narrow: one icon, one prompt, one output format. There is no batch mode, no style library, no configuration surface beyond the prompt itself. Whether that is a limitation or a feature depends entirely on what you need. For someone who wants a polished animated icon without building a multi-step image-to-video workflow from scratch, the narrowness is the point.