JaguarScript

Vexora: your moves, choreographed in light Glowing dance edits are everywhere on social media, but each one takes hours of manual rotoscoping and compositing. Vexora does it automatically. You upload a dance video, pick a style, and get back a video with light effects attached to the real person's body: fire trails, water, lightning, auras, sparks and face lines. The original footage is never redrawn. How it works Tracking. MediaPipe follows pose, hands and face on every frame, and filtering keeps the points stable. Deep analysis with Meta Muse Spark. Muse Spark watches the whole video like an animator. It reads the intent of each move and the story of the piece, then writes a choreography: key moments, each with a color and an intention. Muse Spark doesn't hear audio, so Muse Voice Transcribe gives it the lyrics with timestamps, and the effects can follow the words of the song. Precision. The model decides what happens and why; our tracking decides exactly when. Each AI moment snaps to the real gesture (within 0.3 s) or to the music beat (within 0.15 s). With several dancers, the model says who each moment is for. Rendering. A procedural engine draws HDR transparent layers and blends them over the untouched video. Safe and affordable by design. The model never generates pixels. Its answer is checked against our effect catalog, and anything unknown is rejected, which limits prompt injection through the video or audio. Effects the model invents but we don't have yet are logged as proposals for future work, and the closest existing effect is used meanwhile. Responses are cached, and daily quotas cap the cost. The product. Users sign in with Google, upload clips of up to 3 minutes and choose from styles such as Avatar, Fire, Water and Lightning. They can choose a fast economic analysis or the deep one with Muse Spark, see the AI's intention for each moment in a choreography panel, edit the effects, create new versions and share them. It is built with Python, FastAPI, React, Firebase and Google Cloud Run. What we learned. Multimodal models are excellent at meaning (story, intent, lyrics) and weak at frame-accurate timing. Vexora combines both: Muse Spark's understanding with our own motion tracking.