Enjoying this issue?
Get tomorrow's AI & engineering digest in your inbox — hand-picked, summarized, and always spam-free.
TLDR
PhotoCraft, a free Photoshop clone, uses PSD files as answer keys to grade itself: it compares its rendering to the final composite image that Photoshop already saved inside each file. The clone passes 258 of 258 round-trip tests (open, save, reopen) but only about half of the oracle tests that check whether the output matches Photoshop's own picture. The two scores reveal different aspects of compatibility, and the founder claims the implementation is clean room, built from public specs and observed behavior.
Key points
PhotoCraft passes 258 of 258 round-trip tests but only at least 130 of 256 oracle tests.
PSD files store the final merged image, which the clone uses as an answer key for grading.
The founder claims the clone is clean room, built from public specs and observed behavior.
The grading method allows AI agents to optimize code by chasing a score.
Without maximize compatibility enabled, PSD files lack the answer key and cannot be graded.
Tools mentioned
Techniques
- clean room reverse engineering
- using PSD composite as answer key
Stop scrolling. Start reading smarter.
Receive the day's most important AI & engineering updates in one concise email. No spam.
Transcript (captions)
PhotoCraft, the free Photoshop clone, passes 258 of 258 Photoshop saved test files on one check and at least 130 of 256 on another. Same corpus of Photoshop files, two very different scores. How
does a clone grade itself against Photoshop? The obvious guess is that it read Adobe's code. The readme says it's clean room, built from public specs and observed behavior. That's the team's
claim. No independent audit turned up. The grading comes from the files. Adobe's public spec says a PSD stores the complete merged image in the last section of the file. Photoshop already
did the rendering and saved the answer inside. Think of the answer key at the back of a textbook. The clone does the problem, flips to the back, compares. It breaks in one place. The key is only the
final image, never the steps. The team built a corpus, 258 small PSD files, generated by scripts and saved by Photoshop. Each no bigger than 320 by 320 pixels. PhotoCraft downloads them at
a pinned commit, checks their SHA-256 hashes, renders each one, and compares the result with Photoshop's own composite. That turns does it look right into a score. A score is something a
coding agent can chase, and agents usually work like that. Change code, run tests, read the number, repeat. The founder says he vibe coded these. How his agents were run isn't documented.
Now the proof from PhotoCraft's v0.3 0 release notes, the 7th of October, round trip, meaning open, save, reopen, 258 of 258. Oracle, meaning match Photoshop's own picture, at least 130 of 256.
Do the division and it's about half. A file can survive its own save and still not look like Photoshop's picture. Both scores are real. One just asks an easier question. Those
are the 7th of October floors, and v0.7.0 landed the 11th of October. So, if you pay for Photoshop, watch the second score, not the first. Where a file hands
you the answer, a clone can grade itself fast. Saved without maximize compatibility, a PSD has no answer, and there's nothing to grade against. Same files, two scores. 258 of 258 says the
file survived. At least 130 of 256 says it matches Photoshop's own picture. So, when a clone says it passes, which test is it passing?