Spaces:
Sleeping
AI Deepfake Analyzer Video Script
This script is designed for a polished concept + product walkthrough video. The primary version is about 5 minutes, long enough to explain the idea, show the app, walk through the report, and still sound human. A shorter social cut is included at the end.
Use safe demo media only. For the real-image control, use a consenting public-domain or self-owned non-sensitive image. For the AI-image control, use a clearly synthetic generated image. Do not demo with actual intimate or harassment content.
Video Goal
The video should make one thing clear: this app is not just another "AI detector score." It is a privacy-first evidence workflow for people who need to understand whether an image may be real, synthetic, or manipulated, and why the system reached that conclusion.
The tone should be serious but not gloomy. The viewer should feel that the project is technically thoughtful, victim-aware, and honest about uncertainty.
Recommended Structure
Target length: 4:30 to 5:30.
Recording format: screen recording with voiceover, optional webcam intro/outro.
Suggested flow:
- Open with the problem.
- Explain why a single detector score is not enough.
- Show the app landing page and architecture.
- Upload an AI-generated image and run analysis.
- Walk through the verdict and metrics.
- Walk through the explainable report.
- Download the PDF/JSON report.
- Explain privacy, limitations, and why inconclusive is a valid result.
- Close with demo link, GitHub link, and article.
Full Script
0:00 - 0:20: Hook
Visual: Start on the 5:2 article header image or the app landing page. Slowly zoom into the evidence pipeline or the analyzer hero section.
Voiceover:
"Deepfakes and AI-generated images are no longer just a future problem. Someone can create a fake image, post it online, and the person affected may have no easy way to understand what happened or how to explain it to someone else."
"Most tools answer with a single score. Something like, 82 percent AI. But if you are a real person dealing with harassment, that is not enough. You need evidence. You need limitations. You need to know why."
On-screen caption: "A deepfake analyzer should explain its evidence, not just output a score."
0:20 - 0:55: Problem and Product Idea
Visual:
Show Figure 1: 01-why-a-single-deepfake-score-is-not-enough.png.
Voiceover:
"That is why I built AI Deepfake Analyzer. The goal is simple: upload an image or submit a public URL, and the app gives you a layered authenticity report."
"It does not claim perfect truth. It does not say pixels alone can prove everything. Instead, it checks multiple evidence layers: model opinions, metadata, provenance, forensic signals, input quality, and detector disagreement."
"If the evidence is strong, it can return likely real, likely AI generated, or likely manipulated. If the evidence is weak, it can say inconclusive. That abstention is important, because a wrong confident accusation can hurt people too."
On-screen caption: "Labels: likely real, likely AI generated, likely manipulated, inconclusive."
0:55 - 1:35: Architecture Overview
Visual: Show Figure 2, then Figure 4, then Figure 5. Move slowly. Do not rush the diagrams.
Voiceover:
"At the product level, the app looks simple. The user uploads an image, confirms they have the right to submit it, and starts analysis."
"Behind that, the system creates an evidence packet. One image fans out into separate lanes. Metadata and provenance are checked. Hashes are calculated. Compression, noise, frequency, and regional anomalies are inspected. Then pretrained open-source detector models give their opinions."
"The model architecture is an application-level mixture of experts. I am not training a huge custom model here. Instead, multiple pretrained detectors and non-model experts produce opinions. A safety arbiter combines them, applies reliability weights, checks disagreement, and decides whether the evidence is strong enough."
"The key idea is that a model score is treated as an opinion, not as the final verdict."
On-screen caption: "Model output is one evidence layer, not the whole decision."
1:35 - 2:20: App Walkthrough - Upload and Analyze
Visual: Open the live app or local app. Show the landing page briefly, then click "Analyze" or scroll to the analyzer. Use a safe AI-generated image for this demo.
Voiceover:
"Now let us walk through the app."
"The first screen explains the privacy-first posture and the architecture. For the actual analysis, I can upload a file or submit a public URL. For this demo, I am using a safe AI-generated test image."
"The preview is blurred by default because this tool is designed for sensitive media. Before analysis, the user has to confirm they have the right to submit the media."
"Once I click analyze, the backend validates the file, prepares the temporary media packet, runs the evidence layers, and returns the result."
Screen actions: Open app. Scroll to analyzer. Choose upload. Select AI-generated test image. Show blurred preview. Check consent. Click Analyze.
On-screen caption: "Sensitive previews are blurred by default."
2:20 - 2:55: Verdict Walkthrough
Visual: Show the result card: final label, AI probability, manipulation probability, detector disagreement, confidence band.
Voiceover:
"The first thing the user sees is the victim-friendly summary. It shows the final label, AI probability, manipulation probability, detector disagreement, and confidence."
"But the important part is that the report does not stop here. A score alone is not enough. So below the verdict, the app explains the decision."
"For example, if the image is likely AI generated, I want to know what drove that conclusion. Was it model consensus? Metadata? Forensic evidence? Or was the system uncertain?"
On-screen caption: "The score is the start of the report, not the end."
2:55 - 3:55: Report Walkthrough
Visual: Scroll through the report sections in this order: Decision Summary, Mixture of Experts, Evidence Layers, Explainability, Region Evidence Map, Analytical Layer Ledger, Technical Appendix.
Voiceover:
"The Decision Summary translates the analysis into plain language. It separates the primary drivers, counter-evidence, uncertainty factors, and what would help improve confidence."
"The Mixture of Experts section shows how different parts of the system voted. There is a visual detector ensemble, a portrait specialist when relevant, metadata and provenance, forensic residuals, input quality, and finally the safety arbiter."
"The Evidence Layers section is where the app becomes more explainable. Instead of saying 'trust the model,' it shows what each layer checked: input source, metadata, provenance, hashes, compression and noise, and detector outputs."
"The Explainability section shows the model consensus and calibration gate. This is where disagreement matters. If models fight each other, or if the image is too compressed or low-quality, the app should lower confidence or return inconclusive."
"The Regional Evidence Map is an abstract 4-by-4 map. It does not show the uploaded image. It only shows which regions had stronger anomaly signals relative to the rest of the image. That keeps the report useful without exposing sensitive media again."
"Finally, the Technical Appendix gives hashes, detector names, detector labels, weights, and reproducibility notes. This is for reviewers, moderators, or technical helpers who need to understand how the result was produced."
On-screen captions: "Decision Summary: why the verdict happened." "Mixture of Experts: model and non-model opinions." "Regional Map: anomaly evidence without exposing the raw image." "Technical Appendix: hashes, detector outputs, reproducibility."
3:55 - 4:25: PDF and JSON Report Export
Visual: Click JSON, show file download briefly. Click PDF and open the PDF report if convenient. If opening the PDF slows the recording, show the report preview section or a previously opened PDF.
Voiceover:
"The report can also be exported."
"JSON is useful for developers, researchers, and structured review workflows. PDF is useful when a person needs to preserve the analysis or share it with a trusted helper or platform moderation team."
"The PDF is not a legal certificate. It is an evidence summary. That distinction matters. The report says what the app checked, what it found, what it could not know, and what could improve confidence."
On-screen caption: "PDF = shareable evidence summary. JSON = structured technical export."
4:25 - 4:55: Privacy and Safety Boundaries
Visual:
Show Figure 3: 03-the-input-boundary.png, then return to the app and show Delete button.
Voiceover:
"Privacy is a core part of the design. The app supports uploads and public URLs, but it does not do face search, doxxing, login scraping, or private identity inference."
"Public URL context is limited to visible public information. Raw media is not returned by the API. Sensitive previews are blurred. Temporary media is short-lived. And the user can delete an analysis early."
"This matters because the target user may already be dealing with harassment. A defense tool should not create a new privacy risk."
On-screen caption: "No face search. No doxxing. Public links only. Short media retention."
4:55 - 5:25: Limitations and Honesty
Visual:
Show Figure 7: 07-the-verdict-policy.png.
Voiceover:
"The honest part is this: no deepfake detector can guarantee perfect accuracy on every image. Screenshots, crops, heavy compression, filters, and new generator families can all confuse detectors."
"That is why this app is designed around calibrated evidence. It can say likely real, likely AI generated, likely manipulated, or inconclusive. Inconclusive is not a failure. Sometimes it is the safest answer."
"The next step is larger calibration sets, more model diversity, stronger provenance tooling, and eventually video analysis with frame sampling and timeline-level reporting."
On-screen caption: "The app does not promise certainty. It explains evidence."
5:25 - 5:45: Closing
Visual: Show the final landing page, GitHub repo, live demo URL, and article header.
Voiceover:
"So that is AI Deepfake Analyzer: a privacy-first, explainable authenticity workflow for sensitive media abuse."
"The live demo, GitHub repo, and full technical article are available. The main idea is simple: when the stakes are human, a black-box score is not enough. The system should show its work."
On-screen caption: "AI Deepfake Analyzer: explainable evidence, not blind scoring."
Report Walkthrough Talking Points
Use this section if you want a slower separate video focused only on the report.
Start with:
"The report is the most important part of this app. The detector score is only one signal. The report explains the whole decision."
When showing the verdict:
"This section gives the top-level label and confidence. I treat this as triage, not a final forensic certificate."
When showing AI probability:
"AI probability estimates how strongly the evidence leans synthetic. It should be read with the confidence band and detector disagreement."
When showing manipulation probability:
"Manipulation probability is separate because an image can be real but edited, composited, or recompressed in suspicious ways."
When showing detector disagreement:
"Disagreement is important. If the model panel does not agree, the app should not pretend the answer is obvious."
When showing Decision Summary:
"This is the plain-language explanation. It separates what drove the result, what pushed against it, and what evidence would improve confidence."
When showing Mixture of Experts:
"Each expert gives an opinion. Some are model-based, some are not. The arbiter combines them rather than blindly trusting one detector."
When showing Evidence Layers:
"This is the layer-by-layer analysis. Metadata, provenance, hashes, compression, noise, and model evidence are kept separate so the user can inspect them."
When showing Regional Evidence Map:
"This map is abstract. It does not expose the submitted image, but it shows whether some regions had stronger anomaly signals than others."
When showing Technical Appendix:
"The appendix is for technical review. It contains hashes, detector outputs, model labels, weights, runtime, and reproducibility notes."
When showing download buttons:
"The PDF is for sharing with a person or platform. The JSON is for structured review, debugging, or research workflows."
Screen Recording Checklist
Before recording, pre-warm the app by running one analysis so model downloads do not happen on camera.
Keep these tabs ready:
App: https://devraj1990-deepfake-moe.hf.space
GitHub: https://github.com/Devrajsinh-Jhala/DeepFake-MoE
Article: docs/deepfake-analyzer-final-article.md
Assets folder: docs/social-diagrams
Use these visual assets in order:
article-header-5x2.png01-why-a-single-deepfake-score-is-not-enough.png02-the-product-in-one-flow.png03-the-input-boundary.png04-the-evidence-bus.png05-the-mixture-of-experts-design.png06-the-report-is-the-product.png07-the-verdict-policy.png08-deployment-architecture.png
Recording tips:
Use browser zoom around 90 to 100 percent.
Use a clean desktop and hide personal bookmarks.
Use safe demo images only.
Blur or crop anything personal.
Pause for 1 to 2 seconds before clicking important buttons so the viewer can follow.
Do not overclaim accuracy. Say "evidence," "signals," and "confidence" more than "proof."
90-Second Social Cut
This version is for X, LinkedIn, Instagram, or a short product demo.
Visual: Start with header, then quick app walkthrough, then result/report.
Voiceover:
"I built an AI Deepfake Analyzer for a problem that is becoming very real: fake or manipulated images being used for harassment."
"Most tools give you a single AI probability. But for a victim, a black-box score is not enough. You need to know what was checked and why the system reached that conclusion."
"This app accepts an upload or a public URL. It checks metadata, provenance, hashes, compression, noise, regional anomaly signals, and a mixture of pretrained open-source detector models."
"The final verdict can be likely real, likely AI generated, likely manipulated, or inconclusive. That last label matters, because if the evidence is weak, the system should not force a confident accusation."
"Here is the walkthrough. I upload a safe AI-generated test image, confirm consent, and run analysis. The app returns a verdict, confidence, AI probability, manipulation probability, and detector disagreement."
"Then the report explains the result: decision summary, mixture of experts, evidence layers, regional anomaly map, and technical appendix."
"The report can be exported as PDF for sharing or JSON for technical review."
"The goal is not to promise perfect detection. The goal is to make deepfake analysis more explainable, privacy-first, and useful for people dealing with synthetic media abuse."
"Live demo and GitHub are linked."
30-Second Hook Cut
"Deepfake detection should not be a black box."
"I built an AI Deepfake Analyzer that does more than return one AI score. It checks model opinions, metadata, provenance, forensic layers, input quality, and detector disagreement."
"Then it gives a victim-friendly report with a technical appendix, PDF export, JSON export, and clear uncertainty."
"The app can say likely real, likely AI generated, likely manipulated, or inconclusive."
"Because when the stakes are human, the system should show its work."
Suggested Title Options
"I Built an Explainable AI Deepfake Analyzer"
"A Deepfake Detector That Shows Its Evidence"
"Building a Privacy-First AI Deepfake Analyzer"
"Why Deepfake Detection Needs More Than One Score"
Suggested Opening Caption
"I built a privacy-first AI Deepfake Analyzer for synthetic media abuse. It does not just return an AI score. It explains the evidence layer by layer."
Suggested Closing Caption
"This is not a claim of perfect detection. It is an explainable evidence workflow: model opinions, metadata, forensics, provenance, uncertainty, and a report people can actually use."