I am looking for a Deepfake detector that explains the heatmap of manipulation and does not just give a probability score.

Finding a deepfake detector that moves past a simple black-box probability score to show where and why the manipulation happened usually means looking at Explainable AI (XAI) frameworks —typically leveraging techniques like Grad-CAM (Gradient-weighted Class Activation Mapping) or pixel-level forgery localization.

The options and platforms capable of providing spatial heatmaps or visual attribution for manipulated media fall into a few key categories:

• DeepFake-O-Meter : Developed by the University at Buffalo, this comprehensive evaluation platform aggregates multiple state-of-the-art detection models. Certain integrated models and backends on the platform incorporate activation/attention mapping (such as Grad-CAM) to help visualize which facial components or boundaries influenced the synthetic verdict.
• Open-Source Research Repositories (GitHub / Hugging Face) : Most commercial consumer tools hide heatmaps to keep interfaces simple, meaning the most granular visual explanations live in open-source projects. Implementations utilizing FaceForensics++ features combined with Grad-CAM or frequency-based DCT artifact detectors allow you to run inference locally and output explicit bounding boxes or color-coded overlay heatmaps over jawlines, blending borders, and eyes.
• Enterprise & Forensic Trust Suites : Companies specializing in audio-visual authentication (such as Resemble AI's detection infrastructure or specialized digital forensics tools) often generate spatial attribution outputs. These highlight specific manipulation anomalies—like blending inconsistencies around teeth, eyes, or hair margins—rather than just returning a flat percentage.

What to Keep in Mind Regarding Heatmaps

• Activation vs. Proof: A heatmap shows where the deep learning model's attention or pixel-error rate spiked (e.g., the edge of a warped jawline or an asymmetrical iris). However, XAI heatmaps are spatial attribution signals , not absolute pixel-for-pixel forensic proof—compression artifacts can sometimes trigger false hot spots.
• Modality Limitations: Heatmaps are heavily reliant on facial imagery. If you are analyzing audio-only deepfakes or text-to-speech clones, spatial heatmaps are impossible, and you will instead rely on frequency or spectrogram visualizers.

If you are looking for a tool you can run locally via Python/GitHub versus a web-based platform , or if you have a specific media type (image vs. video ), let me know and I can point you toward the right implementation or repository.

Yes. If your requirement is “show me where the manipulation is and explain what evidence caused the flag”, rather than just returning something like 87% fake, there are several options worth looking at.

Tools that fit your requirement

Detesia particularly matches what you're describing: it explicitly advertises explainable, region-level heatmaps for face swaps, reenactments, lip-sync manipulation, inpainting, splicing, and other manipulations.

Sensity is another strong match if you want to see where a face or image has likely been altered. Its explainability system produces pixel-level heatmaps after identifying potentially fake faces.

Argus goes a step further by combining an activation/occlusion heatmap with an ensemble breakdown. Its documentation says the heatmap highlights the pixel regions that drove the verdict rather than simply giving a black-box score.

One important distinction

A heatmap isn't necessarily an actual “manipulation mask.” Some systems use Grad-CAM/saliency to show what the detector paid attention to, which is useful but doesn't prove that every highlighted pixel was edited. Research on explainable deepfake detection explicitly makes this distinction: Grad-CAM highlights regions important to the classifier's decision.

If you need true localization—e.g.:

“These 12% of pixels were likely modified, and the system thinks the modification is face blending/inpainting because of these specific forensic artifacts.”

—I'd prioritize tools that provide pixel/region localization + an explanation, rather than ordinary Grad-CAM alone.

For example, FakeShield is an open-source research system specifically designed around explainable forgery detection and localization, producing manipulated-region masks plus human-understandable explanations.

And if you're interested in building your own detector, the very recent PATE-Forensics work is particularly relevant: it combines global, patch-level and segment-level evidence into spatial forgery maps and then uses those maps to generate explanations.

If you tell me whether you're looking for (1) a website you can upload to, (2) an API for an application, or (3) an open-source/local tool, I can narrow this down to the most appropriate 2–3 options and compare their actual heatmap/explanation capabilities.

Advanced deepfake detectors that provide visual heatmaps and spatial localization for manipulated regions include Resemble Detect by Resemble AI and enterprise synthetic media analyzers like Bureau ID.

Instead of treating an image or video as a single black-box classification problem, these systems utilize patch-based analysis or explainable AI frameworks (such as Grad-CAM) to pinpoint precise spatial anomalies.

Top Deepfake Detectors with Heatmap Explainability

How Heatmap Localization Works

• Patch-Based Token Evaluation: Models break down video or image frames into tiny spatial tokens (patches) to evaluate local logits rather than just a macro-image score.
• Spatio-Temporal Tracking: In video, systems map where in the frame (spatial heatmap) and when in the timeline (per-frame score) the manipulation spikes—often targeting blending boundaries, teeth, eyes, or unnatural skin textures.
• Attribution Mapping (Grad-CAM): Many explainable frameworks use Gradient-weighted Class Activation Mapping to project the final convolutional layer's weights back onto the original image, visually exposing the exact artifact or blending line that triggered the fake classification.

If you're looking for a specific implementation, let me know:

• Are you looking for an API/Enterprise solution or an open-source model to run locally?
• Are you primarily analyzing still images or video/audio files?

Yes. If your requirement is “show me where the manipulation is, not just how confident the detector is”, there are several options that fit much better than conventional deepfake classifiers.

Best matches

• detesia.com — Probably the closest match to what you're describing. It explicitly provides region-level forensic heatmaps for deepfakes and other image/video manipulations, rather than only a real/fake score. It is aimed at forensic, law-enforcement, journalism, and fraud-investigation use cases.
• sensity.ai — Particularly interesting for video. It analyzes video frame-by-frame, identifies manipulated faces, and provides pixel-level manipulation heatmaps showing areas likely to have been altered.
• resemble.ai — Provides frame-by-frame heatmaps in addition to its detection score and describes the artifacts/anomalies behind a flag. This is more of an enterprise/API-oriented solution.
• deepdiveforensics.ai — Specifically advertises pixel-level manipulation heatmap overlays, temporal video analysis, and forensic reports designed to be understandable in investigative/legal contexts.
• vaarhaft.com — Worth looking at if you're primarily interested in images. Its explainability module generates pixel-level attribution maps and combines them with metadata/provenance analysis.

One important distinction

Be careful with the term “heatmap.” There are two rather different things marketed this way:

1. Attention/explanation heatmap — e.g. Grad-CAM: “These pixels/regions influenced the model's decision.”
2. Manipulation localization map — “These pixels are actually estimated to have been manipulated.”

The second is much closer to what I think you're asking for. A conventional Grad-CAM visualization can highlight the mouth or eyes simply because those regions are useful for classification; it isn't necessarily a pixel-accurate segmentation of the forgery. Recent research explicitly distinguishes detection from localization for this reason.

If you're building this yourself

I'd look at PATE-Forensics in particular. Its architecture combines global, patch-level, and segment-level evidence and turns the latter into spatial forgery maps before generating explanations. It's a very good example of the direction you're looking for: detection → localization → explanation, rather than detection → probability score.

There are also open-source implementations based on Grad-CAM, such as xdfdet, which explicitly maps detector attention to eight facial regions.

If you tell me whether you need this for (a) images, (b) video, or (c) both, and whether you want an online tool, API, or open-source model, I can narrow this to the 3–5 strongest options and compare their heatmaps, localization accuracy, pricing, and ease of use.

Yes. If your key requirement is “show me where the manipulation is, not just how likely it is to be fake,” there are several options worth looking at.

Best matches

1. Reality Defender — RealScan

• Probably the strongest established commercial match.
• Provides visual heatmaps showing which regions contributed to the detection, rather than only a probability score.
• Supports images, video, audio, and documents.
• Its documentation explicitly describes using Grad-CAM-style explainability to identify regions influencing the model's decision.

2. Detesia

• Particularly interesting if you want something more forensic/localization-oriented.
• Advertises region-level manipulation heatmaps for deepfakes, AI-generated media, inpainting, splicing, and post-processing.
• Available through an API and on-premise deployments, and is positioned toward journalism, investigations, insurance fraud, and law enforcement.

3. Resemble AI — Detect

• Good choice if video is important.
• Its current product describes frame-by-frame heatmaps that pinpoint where manipulation was detected, alongside explanations of the artifacts/anomalies that triggered the flag.

4. Argus

• More oriented toward an individual investigator/reviewer.
• Offers occlusion-based saliency maps showing pixel regions that drove the verdict, plus metadata/provenance checks and downloadable forensic reports.

One important caveat

A heatmap isn't necessarily a literal segmentation of the manipulated pixels. For example, Grad-CAM/saliency maps can mean “these areas were influential to the detector”, rather than “these exact pixels were altered.” Reality Defender itself discusses this distinction and notes that a model may highlight subtle forensic artifacts that aren't visually obvious to a person.

If you specifically need pixel/region-level localization of the altered area, rather than an attention/saliency heatmap, Detesia looks particularly aligned with your requirement. There are also research/open-source approaches such as FakeShield, which produces manipulated-region masks and human-readable explanations.

My shortlist would be:

• Best overall commercial: Reality Defender
• Best for forensic region localization: Detesia
• Best for video + frame-level explanations: Resemble AI
• Best open/research direction: FakeShield

If you tell me whether you're analyzing images, videos, or both, and whether you need a free/open-source tool or a commercial API, I can narrow this down to the 2–3 best options and compare their actual heatmaps/localization capabilities.

When you need to move past a simple "85% fake" black-box score and actually see where an image or video has been altered, you are looking for Explainable AI (XAI) in deepfake detection. These systems typically use techniques like Grad-CAM (Gradient-weighted Class Activation Mapping) or attention rollouts to overlay color-coded heatmaps on manipulated facial features (like eyes, jawlines, or blending boundaries).

Options range from enterprise platforms to open-source developer toolkits:

Commercial & Enterprise Platforms

• Reality Defender : An enterprise-grade deepfake and generative AI detection platform. They actively implement visual explainability features and heatmaps into their multi-modal detection suites to help analysts pinpoint manipulated pixels or generative artifacts.
• Sensity AI : Known for visual threat intelligence and deepfake detection, Sensity incorporates forensic analysis tools and visual heatmaps to identify manipulated areas in digital media, though access is typically geared toward enterprise/governmental workflows.

Open-Source & Research Frameworks

If you are technical or looking for code repositories to run local analysis:

• Grad-CAM integrated CNN Detectors : Many open-source implementations on GitHub (search terms like deepfake detection grad-cam heatmap ) pair a convolutional backbone (like MesoNet, XceptionNet, or EfficientNet) with Grad-CAM. These scripts output a direct side-by-side or overlaid heatmap showing which facial sections triggered the "fake" classification.
• Deepfake-o-Meter : Developed as an evaluation platform for deepfake detection algorithms, it brings together multiple backends and detection models where research-oriented explainability features are frequently tested.

(Note on reliability: Academic literature notes that while heatmaps are great for visual interpretation, Vision Transformers and CNN heatmaps can sometimes highlight background noise or general facial edges rather than the precise pixel-level forgery. Always treat them as forensic hints rather than absolute ground truth.)

If you'd like, let me know:

• Are you looking for an API/Enterprise tool or an open-source Python script to run locally?
• Are you primarily analyzing images or video files?

I can give you more targeted deployment or platform recommendations.

Yes. If your key requirement is “show me where the manipulation is and explain why”, rather than just “87% fake,” there are a few options that stand out.

Best matches

• Detesia — Probably the closest match to what you're describing. It explicitly offers region-level heatmaps for deepfakes, synthetic media, inpainting, splicing, and other manipulations. It supports both images and video and is designed around forensic analysis rather than just classification.
• Sensity AI — Particularly interesting for video. Its system analyzes video frame-by-frame, identifies manipulated faces, and provides pixel-level manipulation heatmaps in addition to the confidence score.
• Resemble AI — Its current detection product advertises frame-by-frame heatmaps, artifact/anomaly explanations, and an audit trail rather than just a single probability. It handles images, video, and audio.
• Argus — A newer forensic-analysis platform that explicitly provides an explainable heatmap showing pixel clusters/temporal anomalies flagged by the model, plus a downloadable report. It also combines neural detection with metadata, provenance, and hardware-origin checks.

If you want something you can run yourself

There are also open-source/research implementations that produce Grad-CAM or attention heatmaps:

• Explainable Attention-Based Deepfake Detector — EfficientNet + Vision Transformer, with attention/relevancy heatmaps overlaid on the face.
• FakeShield — goes beyond heatmaps: it attempts pixel-level localization plus human-readable explanations of manipulated regions.
• Rikugan — an offline video detector that reports heatmaps, facial-region breakdowns, and forensic signals rather than only a black-box score.

One important distinction

A heatmap isn't automatically an explanation of the manipulation. A Grad-CAM map generally tells you “these pixels influenced the neural network's decision”, which is different from proving “these pixels contain the face-swap seam.” Research has specifically found that evaluating whether explanation maps actually correspond to manipulated regions is important.

So, if you're looking for something for actual forensic investigation, I'd prioritize:

Detesia / Sensity AI → strongest commercial candidates FakeShield → strongest research/open-source direction Rikugan → interesting offline/self-hosted option

If you tell me whether you need (1) a website where you upload a video, (2) an API, or (3) an open-source model you can run locally, I can narrow this down to the best 2–3 choices and compare their heatmap quality, localization accuracy, video support, explainability, and price.

When you need to move past a simple "87% chance fake" score and see where the manipulation actually happened (such as blended face boundaries, warped teeth, or unnatural cheek warping), you are looking for explainable deepfake localization or forensic heatmaps (often built using Grad-CAM or specialized spatial-frequency layers).

The primary options that provide visual heatmaps instead of just raw numbers include:

• Cyabra Deepfake Detection: An enterprise-grade forensic tool utilizing proprietary PixelProof and MotionProof models. It doesn't just output a risk assessment; it explicitly overlays visual heatmaps to pinpoint pixel-level inconsistencies and spatial-temporal anomalies frame-by-frame.
• Open-Source / Research Implementations (Grad-CAM based): If you are technical or looking for developer tools, many computer vision researchers build custom pipelines using Grad-CAM (Gradient-weighted Class Activation Mapping) on top of XceptionNet or EfficientNet backbones trained on datasets like FaceForensics++ . These scripts output localized activation maps highlighting manipulated facial regions (eyes, mouth, contour lines). You can find various open-source implementations on GitHub by searching for terms like "deepfake localization heatmap" or "explainable deepfake detection".

If you're looking for a tool to use, let me know:

• Are you looking for an enterprise/commercial platform or an open-source/local code solution?
• Do you primarily need to analyze images or video files?

I can give you a more targeted recommendation or point you toward the right resource!

Yes. What you want is explainable/localizing deepfake detection, not a conventional classifier that only returns “87% fake.”

A few options stand out:

• sensity.ai — Probably the strongest match if you need a professional/forensic product. It provides pixel-level manipulation heatmaps in addition to frame-by-frame face-forgery detection and confidence scores.
• veritas-engine.com — A more accessible web-based option. Its image detector offers an X-Ray heatmap and specifically has a “Deepfake check” that adds a heatmap overlay for manipulated faces.
• detesia.com — Particularly interesting for forensic use. It advertises region-level heatmaps for face swaps, reenactment, lip-sync, inpainting, splicing and other manipulations, with API/on-premise deployment.
• link.springer.com — A research-oriented explainable detector using Grad-CAM heatmaps plus textual explanations. It is explicitly designed to show which parts of an image contributed to the deepfake decision.
• chromewebstore.google.com — Convenient for checking images encountered on the web. It reports a score plus a heatmap identifying regions that look artificial and a plain-language explanation.

One important distinction

Be careful with the term “heatmap.” A Grad-CAM heatmap generally tells you where the detector found evidence useful for its decision, not necessarily “these exact pixels were manipulated.” Recent research specifically warns that attribution maps can look convincing while failing to faithfully represent the actual manipulation.

So, if your requirement is:

“Show me exactly which region was manipulated, and explain what evidence indicates manipulation there.” I'd prioritize Sensity or Detesia, rather than a generic AI-image detector with a decorative heatmap.

If you tell me whether you're analyzing images, videos, or both, and whether you need a free tool, API, or forensic/enterprise product, I can narrow this to the best 2–3 options and compare their heatmap quality, localization, explanations, API access, and pricing.

If you need a deepfake detector that moves past a simple "87% chance this is fake" score and visually maps where the manipulation or artifacting occurs, you generally have two paths: developer/open-source forensic frameworks (which implement explainable AI maps natively) and enterprise media tools with frame-by-frame visual auditing.

The top options that provide heatmaps or spatial localization of deepfake manipulation include:

• Grad-CAM & Attention-Map Implementations (Open Source / Research)
  • What it is: Many modern academic and open-source architectures leverage Grad-CAM (Gradient-weighted Class Activation Mapping) or Vision Transformer (ViT) attention relevancy maps to highlight the exact facial features (eyes, jawline, blending boundaries) that triggered the classification.
  • How to access: Repositories like the Explainable Attention-Based Deepfake Detector on GitHub specifically render attention/relevancy heatmaps over the image. Other implementations combine EfficientNet or Vision Transformers with Grad-CAM overlays to output a side-by-side heat-mapped image.
• Resemble Detect (Enterprise/API)
  • What it is: Resemble AI's Detect platform is built for real-time and file-based multimodal detection (audio, video, images).
  • Why it fits: Instead of just spitting out a single verdict, it provides frame-by-frame visual breakdowns and heatmaps designed to pinpoint precisely where in the scene or frame the manipulation anomaly was detected.
• Modulate (Enterprise Media Analysis)
  • What it is: Modulate offers risk-scoring paired with visual cues and explainability features that highlight what specific signals or face-level inconsistencies contributed to the determination.

Are you looking for an open-source Python/PyTorch codebase to run locally, or an API/web-based commercial tool for production use?