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

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?

Yes. If your requirement is “show me where the manipulation is, not merely how likely the file is fake,” there are several better fits than generic deepfake scorers.

Best matches

• sensity.ai — Probably the strongest fit for video/face deepfakes. Its explainability layer produces pixel-level manipulation heatmaps, while its detector works frame-by-frame and identifies manipulated faces. Sensity AI Best for: professional/forensic video analysis.
• detesia.com — Explicitly designed around explainable deepfake and synthetic-media detection, including region-level heatmaps for face swaps, reenactments, lip-sync manipulation, inpainting and splicing. It also offers on-premise/API deployment. Detesia Best for: organizations that need forensic localization rather than a consumer “AI score.”
• resemble.ai — Particularly interesting if you want the explanation to go beyond the heatmap. It provides frame-by-frame heatmaps, identifies artifacts contributing to the result, and provides human-readable forensic explanations. Resemble AI Best for: video, audio, and enterprise workflows where you want an explanation/audit trail.
• turbolens.io — Very good match if you're primarily interested in images rather than videos. It reports suspicious-region coordinates, tampered-pixel proportion, and a visual heatmap for local manipulation such as splicing, copy-move and retouching. TurboLens Best for: determining which pixels/regions were altered in an otherwise genuine photograph.
• artificialimagedetector.com — Provides an occlusion-based saliency heatmap showing the pixel regions that influenced the verdict, alongside model/forensic breakdowns and reports. It supports images and video. Artificial Image Detector Best for: an accessible analysis interface with visual explanations.

If you want something open-source

FakeShield is particularly interesting academically. It was designed specifically for explainable image forgery detection and localization, producing manipulated-region masks plus human-understandable explanations based on pixel-level artifacts and semantic inconsistencies. It covers DeepFakes as well as splicing, copy-move, removal, and AI-generated imagery.

There's also an Explainable Attention-Based Deepfake Detector on GitHub that generates Transformer relevance/attention heatmaps over faces. It's more of a research/developer project than a polished forensic product.

One important distinction

I'd separate “heatmap” into two categories:

1. Saliency/explanation heatmap — “These pixels caused my classifier to say fake.” This is what Grad-CAM, attention maps, and occlusion maps generally provide.
2. Manipulation localization map — “These pixels are actually likely to have been edited.” This is much closer to what you're asking for and is generally more useful for forensic investigation.

For example, Sensity describes its heatmap as pointing to pixels more likely to be manipulated, while TurboLens actually returns suspicious-region coordinates and tampered-pixel proportions.

If you tell me whether you're analyzing (a) faces/photos, (b) videos, or (c) both, and whether you need a web interface or an API/local model, I can narrow this down to the 2–3 best options and compare their actual heatmap output.

Finding a deepfake detector that goes beyond a simple percentage score to provide visual localization (heatmaps using techniques like Grad-CAM or attention maps) usually moves you from "consumer-facing tools" into explainable AI (XAI) frameworks and open-source forensic repos . Most commercial enterprise APIs hide the underlying mechanism to protect proprietary models, but several academic, open-source, and specialized platforms provide visual proof of where the manipulation occurs.

Top Options for Heatmap-Based Deepfake Detection

- **What it is:** Emerging human-centered forensic platforms like DeepDect integrate dedicated Explainable AI modules explicitly designed for this.
- **How it explains:** It pairs a percentage score with **Grad-CAM heatmaps** overlaid on the face, visually highlighting pixels/regions (like blending boundaries around the jawline or eyes) alongside short textual descriptions of the anomaly.[](https://google.com/goto?url=CAESXwHrOzAVRNl-5SlbxShVmN29hbocNks92WamNJHE5tfZhNy2UGZkoP-i1WtoDG5bCyHImh6VGYCpVmxJ2swzEUEa6UscoDz0dNb9CcUJr2an4kEB-lTDZIu6De4Rj4yV) [[1]](https://google.com/goto?url=CAESXwHrOzAVRNl-5SlbxShVmN29hbocNks92WamNJHE5tfZhNy2UGZkoP-i1WtoDG5bCyHImh6VGYCpVmxJ2swzEUEa6UscoDz0dNb9CcUJr2an4kEB-lTDZIu6De4Rj4yV)[[2]](https://google.com/goto?url=CAESXwHrOzAVRNl-5SlbxShVmN29hbocNks92WamNJHE5tfZhNy2UGZkoP-i1WtoDG5bCyHImh6VGYCpVmxJ2swzEUEa6UscoDz0dNb9CcUJr2an4kEB-lTDZIu6De4Rj4yV)

- If you have a technical background or a local Python environment, open-source implementations on GitHub provide absolute transparency.
- **Projects to look at:** Look at repositories utilizing EfficientNet or Vision Transformers (ViT) paired with **Grad-CAM visualization** (such as [thourihan/DeepfakeDetection](https://google.com/goto?url=CAESUQHrOzAVVGWar1jBvv2LQ8xz8PJIFyNNOqk00rWeK-xonaSz_tGFShqOdkCVOcwoE6QFag3Ucg_-YRWtIitr3cwjtgvpuIQqUxRtXwgyJKAICw) or lightweight edge implementations featuring [krassiaa/deepfake-detection](https://google.com/goto?url=CAESUQHrOzAVXjhD66MOd5cgRmNuP_pHqDzxljLdeqabwZfpANeAVE4m2zVC8Nl9BVi9Ol4R2u1yTbh4LRkEGteBkmjD7crcOfNsMidiA3Tqmi3yWQ)).
- **How it explains:** These scripts run inference on an image or video frame and output a secondary saved image featuring a color-coded heatmap (red/yellow indicating high manipulation contribution) mapped directly over the suspect facial features.[](https://google.com/goto?url=CAESZgHrOzAVtG46dhGuBifrQg-5Iq6pO68KfrtQLOPFYwawPuTQG5L84cu1N40l0op2bNq-cIdQGFWonmsMhcb_P6rAB_4x08NF1aUpSm46aZDZP-1bDtC5gn8kFXDvG0xMjF_6_cTRwA) [[1]](https://google.com/goto?url=CAESZgHrOzAVtG46dhGuBifrQg-5Iq6pO68KfrtQLOPFYwawPuTQG5L84cu1N40l0op2bNq-cIdQGFWonmsMhcb_P6rAB_4x08NF1aUpSm46aZDZP-1bDtC5gn8kFXDvG0xMjF_6_cTRwA)[[2]](https://google.com/goto?url=CAESUQHrOzAVVGWar1jBvv2LQ8xz8PJIFyNNOqk00rWeK-xonaSz_tGFShqOdkCVOcwoE6QFag3Ucg_-YRWtIitr3cwjtgvpuIQqUxRtXwgyJKAICw)[[3]](https://google.com/goto?url=CAESUQHrOzAVXjhD66MOd5cgRmNuP_pHqDzxljLdeqabwZfpANeAVE4m2zVC8Nl9BVi9Ol4R2u1yTbh4LRkEGteBkmjD7crcOfNsMidiA3Tqmi3yWQ)[[4]](https://google.com/goto?url=CAESYgHrOzAVxeTN9xmWUxl_IEcPol50vY8NA4R-OXOLiHDBKBzQBk00sMx1uy0NWXDbP5hrdZfCtQ--EpAI7DxiPK8_QLzp1-xtmNxap-ec-LvU7-0Y8OdeBqC6a2xWCNNuwtQP)[[5]](https://google.com/goto?url=CAESkQEB6zswFRa9OxTTUyjZKRb5X3kRIx1nppXsn4FCWJG-jkg8BlvX4N7-hcMoGnkU-7xi0R27yKeHM-kEiZkKmSK8Gu776b7vAN-Ko2voil8UjI560S7nLgZ5kNuuKDltBmpkN5kxB247p7P7e9RYFgm-SQajSLGjWKOM8ckGcvQnzecP3ozGZy18Erc14ZjTFjtJ)

- **What it is:** Advanced enterprise tools like Reality Defender or Sensity AI.
- **How it explains:** While their default user dashboards prioritize fast verdicts for corporate risk teams, their deep-dive forensic analyst views often include spatial artifact localization, frame-by-frame breakdown graphs, and pixel-anomaly highlights to show *which* part of the media triggered the flag.[[1]](https://google.com/goto?url=CAESYQHrOzAVU0YgNF3GVze0XFEsziN16HyOaqJiSpCrtlLeiM-k3APdmuLk9dhrJ-83x-CxxikDhBKD-yDI6bJBNcpoAzhkDA3dy3L0noZdzzr1H1SGBfPDT7IA6gNQljkTVc4)[[2]](https://google.com/goto?url=CAESQwHrOzAVm7Pj3uNtQ1YVm5QTRiDBV1eu5oSuEA5LWNKcAJMAuC5R_LrDVDxF9mwgSWUWX-7LpyvrVWKFkevPwz9bTQ4)[[3]](https://google.com/goto?url=CAESfwHrOzAVuXgeUTGEd1QTM4IB2uZYWKpUBZjsUEHh7_YdUn-R02O_pNKjnX2MKzfdv1BvMRXuPJdW_U_0c2tt6SEFuSZ9y8kUJbk9AC4v2aNrVF1q-BVqQKG3rEh2_pggV5HkXASjP2rbK4xTMtC4ZWMkk6bhaqb-yP4VO9VKORQ)[[4]](https://google.com/goto?url=CAESfQHrOzAVTU1uJ6Cmm2A7HX9yOA7pWI2F_THmNNKHCAper88uCzHI1rkb8wuW5P-Hj4a1WMadLMmklLSN6TMXZg0zyiupmZc3FK6NCxX_M6z4u8fUKHbmIjJ6wisROrCdeB4b-_saAg9zuIQOwQ-gffm-1VuSFcj6PwwV9Dun)[[5]](https://google.com/goto?url=CAESbAHrOzAVyPSfiC16AlMMgYPlHEGm5wa1GuVzOYsotoM-Clr1VxN9HN9G8ZnhHh1Ajk3oPbVWg2bCMO0mIJRs4Ume0hYs0NoafGniTvcmElcGHPmgeeDP9OlwFi_Ei6mjWC2D_dNbyMCs7NIqdQ)

How Heatmaps Work Under the Hood Here

If you decide to build or run a tool locally, they almost universally rely on Grad-CAM (Gradient-weighted Class Activation Mapping) or SHAP (SHapley Additive exPlanations).

• The model looks at the gradients of the classification score flowing into the final convolutional or attention layer of the neural network.
• It projects a coarse localization map highlighting the exact spatial regions (such as unnatural blending around the ears, warped teeth, or inconsistent eye reflections) that tipped the model off to a fake.

Are you looking for a ready-to-use web application/software , or do you need a Python/Open-Source code repository that you can run locally to generate these heatmaps yourself?

Finding a production-ready, consumer-facing tool that consistently outputs an intuitive pixel-level manipulation heatmap (rather than just a single confidence percentage) can be tricky because most commercial enterprise APIs treat their outputs as binary scores to prevent reverse engineering.

However, explainable deepfake localization falls into two primary categories depending on whether you are looking for open-source research models or forensic/enterprise software:

1. Open-Source & Academic Frameworks (True Heatmaps)

If you have technical capability (Python/PyTorch) or are looking for methodologies that explicitly generate spatial heatmaps using Explainable AI (XAI) like Grad-CAM, LIME, or frequency-domain masking, look at these prominent open-source architectures:

• AUFF-Net / Grad-CAM implementations: Academic frameworks built on standard backbones (like EfficientNet or Xception) coupled with Grad-CAM (Gradient-weighted Class Activation Mapping). These specifically highlight facial regions (eyes, mouth, jawline) where blending artifacts occur and project a color-coded heatmap over the manipulated frame.
• TruthLens: A recent explainable deepfake framework designed to go beyond binary classification by providing both spatial localization and structured reasoning regarding which facial features (eyes, nose, mouth) appear altered.
• EXAI Image Forgery/Deepfake Repositories on GitHub: Searching GitHub for "Explainable Deepfake Detection GradCAM" or "Deepfake Localization Mask" will point you to community implementations that output bounding boxes or pixel-wise heatmaps visualizing manipulated frequency or noise patterns.

2. Enterprise & Forensic Software (Advanced Breakdown)

Commercial options rarely show a raw "heatmap" out of the box for general consumers, but high-end forensic suites do offer advanced spatial and temporal attribution:

• Fraunhofer IOSB (Visual Explanation / XAI research): Applied research groups like Fraunhofer IOSB actively develop tools that merge deepfake detection with heatmap-based and segment-based XAI approaches to visually display the exact structural or frequency patterns triggering a fake flag.
• Resemble AI (DETECT-World): Focuses on forensic-style breakdown of audio/video content by evaluating real-world physics, lighting, and timing anomalies rather than just black-box scoring, giving a much deeper context to where and how the discrepancy occurs.

Are you looking for an open-source code implementation you can run yourself, or an API/software service for a workflow or application? Let me know so I can point you to the right resource.

Yes. If your key requirement is “show me where the manipulation is, not merely how likely the file is fake,” there are several options that fit much better than ordinary deepfake classifiers.

Best matches

• detesia.com — Probably the closest match to what you're describing. It explicitly offers region-level forensic heatmaps for images and video, including face swaps, reenactments, lip-sync manipulation, inpainting, splicing, and post-processing. It is aimed at forensic/investigative use rather than just consumer “AI or not” scoring.
• sensity.ai — Particularly interesting for video. It analyzes video frame-by-frame, identifies suspected fake faces, and provides pixel-level manipulation heatmaps showing areas likely to have been manipulated.
• resemble.ai — Supports image, video, and audio. Its documentation specifically describes frame-by-frame heatmaps that pinpoint where manipulation was detected, along with explanations of the artifacts/anomalies.
• artificialimagedetector.com — A more accessible option. It provides an explainable heatmap, model/ensemble breakdown, and downloadable forensic report. Its heatmap is intended to show the pixel clusters or temporal anomalies that caused the system to flag the media.
• deepguard.org — Provides heatmaps plus an engine breakdown and forensic PDF. It supports image, video, audio, and other media.

One important distinction

A heatmap isn't necessarily a true manipulation mask.

Many detectors use Grad-CAM/saliency, which answers:

“Which parts of the image influenced the detector's decision?” That is somewhat different from:

“These exact pixels were manipulated.” For example, research on explainable deepfake detection explicitly describes Grad-CAM as highlighting regions important to the classifier, while newer forgery-localization systems attempt to produce actual tampered-region masks.

So if you're evaluating these for forensic investigation, I'd prioritize a tool that provides localization + explanation + original/heatmap overlay, rather than simply a Grad-CAM image.

If you want to run it yourself

There are also open-source approaches. FakeShield is particularly relevant: it was designed for explainable image forgery detection and localization and produces tampered-region masks plus human-readable explanations.

There's also Rikugan, a local/offline video detector that reports heatmap overlays and per-region facial analysis, which could be attractive if sending sensitive video to a cloud service is undesirable.

My shortlist would be:

1. Detesia — best fit for serious forensic/explanatory use.
2. Sensity — strong choice if your primary concern is video.
3. Resemble AI — good multimodal/enterprise option.
4. Argus — convenient option with visible heatmaps and reports.
5. FakeShield — best direction if you want an open/research-based system you can inspect or deploy yourself.

If you tell me whether you're analyzing photos, videos, or both, and whether you need a free tool, API, or forensic/enterprise product, I can narrow this to the 2–3 strongest choices and compare their actual heatmap/localization capabilities.

That is somewhat different from:

“These exact pixels were manipulated.” For example, research on explainable deepfake detection explicitly describes Grad-CAM as highlighting regions important to the classifier, while newer forgery-localization systems attempt to produce actual tampered-region masks.

So if you're evaluating these for forensic investigation, I'd prioritize a tool that provides localization + explanation + original/heatmap overlay, rather than simply a Grad-CAM image.

If you want to run it yourself

There are also open-source approaches. FakeShield is particularly relevant: it was designed for explainable image forgery detection and localization and produces tampered-region masks plus human-readable explanations.

There's also Rikugan, a local/offline video detector that reports heatmap overlays and per-region facial analysis, which could be attractive if sending sensitive video to a cloud service is undesirable.

My shortlist would be:

Yes. What you want is deepfake detection with manipulation localization/explainable AI, rather than a conventional classifier that only says “87% fake.”

Here are the strongest options I found:

1. Detesia — probably the closest match

Detesia explicitly advertises region-level heatmaps for image and video forensics. It covers face swaps, reenactment, lip-sync manipulation, synthetic media, and more, and is designed around forensic analysis rather than just a probability score.

Why I'd investigate it first: the output is intended to answer where was the manipulation?, not merely is this fake?

2. Argus — particularly interesting for forensic review

Argus provides an explainable heatmap showing pixel regions that drove the AI verdict, alongside an ensemble breakdown and an audit-ready report. It also supports video/temporal analysis.

This is a good fit if you want something closer to an investigator's report than a consumer “AI detector.”

3. Resemble AI Detect — strong if you need video/audio

Resemble says its system provides frame-by-frame heatmaps pinpointing where manipulation was detected, plus human-readable explanations of the artifacts contributing to the result. It analyzes image, video, and audio.

I'd consider this if your primary use case is video rather than still images.

4. FakeShield — best if you want research/open-source control

FakeShield is an open-source research system specifically built for explainable image forgery detection and localization. It generates manipulated-region masks and human-readable explanations based on pixel-level artifacts and semantic inconsistencies.

This is particularly interesting if you have technical resources and want to run or customize the detector yourself.

---

One important distinction

Don't assume that every heatmap means “this is the manipulated area.”

A conventional Grad-CAM heatmap generally shows which regions influenced the classifier's decision. That's subtly different from a true manipulation-localization map, which attempts to identify the pixels/regions that were actually altered.

For example, recent research distinguishes heatmap-based explanations from segment-based localization, and specifically notes that heatmaps visualize the areas contributing to the model's decision.

If your requirement is:

“Show me the actual suspected manipulated region, preferably at pixel/segment level, and explain what artifact caused the flag.”

I'd prioritize Detesia, Argus, and FakeShield over ordinary deepfake detectors that merely bolt Grad-CAM onto a fake/real classifier.

There is also academic work showing that LIME can outperform other explanation methods for identifying regions that actually influence a deepfake detector's decision.

If you tell me whether you're analyzing (1) photos, (2) videos, or (3) both, and whether you need a web interface vs. API/local software, I can narrow this to the best 2–3 options and compare their heatmaps, localization accuracy, explainability, pricing, and forensic suitability. 

If you need a tool that goes beyond a simple percentage score and visually highlights where the manipulation or artifact occurs, you are looking for Explainable AI (XAI) and forgery localization in deepfake detection.

Options vary from enterprise commercial platforms to academic/open-source tools that utilize techniques like Grad-CAM (Gradient-weighted Class Activation Mapping) or attention maps to render heatmaps:

• Reality Defender is a leading commercial enterprise and forensic platform that employs a multi-model approach. For visual media, it provides granular, explainable indicators and color-coded heatmaps that spotlight the specific facial sections (such as eyes, nose, or mouth) or frequency patterns triggering the alert, moving past a basic black-box percentage score.
• DeepFake-o-Meter is an open-source collaborative research platform hosted by the University at Buffalo Media Forensics Lab. It aggregates multiple state-of-the-art detection backbones, several of which output spatial activation maps or frame-by-frame temporal tracking to show where manipulation is concentrated.
• DeepDect is an academic and applied platform designed specifically around human-centered Explainable AI. It pairs deepfake classification with an explicit XAI module that projects heatmaps over face-swapped or generated regions alongside short descriptive text summaries.
• Explainable Attention-Based Detectors (Open Source) are readily available on repositories like GitHub (such as hybrid EfficientNet and Vision Transformer implementations). These models produce attention or relevancy heatmaps natively from transformer layers, allowing you to see the exact spatial weights driving the classification.

If you'd like, let me know:

• Are you looking for a commercial web application or an open-source/code-based tool?
• Do you primarily need to analyze images, videos, or live video feeds?

I can help narrow down the exact platform that fits your workflow.

Yes. If your requirement is “show me where the manipulation is and explain why it was flagged,” rather than just “87% fake,” there are several much better fits.

Best options

1. Detesia — probably the closest match

   • Handles images and video.
   • Provides region-level forensic heatmaps showing suspected manipulation.
   • Covers face swaps, reenactment, lip-sync, AI-generated media, inpainting, splicing, and post-processing.
   • Designed specifically around explainable/forensic detection, rather than treating explainability as an afterthought.
   • Particularly interesting if you need something for professional investigation rather than casual “is this AI?” checking.

2. Sensity AI — strong for video

   • Performs frame-by-frame face-forgery labeling.
   • Its explainability layer produces pixel-level manipulation heatmaps, in addition to identifying fake faces.
   • This is a particularly good fit if you want to inspect where in a video the suspected manipulation occurs.

3. Resemble Detect — best if you want textual explanations too

   • Supports image, video, and audio.
   • Provides frame-by-frame heatmaps indicating where manipulation was detected.
   • It also generates human-readable forensic explanations describing the artifacts/anomalies that contributed to the result.

4. Argus — good consumer-accessible forensic workflow

   • Returns an explainable heatmap rather than only a score.
   • Uses occlusion-based saliency maps to identify the pixel regions driving the verdict.
   • Also provides model/ensemble breakdowns and downloadable forensic reports.

5. • Provides visual indicators of which portions of a file were adapted, alongside its manipulation score.
   • Supports images, video and audio and is oriented more toward enterprise/forensic use.

One important distinction

A heatmap isn't automatically proof that those pixels were manipulated.

Many detectors use Grad-CAM/saliency/attention maps, which essentially mean “these areas influenced the model's decision.” That's weaker than a true manipulation localization mask. Research specifically distinguishes these approaches and has found that localization maps need to be evaluated against known manipulated regions.

So, if you're evaluating these tools for forensic or investigative use, I'd prioritize:

Detesia → Sensity → Resemble Detect

and ask specifically whether their heatmap represents:

• model attention (“this area influenced the decision”), or
• manipulation localization (“this area contains evidence of alteration”),
• whether it works frame-by-frame for video,
• and whether the system can explain the type of artifact (face boundary, warping, inconsistent texture, blending, etc.).

If you tell me whether you're looking for images, videos, or both, and whether this is for research, journalism, legal/forensic work, or general use, I can narrow this down to the 2–3 strongest detectors and compare their heatmaps, accuracy, pricing, and accessibility.

For an explainable deepfake detector that provides visual heatmaps instead of just a flat probability score, Reality Defender and Resemble AI Detect are the leading commercial platforms. Both platforms use multi-model architecture to parse inputs and generate pixel-level or frame-by-frame heatmaps (using Grad-CAM or similar saliency techniques) to explicitly show which regions of a face, image, or video triggered a manipulation flag.

Top Commercial & Enterprise Heatmap Tools

• Reality Defender : Combines an ensemble of multiple deep-learning models and outputs visual localization maps. Through its developer tools and enterprise dashboard, it provides pre-signed URLs for image heatmaps and scene-by-scene temporal breakdowns for video, translating a raw score into a defensible forensic audit trail.
• Resemble AI Detect : Offers a multi-modal real-time and file-upload detection suite that generates frame-by-frame video heatmaps, audio segment breakdowns, and visual image heatmaps to pinpoint exact zones of synthetic insertion. They also maintain a consumer-facing Resemble AI Deepfake Detector Chrome Extension that renders image heatmaps directly in the browser.
• DuckDuckGoose AI : Provides enterprise screening via platforms like Phocus , breaking down media authenticity with visual indicators and localized inspection maps for fraud and compliance analysts.

Open-Source & Research Frameworks

If you have technical resources and prefer an open-source or customizable model rather than a commercial service:

• Explainable Attention-Based Deepfake Detector : Available via repositories like the Explainable Attention-Based Deepfake Detector GitHub Project , this hybrid EfficientNet and Vision Transformer architecture renders attention/relevancy map visual outputs directly.
• Academic / Custom Pipelines : Many researchers build local implementations using XceptionNet or EfficientNet paired with Grad-CAM (Class Activation Mapping) and MediaPipe Face Mesh to map manipulated pixel density onto facial features (like eyes, jawline, or blending borders).

Would you prefer an API-driven enterprise platform (like Reality Defender or Resemble AI) or an open-source Python/GitHub code implementation that you can run locally?