slimdeer: the ioredis author had AI write an image compressor that beats TinyPNG; 12 days, 500 stars

2026 Oct 27, early morning 8 (Beijing time), Zihua Li (GitHub luin), creator of ioredis and former CTO of Shimo, posted in V2EX's "share your creations" node with the title "pngdeer — the best open-source PNG compression library (beats TinyPNG by 8%–12%)." The first sentence of the post: "I originally just wanted to test the limits of AI: I handed every PNG file on my computer to Opus 5.5 and asked it to build a compression tool that beats TinyPNG." The library got 113 stars that day, 284 the next day, and 12 stars in 502 days since the repo was created (GitHub, 11-06). Image compression is an old category that has been crowded for over a decade. This one broke out partly on the author's reputation and partly on a pattern anyone can copy: pick a closed-source benchmark everyone has used, make yours open-source and local, and prove you win with a reproducible leaderboard.

What it is

An open-source (MIT) image compressor written in Rust, with a command line (installable via brew install) and a browser version. The browser version is WebAssembly, so images are not uploaded. It started as PNG-only, then gained WebP, JPEG, AVIF and APNG animation support. In Nov 4 it was renamed slimdeer, and the author joked on X about "daily regret over the project name +1".

Its biggest difference from others is that there is no quality slider. For each image it tries 2 to 256 color counts, with and without dithering, and two palette algorithms. It scores results with the perceptual quality metric SSIMULACRA2, then picks the smallest file that still meets the quality bar. Users don't have to guess whether to set quality to 70 or 80.

The cost is speed: each image is encoded several times. A V2EX user, NickyPP, said: "With no visible difference, it's about 20% smaller than pngquant, which I normally use, but it is a bit slow." The author's answer: sacrificing speed for quality is worth it, since you compress once and use the result for a long time, and the bandwidth saved is worth more.

Why this one

People knew the name first. luin has 4944 followers on GitHub, and his V2EX bio reads "former CTO @Shimo, author of ioredis". The first post got 33 replies. The same day, a Chinese AI news site reposted it under the title "AI-built open-source PNG compression library", and someone posted a YouTube video the same day (204 views). It wasn't on Hacker News or Product Hunt, and no Reddit hits turned up. It also didn't make GitHub trending, since trending excludes repos under about two months old. The first wave was almost entirely the author's own Chinese developer circle.

The headline travels in one line. "Open source, and even smaller than TinyPNG". TinyPNG is an image compression site nearly every web and app developer has used. It is closed-source, requires uploads, and charges for its API. Using it as the target means no explanation is needed. The repo topics also include tinypng-alternative and tinypng-cli, which match searches for "TinyPNG alternatives" (inferred).

The README reads like a research report. 180 images, split into a tuning set of 109 and a held-out set of 71 collected only after tuning. For open-source rivals, each image gets the best of ten settings picked after the fact; slimdeer only gets whatever its own search finds at compression time. It also re-runs the test on butteraugli, a metric it never optimized for, and still comes out 6–9% smaller. The whole suite can be re-run. English site LockerGnome 10-29 wrote it up and singled out the held-out set. Technical readers hate self-promotion, and this leaderboard shuts down doubt up front, so people actually star it.

AI has another layer of spread. "Hand a PNG to Opus 5.5 and have it beat TinyPNG" is good material for AI coding discussions, and people who don't care about image compression will still share it.

Of these, only the first can't be copied by a follower. "8–12% smaller than TinyPNG" by itself isn't new: a tuned pngquant roughly ties TinyPNG too (on its own leaderboard the gap is 0.0 points). What makes people share is the combination: open source, reproducible, beats the closed-source benchmark, and was built with AI.

Where the stars came from

The peak didn't drop to zero, thanks to the author's rhythm: for over a week after launch, a new format every two or three days, with an X post each time; finally a rename and another round, which pulled in another +42 day.

The V2EX launch thread was itself a growth engine. Someone said "you could add an mcp feature", and the author replied "Great idea! Adding it now"; someone asked about compressing a whole folder, and a few floors later the author replied "already supported"; someone asked about APNG animations, and again a few floors later "already supported"; someone posted a three-way comparison table across 39 images. User iHongRen wrapped it into a macOS menu bar app, Mite, the same day, and someone said they'd wrap a Python SDK "for production use". By 11-06 there were 7 satellite libraries, each with 0–3 stars, and no feature copycats.

There's another layer: slimdeer's site sits at heydeer.ai/open-source/, and HeyDeer is the author's company in Singapore, whose main business is AI code review (pricing page from $75/month). HeyDeer also turned one of slimdeer's PRs into a "real AI code review" example page. This is using a small open-source tool to build reputation for the core business; whether it converts is unknown.

What a follower can actually take

Building yet another "smaller than TinyPNG" compressor probably gets you nothing. In this space Squoosh has 26k stars, ImageOptim 10k, Caesium 6000+, pngquant 5700+. slimdeer already owns the spot "open source beats TinyPNG", and the author is still iterating fast.

What you can take is the playbook. It has three parts, and only one depends on fame:

• Target: a closed-source tool everyone has used that charges or requires uploads.
• Evidence: a reproducible leaderboard where even the rival's best settings lose, placed on the first screen of the README.
• Rhythm: adding features and posting updates repeatedly in the two weeks after launch, and taking requests in comments on the spot.

Fame decides how many people see it on day one. The last two decide whether those people star and share.

If it were you

B. "Open source + local + reproducibly beating a closed-source benchmark" is a pitch that hasn't been overused, and with AI writing code, one person can build a convincing leaderboard in weeks. But slimdeer's first wave came from the author's fame; 500 stars is only middling for an author with 5000 followers, and anyone without their own channel should discount that.

Good fit for: developers who know a file-processing niche (PDF, video, audio, fonts, GIF); people willing to build test sets and run benchmarks; people with some presence on V2EX, the Chinese X community, or an English community.

Bad fit for: anyone who just wants to build another image compressor; anyone without hard numbers who only wants to write "faster and easier"; anyone who posts and then ignores the comments.

How: pick a closed-source benchmark that people use daily but that requires uploads or payment: web tools like Smallpdf and iLovePDF for PDFs, remove.bg for background removal, and the sites for video, GIF compression and format conversion (these targets are inferred; nobody has verified they'd earn stars). Use AI to build the core algorithm, ship a CLI plus an in-browser version, and go zero-upload. Put three things on the README's first screen: one sentence "N% smaller at the same quality as X", a live demo link, and a leaderboard. Split the test set into tuning and held-out images, give rivals their best settings after the fact, and re-run with a second metric. You could also target what slimdeer does poorly: it's slow, with a large image in the browser taking ten seconds to a minute, so a "slightly worse but ten times faster" version may have its own audience (inferred).

How to get users: on launch day, post to both V2EX ("share your creations") and X, with the target's name and the percentage in the title; open the body with how you had AI build it. After that, post once for every feature you add, build comment requests the same day when you can, and reply "already supported" when you do. Put "<target>-alternative" in your topics.

Deadline: the story "AI alone built something that beats an established closed-source tool" is still fresh, with maybe a few months left; once posts like this multiply, the AI angle's pull will fade first.

Where this estimate could be wrong: ① someone without fame follows this playbook for a different category, an "open-source X", and passes 300 stars within a month, which would mean the playbook is enough and B is too low; ② conversely, a famous author copies it and it doesn't take off, which would mean slimdeer leaned mainly on the ioredis brand; ③ TinyPNG updates its algorithm and pulls ahead, so "beats the benchmark" stops being true and the pitch loses force.

Data and sources

• GitHub: heydeer/slimdeer's README, topics, stars and forks (recorded daily since 10-27, 11-06 live with 502 stars, 43 forks); author luin's profile.
• V2EX: t/1252084 (10-27 launch post and replies), t/1254634 (11-05 second post).
• X: @luinlee posts on going open source (10-27), WebP (10-28), JPEG (10-30), and the rename (11-04).
• Coverage and reposts: LockerGnome (10-29), 80aj, it8090, Baffox, youtol; YouTube Alex Hitt (10-27,204 views).
• HeyDeer: heydeer.ai/open-source/slimdeer, heydeer.ai/examples/slimdeer-pr-6, heydeer.ai/pricing.
• Rival star counts: GitHub (around 11-06).