CLIPSIFTALPHA / 0.1.0
LOCAL REVIEW SYSTEM · WINDOWS
ClipSift samples representative frames and uses Gemma locally to help surface clips likely to contain a person—while keeping a human in control.
LOCAL PROCESSING / ORIGINALS UNTOUCHED / NO FACIAL RECOGNITION
SCANRESULTSSYSTEMCCTV clips / selected
INTERFACE CONCEPT · CONTROLS AND CLASSIFICATIONS ARE ILLUSTRATIVE
Classifications can be wrong. ClipSift is a review aid—not proof that a person is or is not present, and not a security decision system. Always review important footage yourself.
01 / DOWNLOAD
Download the complete Windows package, verify it, extract it to a local folder, then complete the separate Gemma setup before your first scan.
This build is unsigned, so Windows may warn. Download only from this page and verify the checksum.
ce3380b629086c00930928d0038e73f805d1a57f93794ebaafd95d6a1e6bd52502 / WORKFLOW
Select local CCTV clips and a separate output location.
Cover each readable timeline; motion can prioritise candidates, never prove a person.
Gemma observes selected frames. Application policy turns observations into review states.
Filter the queue, inspect evidence, and use CSV and benchmark metadata.
03 / CLASSIFICATIONS
ClipSift errs toward human review when an assessment is uncertain, malformed or low-confidence.
Gemma observed a likely visible person. Inspect the saved evidence and source clip.
The result is uncertain or unsuitable for a confident automated classification.
No person was observed with sufficient confidence. This is not proof that nobody appears.
04 / TWO WAYS TO WORK
05 / BEFORE YOU START
safe for lower-memory GPUs; balanced needs at least 12 GiB VRAMSEPARATE MODEL SETUPhf auth login.Gemma is downloaded separately through Hugging Face and is not inside the ZIP. The first scan downloads it to your normal per-user cache. CPU mode is available but may be extremely slow and memory-heavy.
06 / PRIVACY & SAFETY
LOCALCCTV footage is processed locally. Hugging Face is contacted for authentication and the separate model download—not to upload your media.
NON-DESTRUCTIVEOriginal footage is never modified, moved or deleted. Generated evidence, reports and clip copies go to your selected output folder.
HUMAN-LEDNo facial recognition and no identity or intent inference. ClipSift does not monitor live feeds or make security decisions.
07 / DEMONSTRATION
A demonstration using privacy-safe footage. Playback is embedded with YouTube's privacy-enhanced domain.
No. Video inference runs locally. Hugging Face network access is for authentication and downloading Gemma separately.
Gemma is a gated model under its own terms. Accept access and authenticate, then it downloads to your normal per-user cache.
No. There is no facial recognition, identity inference or intent inference. It assesses whether a person may be visible.
No. Originals are only read. ClipSift writes reports, evidence and flagged/review clip copies to the output folder.
CPU mode exists, but may be extremely slow and memory-heavy. A compatible NVIDIA CUDA GPU is strongly recommended.
Extract the entire ZIP and keep _internal beside ClipSift.exe. Confirm Gemma terms and hf auth login, available disk/VRAM, supported video types, and a writable output folder. Then run System Check. Reports are written to your selected output folder.