Invisible Dots Review 2026: AI Image Watermarking

What Invisible Dots does in 2026

Invisible Dots is a self-hosted browser agent, described in its repository as an open-source alternative to OpenAI Dots and Grok Bot. The GitHub project at feder-cr/invisible_dots shows 32,114 stars, which is a lot of attention for a browser automation tool. The category listing places it under productivity, with a starting price of $0/mo — the self-hosted model means you run it on your own infrastructure and pay for the compute you consume, not for a seat license.

That positioning matters for this review because the brief asks a specific question: does Invisible Dots function as an AI image watermarking and provenance tool that satisfies EU AI Act Article 50 transparency duties, or is it compliance theater? The honest answer, based on what the documentation and repository describe, is that Invisible Dots is not primarily an image watermarking product. It is a browser agent. The watermarking and provenance conversation is adjacent to what it does, and the confusion is worth untangling before anyone buys it expecting C2PA manifests or invisible pixel-level watermarks.

What the tool actually does is drive a browser in a way that is hard to detect as automation. The repository describes it as undetectable, self-hosted, and open-source. That is a browser-agent capability, not an image-provenance capability. If your compliance problem is "I need to mark AI-generated images so platforms and regulators can identify them," Invisible Dots is the wrong layer of the stack. If your problem is "I need to automate browser workflows without being fingerprinted as a bot," it is the right layer.

This review separates those two things. It looks at what the tool provides, how setup works according to the docs, what the robustness picture looks like for watermarking generally, how detection accuracy tends to behave, what Article 50 actually requires, and how pricing compares to dedicated watermarking vendors. The verdict at the end is a recommendation plus a fallback stack for teams that need provenance specifically.

Watermark types and API surface

Dedicated invisible watermarking tools typically expose two endpoints: embed and detect. Embed takes an image plus a payload and returns a modified image with the payload encoded in a way that survives some transformations. Detect takes an image and returns the payload if present, or a negative result. Capacity — how many bytes you can hide — varies by vendor and by image size. Supported formats are usually PNG, JPEG, and sometimes WebP or TIFF.

Invisible Dots does not present itself that way. The repository describes a browser agent, not an embed/detect API. There is no documented embed endpoint, no detect endpoint, no stated payload capacity, and no format support matrix in the tool block. That is not a criticism of the tool; it is a statement about what it is. A browser agent that controls a real browser session is a different product from a watermarking SDK.

Where the two overlap is in workflow. If you are generating AI images at scale and need to attach provenance metadata before publishing, a browser agent can automate the upload step to a platform that applies its own labeling. It can also automate the process of running images through a separate watermarking service. But the watermarking itself is not something the repository claims to do. Anyone reading a review that implies otherwise should treat that as a red flag.

The practical takeaway: if you need embed and detect endpoints, you need a watermarking vendor, not a browser agent. If you need to automate the pipeline around a watermarking vendor, a browser agent is useful. Invisible Dots sits in the second category.

Setup: key, first watermark, detection round trip

Because Invisible Dots is self-hosted, setup starts with cloning the repository and running it on your own machine or server. The repository is the source of truth for installation steps. There is no hosted API key to request, no dashboard to log into, and no per-request billing from the vendor. The $0/mo starting price reflects that: you pay for your own compute.

For a watermarking tool, setup usually looks like this: sign up, get an API key, install an SDK, call embed with an image and a payload, call detect with the result, confirm the payload comes back. Time to first detect is typically minutes. Rate limits are usually published per tier. None of that applies here because Invisible Dots is not that kind of product.

What setup does involve is getting a browser agent running reliably. That means handling browser binaries, session management, and whatever anti-detection configuration the repository documents. The 32,114 stars suggest a community has worked through these steps and contributed fixes, which is a real advantage over a closed-source agent you cannot inspect. Self-hosting also means you control the data path, which matters if you are processing images that cannot leave your infrastructure.

The detection round trip for watermarking — embed, transform, detect — is not a flow the repository describes. If you need that flow, you are looking at a different tool. The honest framing is that Invisible Dots handles the automation layer, and a watermarking vendor handles the marking layer. Setup for the two is separate.

Robustness: crop, resize, JPEG q60, screenshot, re-upload

This is where watermarking claims usually fall apart, and it is worth explaining why even though Invisible Dots is not a watermarking tool. Invisible watermarks are fragile in specific, predictable ways. A survival matrix across common transforms is the only honest way to present them.

Typical behavior for invisible watermarking, based on how the techniques work:

• Crop. Aggressive cropping removes the region where the watermark signal lives. Small crops often survive; large crops often do not. This is the single most common failure mode.
• Resize. Downscaling destroys high-frequency signal. Upscaling does not restore it. Moderate resizes sometimes survive; heavy downscales usually do not.
• JPEG q60. Lossy re-encoding at low quality strips fine detail. Watermarks embedded in high-frequency coefficients are vulnerable. Quality 60 is a common social-platform setting, so this matters.
• Screenshot. A screenshot re-rasterizes the image at screen resolution, which is effectively a resize plus a re-encode. Survival depends on the watermark's frequency band and the screen resolution.
• Re-upload. Platforms re-encode on upload. The transform is usually a resize plus a JPEG encode, so the same vulnerabilities apply.

No vendor survives all five transforms at full strength. The ones that publish robustness matrices are the ones worth trusting, because they have measured it. A vendor that claims universal survival is selling theater. If you are evaluating Invisible Dots for a provenance use case, the question is not whether its watermark survives cropping — it does not embed watermarks — but whether the automation layer reliably delivers images to a watermarking service that does.

Detection accuracy: false positives and false negatives

Detection accuracy has two failure modes and both matter for compliance. A false negative means a watermarked image is reported as clean, which defeats the purpose. A false positive means a clean image is flagged as AI-generated, which is worse — it accuses legitimate content and erodes trust in the system.

Dedicated watermarking vendors usually report accuracy on a batch of watermarked versus clean images. The numbers vary widely. Some report near-perfect detection on unmodified images and much lower detection after transforms. The gap between those two numbers is the honest measure of the tool.

Invisible Dots does not publish detection accuracy for watermarking because it does not do watermarking. Its accuracy claims, if any, would be about browser automation reliability — whether it completes a workflow without being blocked. That is a different metric and should not be conflated with watermark detection.

For teams building a provenance pipeline, the practical advice is to run your own batch test. Take a set of watermarked images, apply the transforms you expect in production, and measure detection rate. Take a set of clean images and measure false positive rate. Do this before you commit to a vendor. No review can substitute for that, because your transform pipeline is specific to your platform.

Compliance: EU AI Act Article 50, C2PA, platform rules

EU AI Act Article 50 imposes transparency obligations on providers and deployers of certain AI systems. For synthetic image content, the relevant duty is to mark outputs in a machine-readable way so they can be identified as AI-generated. The obligation bites in 2026, which is why this category is suddenly crowded.

Article 50 does not name a specific technology. It requires that outputs be marked in a machine-readable format and detectable as artificially generated or manipulated. That leaves room for multiple approaches: invisible watermarks, C2PA manifests, metadata, or a combination. The regulation sets the outcome; the implementation is up to the provider.

C2PA is the Content Provenance and Authenticity standard, backed by the Coalition for Content Provenance and Authenticity. It embeds a signed manifest in the file describing origin and edit history. It is metadata-based, which means it is robust to some transforms but can be stripped by platforms that re-encode without preserving metadata. Invisible watermarks are pixel-based, which means they survive metadata stripping but are vulnerable to the transforms listed above. The two are complementary, not competing.

Platform rules add another layer. Major platforms have AI-labeling policies that require disclosure of synthetic content. These policies are enforced by the platform, not by the regulation directly, and they change frequently. A compliance stack that satisfies Article 50 may still violate a platform's labeling rule if the platform requires a specific disclosure format.

Where does Invisible Dots fit? It does not. It is a browser agent. It can automate the process of applying C2PA manifests or watermarks through a third-party service, but it does not itself satisfy Article 50. Anyone marketing it as a compliance solution for AI image provenance is stretching the description. The repository describes a browser agent, and that is what it is.

Beetlix is our own product. Where a team needs to coordinate AI workflows that include provenance steps, a workflow layer can sit alongside a watermarking vendor and a browser agent. That is a fair comparison only in the sense that all three occupy different layers; none replaces the others.

Pricing versus Digimarc, Steg.AI, and C2PA tooling

Pricing for watermarking is usually per-image, with volume discounts. The three reference points in this category are Digimarc, Steg.AI, and C2PA tooling.

Digimarc is the long-established player, with enterprise pricing that is typically negotiated rather than published. Steg.AI publishes tiered pricing aimed at developers and platforms. C2PA tooling is often open-source, with costs coming from signing infrastructure and integration work rather than per-image fees.

Invisible Dots lists a starting price of $0/mo because it is self-hosted. That is not a per-image watermarking price; it is the cost of the software license, which is zero. Your actual cost is compute, which depends on your infrastructure. At 1,000 images, 10,000 images, or 100,000 images, the per-image cost of a self-hosted browser agent is whatever your compute costs divided by volume. It does not scale the way a per-image watermarking API does, because there is no per-image fee.

That makes direct price comparison misleading. If you need watermarking, you pay a watermarking vendor per image. If you need browser automation, you pay for compute. Comparing the two on a per-image basis is comparing different products. The honest framing is that Invisible Dots is cheap because it is free software, and watermarking vendors charge because they provide a service you cannot easily replicate.

For a team at 100,000 images per month, the watermarking bill is the dominant cost, and the browser agent is a rounding error. For a team at 1,000 images per month, the watermarking bill may be small enough that the integration work dominates. Either way, the two costs are separate line items, not alternatives.

Verdict: useful compliance layer or theater?

Invisible Dots is not compliance theater because it does not claim to be a compliance tool. It is a self-hosted browser agent with 32,114 stars on GitHub, a $0/mo starting price, and a clear description as an open-source alternative to OpenAI Dots and Grok Bot. Judged on that basis, it is a credible open-source option for teams that need browser automation they can inspect and run on their own infrastructure.

Judged as an AI image watermarking tool, it is the wrong product. It does not embed watermarks, does not detect them, and does not publish a robustness matrix. Anyone who needs Article 50 compliance for synthetic images should look at a dedicated watermarking vendor plus C2PA tooling, and use a browser agent only for the automation around that pipeline.

The fallback stack for a team that needs provenance in 2026: a watermarking vendor for pixel-level marking, C2PA signing for metadata provenance, a browser agent like Invisible Dots for pipeline automation, and a workflow layer to coordinate the three. Each layer does one job. None of them is a substitute for the others, and no single tool satisfies Article 50 on its own.

How this review was researched

This review draws on the Invisible Dots repository at github.com/feder-cr/invisible_dots, the tool listing data including the $0/mo starting price and the 32,114 star count, the public text of EU AI Act Article 50, the C2PA specification, and published pricing pages for Digimarc and Steg.AI. No hands-on testing was performed; the analysis is based on documentation, repository signals, and public regulatory and standards text.