Somebody Just Vibe Coded EVERY Big ADOBE App

0:00 A free Photoshop clone in Rust passes 307 of 309 test files, but that number doesn't show it matches Photoshop. If you pay Adobe each month, you need to know what it does show. The project is

0:13 Photocraft, and it had 13,751 GitHub stars on the 7th of October. If AI can code, where's the Photoshop clone? This repo gives a partial answer. Many PSD files carry Photoshop's own

0:26 flattened picture, so the code can check its work against it automatically. I'll call that check the judge. What does it measure, and where does it stop? Here's the short answer. The 307 is a resave

0:38 check. Photocraft opens a file, saves it, and renders the same picture. The judge, the check against Photoshop's own flattened picture, has an enforced minimum of only 229 of 309 files. The

0:50 readme calls the project early alpha and not yet a Photoshop replacement for daily professional work. Next, how the judge works and where it stops. Start with the file. A PSD stores your layers.

1:03 It can also store one flat picture of the finished result if Photoshop writes one when you save. Programs that can't read every layer use that flat picture. I'll call it the merged image. The

1:13 merged image is real only in some files. When maximize compatibility is off, Photoshop can store white placeholder pixels instead. A flag called has real merge data tells a program which case it

1:25 has. This is general behavior of the PSD format reported in an Adobe community thread. The judge compares two pictures. Photocraft renders a PSD's layers with its own compositor, the code that stacks

1:37 layers into one picture. Then it compares the result with the merged image that Photoshop stored in the same file. Each place where the two pictures differ shows where the clone departs

1:46 from Photoshop. The architecture document calls this a free large conformance test suite because nobody writes expected outputs by hand. The docs don't say how close two pixels must

1:56 be to count as a match. Take a gradient between two colors. The readme lists three interpolation modes: classic, perceptual, and linear. Each mode can make the middle of the gradient come out

2:07 a little different. Photoshop's merged image holds the real result. So, the clone has a known target for each pixel. Think of an answer key printed inside an exam paper. The photocraft code is the

2:18 student and the stored merged image is the key. Each PSD file with a real merged image is one more exam. The comparison breaks in two places. Some papers have no key and the key can't

2:29 grade how a brush feels. A second check runs inside the app. Photocraft has a CPU renderer and a GPU renderer. For every blend mode, adjustment and effect, the two must match within 1255th for

2:42 8-bit output. You say that as 255th. In plain words, no color channel may differ by more than one level out of 255. So, the GPU path can differ from the CPU path by at most one level. A third check

2:56 is a large number of tests. The README claims more than 1,700 tests. They include PSD resave checks, compositor checks against the merged image, and checks at more than one bit depth. The

3:09 tests run automatically in CI, the continuous integration system on three test sets of real files. Each test set has an enforced minimum pass count. An agent also needs a way to run the app

3:20 and see the result. The repo has a registry of more than 500 commands. The UI, the command line, a JSON control channel, and an MCP server all call the same commands. An MCP server is a

3:32 standard way for AI tools to call the app. The control channel accepts only local connections, needs authentication, and can take off-screen screenshots. Claude Opus 5.5 appears in a large share

3:43 of the commits. On the 7th of October, GitHub commit search lists 166 commits with a co-author line naming Claude Opus 5.5. The repo page shows 365 commits in total. Anthropic announced that model on

3:58 the 22nd of September 2026. The team created the repo eight days later. The repo doesn't document how the agents were prompted or how much a person steered them. It does show the setup. A

4:09 judge enforced test floors and a way to run the app. That kind of setup gives an agent feedback it can use. So, this video covers what the setup allows. The agent's own process is a gap. The 307 of

4:22 309 number comes from the resave check. PhotoCraft opens a file, saves it again, and renders the same picture. The readme says this holds for 307 of the 309 files in the PSD tools test set. The resave

4:36 check compares the render before the save with the render after it. It doesn't compare with Photoshop. The harder check is the judge, the comparison with Photoshop's merged

4:45 image. The scorecard sets its enforced floor at 229 of 309 on the same test set. So, the tests require at least 229 files to match the merged image. The sources don't say why the floor is lower

4:59 than 307. Those are CI floors, the minimum the test allow from a scorecard dated the 5th of October. They aren't today's exact pass counts. The second test set has 170 files. There, the

5:12 judge's floor is 146 and the resave floor is 169. The same gap shows again. The third number 116 of 116 is a separate result. It compares shape layers with Photoshop's merged image on

5:26 the project's own set of PSD files. It covers shape layers, not all of Photoshop. The readme lists gradient interpolation, layer effects, shape strokes, clipping, and fill opacity as

5:38 checked features. The test files come from the open ecosystem. PSD tools is an MIT licensed Python library and PhotoCraft uses its 309 file test set. PSD tools itself says the PSD

5:50 specification is far from complete. Photocraft says it's a clean room project built from public specs without Adobe source code. Another number sounds better than it is. The par page says 626

6:02 of 626 Photoshop menu items are present. A script generates that number by comparing Photoshop's menu tree with the live command registry. It proves a command exists. It doesn't prove the

6:13 command behaves like Photoshop. The scorecard also lists 60 of 135 settings that do nothing. It shows only three of 47 performance scenarios meeting their targets. It says document saves are not

6:26 atomic, which in general means a save that stops halfway can leave a partly written file. One layer style operation crashed with an error from WGPU, the Rust graphics library. The suite has

6:38 seven apps. Only Photocraft and Print Craft are labeled early alpha. The other five are labeled in development and one of them is Lightcraft, the Lightroom style RAW app. So Lightcraft, the RAW

6:49 app, isn't a finished release. The site says all seven apps are free and open source. Gigazine reported on the 5th of October that official installers are only planned. Testers in one hacker news

7:00 thread named weak raw processing, two PSDs that change on save, and limited brushes. One wrote, "Lightcraft's image looks like an amateur attempt at raw processing from first principles." In a

7:12 second thread, the reaction split. One tester wrote, "Tested it. Very broken software produced by AI." Another said, "The features they tested worked." A third said, "This could very fast result

7:23 in me canceling my sub." So, who can cancel Adobe today? If you edit layered PSDs, Photocraft is worth testing on copies of your files. The enforced minimum is 229 of 309 files. So some

7:36 files may still differ. This verdict rests on the REDME's gap list, the scorecard and the testers, not on a benchmark. In one thread, a tester said two test files changed on save so much

7:47 that an export render differed. The scorecard also says saves are not atomic, so keep the original untouched. Work on copies of your files. Then compare each export with the original.

7:58 For RAW, keep Lightroom. The judge in this video is the merged image inside a PSD, so it doesn't apply to a RAW file. Testers also called the RAW output weak and Lightcraft isn't a finished release.

8:10 Plugins are the second reason to keep Adobe. The README's own gap list names plug-in compatibility. It also names AI features. About 20 missing tools and typography depth. If your work depends

8:21 on a plugin, check that gap list before you cancel anything. Pro brushes are the third reason. Testers reported limited brushes and the readme lists pro workflows as a gap. A merged image can

8:32 show where paint landed. It can't tell you how the stroke feels as you draw. Now, the bigger lesson, and it's my inference from one case, the team hasn't said it. A clone gets easier to build

8:43 with three things. A public file format, many real files, and an answer stored in many of them. Other apps with those three things could come next. The conditions matter. If a PSD has no real

8:54 merged image, the judge has nothing to compare. If an app has no public format, there's nothing to read. Speed needs a different test and the scorecard shows only three of 47 performance targets

9:05 met. Photocraft shows that a partial Photoshop clone is possible. If you edit layered PSDs, test it on copies of your files. If you shoot RAW, use plugins, or paint with Pro Brushes, keep your

9:17 subscription. Which paid app has a public file format, many real files and an answer stored inside them? That app could get its own judge.