Free video tools. No file uploads.
Free Video Analyzer
Check a video for timing patterns and sampled visual changes. Compare it with a reference when available and optionally run a local AI model on selected frames. Read what was tested and what remains unknown.
Beta 1.0.0-beta.1
Choose a video. A reference is optional.
Your video stays local. This analyzer reads files on your device without an analysis upload. It downloads its program code; it does not send the footage, name, hash or notes. No account is required.
Keep the original and inspect a working copy. Supported decoding depends on the actual codec and your browser.
Use a known earlier or original copy when available. The comparison does not establish that the reference is authentic.
Choose a range and sampling density
Begin with 24 samples across up to the first 60 seconds, or set a different range and density. The maximum is 120 samples in a 60-second range. A short change can fall between samples. Encoded track timing is checked separately.
Choose a video. A reference is optional.
Notes are your observations, not verified findings. They are included in locally saved reports.
Read the observations beside their limits
Start with the supplied-file hash, container records and codec information. The analysis cards then show timing irregularities, repeated sampled pixels, similar frames and substantial scene or color changes. Open the measurements to see the underlying timestamps and settings.
A repeated frame can come from a static scene. A scene change can be an ordinary cut, movement or flash. A color change can follow lighting, exposure, processing or a filter. These observations point to material you can inspect; they do not identify a proven edit.
No reference is required
A single-video check records facts available in that file and its sampled decoded frames. It does not manufacture an earlier version. Regular timing cannot exclude deleted frames or a speed change when an editor has rewritten timestamps.
With a reference, the tool adds ordered candidate frame matches, unmatched sampled runs and timing ratios between aligned content. Read ambiguity warnings before treating a match as useful. The supplied reference has not been authenticated by the tool.
Compare content and presentation timing
For reference comparison, choose corresponding ranges in both files. Matching uses reduced 32 by 32 pixel descriptors rather than every original pixel. Different framing, inserted material, overlays, reencoding and visual processing can change correspondence.
The reported relative presentation rate describes matched sample anchors. It may help you examine a possible speed difference, but sparse samples and repeated content can make the estimate ambiguous. Unmatched samples alone do not establish additions or removals.
Inspect a selected frame
Use the contact sheet to open a full supported decoded frame and inspect channels, contrast, edges or residuals. Difference views require equal dimensions and use the previous sample or a reference sample at the same relative range progress; they do not automatically register moving objects.
The views operate on browser-decoded 8-bit display pixels. They do not preserve original HDR precision or provide sensor PRNU. JPEG ELA and stored PNG hidden-color analysis are not treated as video-frame authenticity tests.
Choose an optional local model
The small generated-image model downloads about 10.1 MiB of weights. CIFAKE, face-image and three-label alternatives each download about 54.1 MiB. Shared runtime code adds about 13.6 MiB on first use. A larger model is not automatically more accurate.
Select one model, read its training task, confirm the limits and choose up to 12 frames. Only Run selected model locally fetches those weights. WASM CPU processing uses one thread. The face model applies the same confirmed square rectangle to all selected frames; check that the face stays inside it.
AI results describe individual frame experiments
These four models classify still images. They have not been evaluated by GDF as temporal video detectors. A brief edit can be missed between selected frames, and an ordinary image can receive a misleading label. No API key or hosted AI service is used.
Each frame retains its raw uncalibrated labels and logits, input and model hashes, crop, preprocessing and runtime information. Errors and cancellations are recorded separately. There is no probability of being fake, calibrated confidence or combined authenticity score.
Save a complete collected record
Use the free PDF, CSV, HTML and JSON buttons to save the observations available from this run. Reports retain the source hashes, track timing, sampling coverage, frame timestamps, applied views, reference correspondence, notes and AI attempts. A display search does not remove records from an export.
Reports contain text records rather than the footage or pictures of frames. Save a view PNG separately. CSV body values use reversible json: encoding; JSON preserves structure. Browser inspection dates rely on the device clock and are not independently verified.
No analysis upload or marketing transmission
The footage, decoded frames, filenames, hashes, crop coordinates and notes remain in the local workflow. This input surface does not load optional analytics or advertising. Website code, fonts and user-selected model files are downloaded without attaching the video data.
Clear closes the local inspection and removes displayed results. Previously saved files remain on your device. Review private details before sharing a report. Local processing helps protect the input from an analysis upload; it does not establish chain of custody or forensic certification.
Browser, format and sampling requirements
The parser recognizes MP4, MOV and WebM. Actual decoding requires browser support for the selected codec and profile. The controlled preview checks use desktop Chromium; other browsers, phones, codecs and devices are not universally validated. An unsupported decoder retains available container information and reports frame work as unavailable.
The bounds are 64 MiB per supplied file, four million coded and display pixels, a 60-second selected range, 2 to 120 requested samples and 12,000 encoded packet records. A local inspection stops at 120 seconds. An AI batch has a separate five-minute bound and 12-frame maximum.
Read incomplete and unavailable results carefully
Sampling leaves intervening frames and all material outside the selected range unexamined. Descriptor and thumbnail reductions are disclosed. Audio content, subtitles, every-frame continuity and temporal AI detection remain outside this beta.
For PDF, limits are 25,000 records, 4,194,304 represented UTF-16 code units, 500 pages, 25 MiB and 60 seconds. No shortened PDF is returned. Unsupported or directional characters are visibly escaped; JSON keeps original structured values. An unavailable result is unknown, not a no-edit finding.
Help test the beta
Version 1.0.0-beta.1 has bounded controlled function checks. A completed study of real-world edit detection, current-generator coverage and demographic performance is still absent. Use a working copy, retain source and handling context and check material findings independently.
Share beta feedback about your browser, version and experience. Keep private footage out of ordinary feedback. The contact link sends none of the selected video, decoded frames or report.
Continue with the right tool
Video Forensics provides a manual-first workflow. Photo Analyzer handles still images; Metadata Analyzer reads broader file properties. Inputs must be reselected on another page and are never carried in a link.
The local media parser is documented by Mediabunny. Browser frame decoding uses WebCodecs. Documentation and local execution do not certify an examination. Browse free tools.
Professional review
Have a video question this screening cannot resolve?
Tell an examiner what is disputed and which source or comparison files are available. Discuss the scope and a secure transfer method before sharing evidence. This inquiry includes none of your selected material.