Bolt photo viewer for Mac advances to the next frame in 31 ms, ApolloOne takes 87
Browsing HEIC photos from an iPhone on a MacBook feels sluggish, so the author spent three evenings with AI agents building Bolt, a photo viewer. From the arrow key press to a sharp image, Bolt's median is 31 ms, while the closest competitor, ApolloOne, takes 87 ms. Switching the system decoder made the difference: an iPhone photo now decodes in 17 ms instead of 151 ms. For a narrow utility, vibe coding with measurements can already beat off-the-shelf software.
AI-processed from Habr AI; edited by Hamidun News
You import a batch of photos from your iPhone in HEIC format (the default for iPhone) and try to flip through them on a modern MacBook. The built-in Preview app and third-party viewers show the next image with a noticeable delay, and they don't launch instantly either. People used to put up with this: nobody would write a custom viewer just to save a few tens of milliseconds.
The author of the Habr article wanted to check whether vibe coding changes that. Over three evenings, using the Fable 5.2 model and Opus 5.
5 sub-agents, they built Bolt, a native viewer in Swift and AppKit with no third-party dependencies. In their measurements, Bolt shows the next photo in 31 ms (the median from arrow key press to sharp image, with presses one second apart), while the closest competitor, ApolloOne, takes 87 ms. Below is how the measurements were made and where the speed comes from.
What Bolt does and how it was compared
The author was inspired by IrfanView, familiar from Windows: open one photo and page through the whole folder from the keyboard. Delete moves the photo to the Trash and immediately shows the next one. There is no confirmation dialog, but any action can be undone with ⌘Z. Keys 1 through 9 move photos into assigned folders, and square brackets rotate the image by changing only the Orientation tag, leaving the pixels untouched. It supports preset-based export, batch conversion, resizing, template-based renaming, and a grid of all photos in the folder. Bolt does not replace a cataloguer like Lightroom or a full editor; it works at the folder level.
You cannot ask another program to report when it has displayed a frame. So the benchmark records the screen with ScreenCaptureKit, presses the keys itself, and searches the recording for the moment a sharp photo appears. The comparison covered Bolt, ApolloOne 5.1 (trial version), XnView MP 1.12.2, Preview 11, and Quick Look, using 202 iPhone 13 photos. 193 of them contain a gain map, a second image inside the HEIC that indicates where to raise brightness so the photo looks like HDR. The machine was the same throughout: a 15-inch MacBook Air with M5, 16 GB of memory, and a 60 Hz display, with three rounds per program.
With presses one second apart, the medians were: Bolt 31 ms, ApolloOne 87, XnView MP 116, Preview 140, Quick Look 172. When the arrow key is held for three seconds, Bolt shows 10.3 photos per second and all frames are sharp; ApolloOne manages 6.2 photos per second. Preview keeps only 6% of its frames sharp, and XnView MP failed to change the image on 59 of 90 presses.
Where the milliseconds go
It takes 41 ms from the key event until the system displays the frame, and only 1.2 ms of that is Bolt's code. On average, about 8 ms more is spent with the frame waiting for the deadline at which WindowServer assembles the next screen frame, and 31–34 ms passes from that deadline to display (roughly two frames at 60 Hz). The agent tried output through Metal instead of contents: it took 43.4 ms versus 41.4, so there was no gain. Pre-placing the neighboring frames shortened the transaction by 0.7 ms, but changed nothing on screen.
One improvement did work. If the photo, window title, file icon, and status bar are sent in a single Core Animation transaction, the image reaches the screen in 54.1 ms, because AppKit lays out and draws the text inside that transaction. When the photo goes in a separate transaction and the title and status bar follow, it takes 41.4 ms. The main culprit is the window title: without it, the photo appears 9.5 ms sooner.
What remains is keeping the next frame ready in cache, and that is the decoder's job. ImageIO, the standard way to read images on a Mac, decodes a non-HDR photo in 45–47 ms and an iPhone photo with a gain map in 140–151 ms, and it cannot produce more than 11 photos per second regardless of thread count. Profiling showed that HEVC itself is decoded in hardware; the time is lost around it.
A 12-megapixel photo consists of 48 tiles of 512 × 512, and for each tile CMPhoto creates a new pixel conversion session. A microbenchmark: 48 tiles with a new session each took 23.4 ms, and with a shared session 11.
0 ms. Core Image receives the whole photo and processes it on the GPU in a single pass, so decoding one iPhone photo sped up from 151 to 17 ms.
- Nov 21, 2026Nectar launches AI agent that finds influencers among 300 million profiles
- Nov 21, 2026How to Test UIs with an AI Agent: 5 Rules from the Creators of e2e, Which Has 13.8K Stars
- Nov 21, 2026opencode Hits 226K GitHub Stars: A Free 10-Tool AI Stack
- Nov 21, 2026GitHub AI Security Catalog Gets 231 Pull Requests in Six Months, 157 Accepted
Key Questions Answered
What Bolt does and how it was compared?
The author was inspired by IrfanView, familiar from Windows: open one photo and page through the whole folder from the keyboard. Delete moves the photo to the Trash and immediately shows the next one. There is no confirmation dialog, but any action can be undone with ⌘Z. Keys 1 through 9 move photos into assigned folders, and square brackets rotate the image by changing only the Orientation tag, leaving the pixels untouched.
Where the milliseconds go?
It takes 41 ms from the key event until the system displays the frame, and only 1.2 ms of that is Bolt's code. On average, about 8 ms more is spent with the frame waiting for the deadline at which WindowServer assembles the next screen frame, and 31–34 ms passes from that deadline to display (roughly two frames at 60 Hz). The agent tried output through Metal instead of contents: it took 43.4 ms versus 41.4, so there was no gain.
Test Frontier LLMs & Autonomous Agents in One Workspace
Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.