अपडेट
केस साक्ष्य: मूल ब्रेकडाउन रिपोर्ट पढ़ें · केस फ़्रेम और शॉट समय देखें
प्रकाशित टूल रिपोर्ट पर आधारित। निर्माण वर्गीकरण विश्लेषणात्मक निर्णय हैं; सुझाए चरण मूल प्रोजेक्ट या सत्यापित रीमेक नहीं हैं।
1. Ask per shot, not per video
Mixed production is the norm now. Our music-video case verdict reads: AI-generated video 53% + AI talking avatar 32% + still image with pan/zoom 13% + code/template rendering 3%; the dialogue case reads: code/template rendering 59% + AI-generated video 41%. A single “AI or real” label would be wrong for both videos. Split the video into shots at every cut first, then judge each one on its own evidence.
2. What stays still when it should not
In the music-video case, the same paper-and-inkstone still life appears twice. In shot #23 the review saw slight parallax between the paper, the inkstone and the window shadow, plus ink texture drifting between frames, and the report judged it AI video. In shot #32 the whole frame moved as one rigid transform with no local motion, so it was judged a still image with pan/zoom. Check frame edges, reflections, foliage, fabric and secondary objects: what moves, what stays frozen, and whether near and far layers shift against each other. A frozen region inside otherwise live motion deserves a closer look; a still photo pushed or panned shows flat parallax — everything slides as one plane.
3. Texture, noise and edges
Real sensors sprinkle grain everywhere, and it varies with the light: brighter areas carry more noise. Generations often look plastic-smooth, with uniform or missing grain, banding in smooth gradients (skies, skin), and shadows that are too clean. For faces — the low-end deepfake checklist — look at the jawline and hairline for blur or texture mismatch, teeth and ears for artifacts, and skin that ignores the scene’s light direction. Any one of these is a reason to check further, not a verdict.
4. Motion that morphs instead of moves
Generated motion interpolates between states, so watch identity across frames: objects that drift into different shapes, limbs that smear, patterns (text, tiles, instruments) that re-shuffle, loops that repeat a touch too perfectly, and physics without inertia or weight. Real filmed motion carries momentum, micro-jitter and background parallax that agree with each other. Slow the clip down and step frame by frame — morphing survives scrutiny worse than filming does.
5. Audio and sync clues
Voice is often the weakest link: TTS cadence that never varies, missing breaths, room tone that cuts in and out between words, lip sync that lags a few frames. In the music-video case the report judged the voice most likely TTS (64%), while the background music stayed unconfirmed — closer to library music (52%) than AI-generated (30%), even though the on-screen title card credits Suno for the music. A report that keeps that tension visible is more useful than one that picks a side. Each clue narrows the search; none is proof on its own.
6. Labels and metadata help — until they don’t
Platform AIGC labels are useful when present, and generation tools may leave metadata — but every re-upload and re-encode can strip it. Absence of metadata proves nothing, and a missing label is not evidence of human authorship. Treat labels as one input alongside the visual and audio signals above, never as the deciding vote.
7. How a breakdown keeps the evidence honest
The reason to write signals down per shot is auditability. A breakdown report keeps a timecode on every claim, separates observation from inference, and stores a confidence value next to each classification — the product-promo case, for example, labels shot #6 as code rendering at only 55% and lists it as an open question. The music-video verdict shares its mixture (53/32/13/3) instead of a single word. You can disagree with any row and re-check it against the frames, which is exactly what a yes/no answer prevents.
8. What no detector can promise
Every AI video detector makes errors, and compression eats exactly the evidence you need. Treat any percentage — from any tool, including ours — as an estimate that prioritizes what to verify, never as proof to publish. If you need certainty for a high-stakes decision, corroborate multiple signals and keep the reasoning visible. This website does not run an online checker; the open-source rea-video workflow runs on your machine, for free, and outputs the shot table with its evidence so you can re-check every row yourself.
अक्सर पूछे जाने वाले प्रश्न
Is there a free AI video detector I can use online?
Upload-based free checkers exist, but none is reliable enough to trust blindly, and uploading sends your video to someone else's server. The open-source rea-video runs locally for free and outputs the evidence per shot, so you can verify its reasoning yourself.
Can this website check if my video is AI-generated?
No. This site publishes guides and real case reports; it does not accept uploads. The analyzer runs on your own machine through Claude Code, Codex or the CLI.
What are signs of a low-end deepfake?
Blur or texture mismatch at the jawline and hairline, artifact teeth and ears, skin that ignores the scene's light direction, no blink or breath rhythm, and lip sync that lags a few frames. Each sign is a reason to check further, not a verdict.
Are AI video detectors accurate?
No detector is fully accurate. Compression and re-uploads erase exactly the evidence detectors rely on, so every percentage output should be treated as an estimate that prioritizes what to verify — never as proof.
Do platforms label AI-generated videos?
More platforms add AIGC labels every year, but coverage is incomplete and labels can disappear after re-uploads or re-encoding. A missing label is not evidence that a video was filmed.
पुन: प्रयोग योग्य विश्लेषण टेम्पलेट
Shot # / timecode Background: frozen region or live parallax? Noise: sensor grain, or smooth and banded? Face/edges: jaw, hairline, teeth, ears? Motion: moves with weight, or morphs? Audio: breaths, room tone, sync Label/metadata: present? (absence proves nothing) Confidence: what I estimate vs. what I can verify
असली रिपोर्ट के साथ अभ्यास
Practice on the music-video case: 118 shots across four production classes (AI video, AI talking avatar, still images with pan/zoom, code-rendered opening cards) and a 53/32/13/3 verdict. Compare shots #23 and #32, then pick any shot and run the checklist above against its frames.
केस खोलें और रिपोर्ट डाउनलोड करें →