From Veo to Motion Transfer: Building Demo Videos for a BJJ Training App
说明与用途
使用 AI 生成 BJJ 训练演示视频时,基础动作无法实现,作者自行拍摄动作,利用 Wan 2.2 进行动作迁移,并通过后期处理得到可循环的高质量视频。
问题
Generating realistic demo videos for BJJ drills using AI failed for basic movements like hip escape because the model had never seen them.
结果
The app now delivers realistic, loopable demo videos for all BJJ drills, offline voice guidance, and a streamlined HIIT experience.
工作流
- Generate starting pose with Gemini‑3‑pro‑image
- Feed pose to Veo 3.1 to create 8‑second loop
- Convert to 960×540 H.264 with ffmpeg
- Film author performing movements
- Transfer motion with Wan 2.2
- Stabilize footage with OpenCV
- Grey background with MediaPipe
- Slow down, interpolate, speed up with ffmpeg
- Repair freeze‑frames with script
- Remove loop seam with crossfade
证据
It worked for push‑ups. It completely failed for the hip escape — the most basic movement in BJJ — because the model had simply never seen it.
I switched to motion transfer with Wan 2.2 (Wan2.2‑Animate on fal.ai): it reads the skeleton and motion from a source video and redraws it onto the reference athlete.
Raw transfer wasn’t shippable either. The post‑processing pipeline that made it work: Stabilization (OpenCV), Background greying (MediaPipe), Slow it down, then speed it back up (ffmpeg), Freeze‑frame repair, Loop seam removal.
- 记录日期: 2026-10-27T00:11:34.237Z