AI-Cloned Photoshop Tested: Can $70 Replace Adobe?

The AI Photoshop Clone That Broke the Internet

Late one night, a single topic took over every feed — an AI-cloned Photoshop had appeared, nearly feature-complete, one-time payment of $70, running across Windows, Mac, and Linux. It sounded like yet another "cry wolf" moment, but the more you dig in, the more substantial it turns out to be.

This isn't one of those toy projects where you ask Claude to build a knockoff Photoshop with a handful of basic tools. Using a combination of what appears to be decompiled original applications and "vibe coding," a wave of projects has emerged from the community: from free clones of the entire Adobe Creative Suite, to turning Modern Warfare 2 into a skateboarding sandbox, to a Frankenstein extraction shooter set in the world of Elder Scrolls.

Reactions split instantly. Some condemned it as outright piracy; others saw it as an inevitable evolution — paid proprietary software dying off as AI-assisted open-source alternatives eventually win out. But one thing was universally agreed upon: there's no going back.

Still, one critical question often gets overlooked: regardless of how the software was built, and regardless of the ethical, moral, and legal controversy, none of it matters if the apps simply don't work. And they're almost certainly never going to receive long-term maintenance. So the only sensible move was to actually buy it and test it. Interestingly, during filming it went from $70 to free — a shift that happened within three hours. That's just how fast things move in AI right now.

How It Was Actually Built

To understand this kind of clone software, you first need to understand the two main construction approaches.

Decompilation: Restoring Source Code from Machine Instructions

The first approach is decompilation. Simply put, it means taking the binary code running on a machine — those 0s and 1s — and reverse-engineering it back into human-readable source code that the original developer wrote. The restored code won't be identical to the original, but if it swims like a duck and quacks like a duck, it's a duck as far as every user is concerned.

Decompilation has always existed. Classic examples include the Wine project, which reverse-engineered closed-source Windows APIs to let Windows software run on Linux, and Samba, which analyzed network traffic to enable file sharing between Windows and Linux systems.

The problem has always been that traditional decompilation is extraordinarily labor-intensive and tedious — only worth doing when a project is small, critically important, or the developer is obsessive enough to push through. AI has completely changed this equation. You can simply tell an AI: "Take these completely unannotated machine instructions made of 0s and 1s and reverse them into the source code that generated them." It will grind away at the task, and the results are surprisingly good — limited only by how much API credit your wallet can sustain.

Wine and Samba are landmark achievements in decompilation and reverse engineering, though their technical approaches differ. Wine (Wine Is Not an Emulator) doesn't emulate — it reimplements Windows' API call layer on Linux/macOS, allowing Windows binaries to run directly without virtualizing an entire Windows system. It relies on years of documented reverse analysis of Windows API behavior rather than directly decompiling Windows itself. Samba, meanwhile, reverse-engineered the SMB (Server Message Block) protocol by capturing and analyzing actual network packets, then implemented Windows network sharing interoperability on Unix systems. Both required over a decade of community collaboration to reach production-ready maturity — a useful benchmark when evaluating the "one more year of polish" expectation for AI-cloned software.

Functional Reconstruction: AI-Generated Approximations

It's worth distinguishing that not everything labeled "decompilation" is the same thing. Tech blogger Kingy's blog outlines several different types. What Photon Studio uses is closer to "functional reconstruction" — newly generated AI source code that aims to reproduce the original software's behavior, but is completely different at the byte level from the original source. In other words, it's more accurately described as "an AI artifact that takes the original program as input."

"Vibe coding" is another key concept for understanding these projects. Coined by AI researcher Andrej Karpathy in early 2025, it describes a coding style that relies heavily on large language models — developers no longer review code logic line by line, but instead describe intent in natural language and let AI generate, modify, and debug code. The developer acts more like someone who "commits when it feels right." This approach dramatically lowers the barrier to reproducing complex software: even if the developer doesn't fully understand the generated code, as long as the AI can keep iterating until it "looks like it runs," the project can move forward. The combination of vibe coding with decompilation and functional reconstruction is especially potent — AI can analyze the low-readability output of decompiled code and continuously generate alternative implementations targeting "behavioral alignment" rather than "code alignment." This is the underlying logic that enabled projects like Photon Studio to take shape so quickly.

Is Any of This Even Legal?

The next unavoidable question: is this legal? The answer is — maybe yes, maybe no.

One classic historical case is worth examining: Phoenix Technologies kick-started the affordable clone PC revolution by creating an IBM-compatible BIOS that didn't infringe on IBM's original source code. Their method had engineers study and document IBM's source code listings from technical reference manuals, then handed that documentation to a separate engineer who had "never read the original source code" to independently write code mimicking the same API. This was ruled legal because the person who wrote the code had never seen the original — and therefore couldn't have plagiarized it. This legal exemption for software reverse engineering still exists today.

Does this apply to AI? In that case, the court ruled that it didn't matter how similar the output was, as long as the author had never seen the original code. And AI has "never seen" the original source code — at least, the data aggregators who created it will never admit to having seen it. On the other hand, patent law and copyright law can also cover functionality and creative expression, and the rules vary significantly across jurisdictions. At best, we're operating in extremely murky gray territory.

The "clean room reverse engineering" principle established in the Phoenix BIOS case (Phoenix Technologies, Ltd. v. IBM, 1989 and related precedents) remains the foundational legal basis for software reverse engineering. Its core lies in information isolation: the team studying the original software's behavior and interface specifications is strictly separated from the team writing the new implementation. The latter can only access functional description documents generated by the former — never any original code. This isolation makes the final product difficult to characterize as "copying" in the copyright infringement sense. Open-source projects like Wine, Samba, and ReactOS all follow similar logic. However, once AI enters the picture, the evidentiary logic of "who saw what" becomes murkier: the composition of a large model's training data is difficult to audit. If the model was ever exposed to the target software's source code, the "clean room" protection could fail — and this is practically impossible to prove or disprove in either direction.

Professional User Testing: Where Does It Fall Short?

Theory aside, the real test is putting it in front of a professional designer. The results were mixed.

Performance issues were immediately noticeable. Importing images was dramatically slower than usual — Photoshop handles it nearly instantaneously, while the clone app noticeably stuttered during loading. Some files wouldn't open at all. Multiple testers encountered various error messages, with one offering the blunt verdict: "Game over — I wouldn't use it."

But in some areas, it performed surprisingly close to the real thing. The menu layout is nearly identical to Photoshop, with Object Selection, Quick Selection, and Magic Wand tools all present. "Five seconds in and I was immediately satisfied — couldn't see any obvious difference at first," one tester noted. But look closer and the problems emerge: Auto Select Subject is noticeably rougher, where Photoshop cleanly isolates a subject and their hat, the clone frequently misses details and grabs elements from underlying layers, requiring extensive manual cleanup. And subject selection is precisely the reason many people subscribe to Photoshop in the first place.

RAW processing also exposed weaknesses. Pushing shadows up, the clone couldn't come close to Photoshop's exposure adjustment even with the shadows slider maxed out. The Color Mixer either didn't work or ground to a halt. When moving layers, the position box would update while the actual image lagged behind, and right-click menus took several seconds to respond. For professional users, "having to wait this long for every small adjustment is an immediate dealbreaker."

Verdict: A Promising Alternative That Needs More Time

The overall assessment was more positive than expected. Multiple testers acknowledged "it's way more similar than I thought it would be," with one apt label being "I can't believe this isn't Photoshop." For users on a tight budget who just need basic Photoshop operations, it's genuinely a viable option. But for professional use, it "probably needs another year of sustained development."

What's really worth thinking about is the broader context. The creative community is generally hostile toward AI, yet the reception to these clone projects has been far more positive than anticipated. Photon Studio is just one of many similar projects, some from individuals and some from teams. This suggests that Adobe's near-meme-worthy software quality and its uncompromising subscription-only model have driven enough creators to the point of wanting to escape — even if that exit leads them into the arms of "Big AI."

From Affinity to Photomator, designers have always been searching for Adobe alternatives. The emergence of AI-cloned software is less a technological miracle and more a genuine response to a user base driven to fury by excessive pricing. Whether these tools will mature remains to be seen, but the signal they send is unmistakably clear.