Direct Tools from Mneme Labs

Turn hours of raw BMX footage into clean riding clips. A local Qwen vision model scans your video, finds every active riding segment, and cuts each one into a separate lossless clip — no standing, talking, or setup dead time. Fast recut mode re-cuts from a saved scan without reloading the model, and line-context mode pads wider (8s in / 4s out) to keep full trick lines intact. Configurable Qwen2-VL and Qwen2.5-VL models. Runs entirely on your own GPU — no subscription, no cloud, no per-clip cost.

A downloadable Python image-workflow source package with image_workspace.py, helper modules, requirements, setup instructions, and prompt examples. Intended for developers building image-generation, enhancement, background-removal, and batch workflows. Some operations require additional models or external services. This archive is source code, not a standalone compiled desktop app.