Free video tools. No file uploads.
Free Video Forensics Tool
Inspect video frames, color channels, residuals and sampled differences. Review container information, exact file hashes and track timing. Add a reference video when you have one, then save a local report.
Beta 1.0.0-beta.1
Choose a video. A reference is optional.
No video upload. The video, filename, hash, decoded frames and notes stay on your device. The website program files download to your browser. No account is needed.
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
Each selected range is limited to 60 seconds and 120 sampled frames. The default checks up to the first 60 seconds using 24 samples. Short edits between samples may be missed. Track packet timing is a separate bounded check.
Choose a video. A reference is optional.
Notes are your observations, not verified findings. They are included in locally saved reports.
Start with one video
Review the recognized container, selected video track, codec, dimensions, duration and original-file SHA-256. Decode a bounded series of samples with their requested and observed presentation timestamps. Inspect changes in color, texture, scene descriptors and repeated sampled content.
Packet timing can reveal irregular intervals or overlaps in the supplied track. Variable frame rate, recording interruptions and ordinary transcoding can produce similar patterns. Rewritten timestamps can conceal removed frames or a speed change. One video cannot reconstruct a missing original.
Add a reference when you have one
The tool compares ordered, reduced frame descriptors from two selected ranges. Candidate matches, ambiguity, unmatched samples and local timestamp ratios help you find sections to examine more closely.
An unmatched sample is not proof that content was added or removed. Sampling, cropping, overlays, motion, filters, compression and different camera views can prevent a match. A time ratio describes matched samples; it does not prove deliberate speeding or slowing. Check both videos and the surrounding context.
Examine a frame in detail
Select a sampled frame to decode its full supported display resolution. Compare the original with grayscale, RGB channels, edges, contrast, equalization, gamma, threshold, channel spread or a median residual. Save the selected view separately as PNG.
Difference views compare equal-sized frames with the previous sample or a reference sample at the same relative range position. They do not automatically register movement or prove that corresponding objects match. Compression, motion, lighting and misalignment can dominate the difference.
Video is different from a JPEG photograph
A decoded video frame does not retain the original JPEG compression structure. The workbench does not label a new JPEG recompression of a video frame as an authenticity test. Stored PNG hidden-color recovery also does not apply to ordinary decoded video frames.
Noise residuals show display differences, not sensor PRNU or a calibrated tampering result. The beta does not analyze audio content, reconstruct every frame, detect every object insertion or reverse a filter.
Optional AI stays on the device
Video Analyzer offers four pinned image classifiers for experiments on up to 12 selected frames. Choose a model and confirm its task before downloading its weights. The smallest model is about 10.1 MiB; the other three are about 54.1 MiB each. The shared runtime adds about 13.6 MiB on first use.
Models run locally on WASM CPU with one thread. No API key, external AI account or installed local server is required. The face model needs a user-confirmed face rectangle; movement can make that rectangle unsuitable for later frames.
Read each model result within its limits
These models were designed for still images. They are not temporal video detectors, and GDF has not established their video accuracy, current-generator coverage or demographic performance. They may miss an edit or flag an ordinary frame.
Raw model labels and logits are not probabilities, confidence scores or an authenticity verdict. Results remain separate by frame and model. No combined truth score is produced. The report retains source and model hashes, preprocessing, crop settings, runtime, timestamps, errors and cancelled attempts.
Your files and reports remain local
The selected videos, filenames, hashes, decoded frames, notes and report data are not sent to an analysis server. This input page does not load optional analytics, advertising or external AI. Program files, fonts and selected model weights download from the website.
Clear cancels active work and removes the displayed records and playback links. It does not delete your original or reports you have saved. Saved reports can contain private filenames, hashes and notes; review them before sharing.
Save the actual collected work
Download PDF, CSV, HTML or JSON with original-file hashes, container and track observations, sample timestamps, coverage, method settings, reference comparisons, notes and any model attempts. Display filtering does not shorten exports.
CSV uses reversible json: cells. PDF is selectable text; directional or unsupported characters are visibly represented. Reports contain text records rather than the video or frame pictures. Download a view PNG separately when needed.
Formats and resource requirements
This beta accepts recognized MP4, MOV and WebM containers when the selected codec can be decoded by the current browser. Container support does not guarantee codec support. Use a current desktop Chromium browser for the broadest tested path; other browsers and phones may lack a needed decoder or sufficient resources.
Each file is limited to 64 MiB and four million coded and displayed pixels. Each selected range is at most 60 seconds, with 2 to 120 sampled frames and at most 12,000 encoded packet timing records. The operation time limit is 120 seconds; optional AI is limited separately to 12 frames and five minutes per batch. Limits stop work with an explanation.
Coverage is recorded, not assumed
Descriptors and contact-sheet previews use explicitly recorded reduced dimensions. Selected inspection and AI frames use full supported decoded display pixels. Decoding and browser color conversion can alter the pixel representation. HDR, original high-bit-depth values and all-frame continuity are outside this beta.
PDF export stops above 25,000 records, 4,194,304 represented UTF-16 code units, 500 pages, 25 MiB or 60 seconds. It returns an explanation rather than a shortened PDF. Use JSON for original structured values or an over-limit PDF report.
Beta feedback and further examination
This working beta has controlled function checks, not a completed real-world video edit detection study. An unavailable or cancelled result is unknown; it is not a no-edit finding. Keep the original and collection context, work on copies and independently check observations that matter to a case.
Send beta feedback with your browser, tool version and the problem you encountered. Do not include private footage or evidence in an ordinary feedback message. This link attaches no selected file or report.
Related tools and methods
Video Analyzer provides the other video workflow. Reselect your videos there; inputs are not put in links. For still images, use Photo Forensics. Metadata Analyzer inspects a wider range of file properties.
Local container parsing uses Mediabunny, with browser decoding described in WebCodecs. These references do not certify this tool. Version 1.0.0-beta.1. All free tools.
Professional video examination
Does a disputed video need a fuller review?
Describe the question and the original or reference material available. GDF can discuss examination scope and a secure transfer method. Your selected video and report are not attached.