Open-source AI photo culling for Mac and Windows.
Drop a shoot: blurry shots, closed eyes and duplicates are set aside, the best photos are kept. On your computer.
Website · Download · Docs · Contributing · Français
Pelli Instant sorts a photo shoot for you. Drop a folder of RAW or JPEG files: it flags blurry and shaken photos, closed eyes, bad exposure and duplicates, keeps the best shot of each burst, and splits the shoot into Selected, To review and Rejected. You go through its choices with the keyboard, change what you disagree with, and export: star ratings and colour labels for Lightroom, Capture One or Bridge, a copy of your selection, or the rejects filed away in a sub-folder.
Everything runs on your computer. Sorting never sends your photos anywhere: the AI models run locally with ONNX Runtime, on the processor of Apple Silicon Macs (the fastest there) and accelerated by DirectML on Windows. No telemetry.
Pelli Instant is made by the team behind Pelli, the platform where photographers deliver and sell their work. The app is free and open source. The official builds ask for a Pelli account, free to create.
Looking for an open-source alternative to Aftershoot, Narrative Select or FilterPixel, one that keeps your photos on your computer? That's what Pelli Instant is for.
- Drop a shoot. Drag a folder onto the window, or choose one, with or without its sub-folders. A RAW and its JPEG twin count as one photo.
- Let it sort. Before the first analysis, Pelli Instant shows what it will check; turn off what you don't want, or just start. It reads the preview your camera embedded in each RAW (it never develops the RAW), measures focus, motion blur and exposure, finds faces and closed eyes, and groups bursts. The grid fills in as it goes.
- Review. Selected, To review, Rejected: each rejection says why (Blurry, Motion blur, Eyes closed, Duplicate…). Fix any call in one key. Your own choices always win over the AI's.
- Export. Ratings for Lightroom, a copy of the selection, or the rejects moved aside. Nothing is ever deleted.
Photos by Flickr photographers under CC BY 2.0 and CC0: credits.
| Feature | What it does |
|---|---|
| Focus you can trust | Measures how wide the edges are, not how contrasted: shallow depth of field, high ISO and soft light are not mistaken for blur. When there is a face, its focus comes first: a sharp face on creamy bokeh is sharp, a sharp background behind a soft face is Subject out of focus. |
| Motion blur | Tells camera shake and moving subjects from oriented textures such as hair, fur or fabric. |
| Closed eyes | Reads the eyes of the faces in the photo, and tells a blink from a smile that narrows the eyes. On Balanced and Lenient, closed eyes send a photo to review, never to the rejects on their own: a kiss or a prayer closes them on purpose. |
| Exposure in stops | Estimates how many stops a photo is off. Low-key and high-key photos made on purpose stay fine; 2 stops off usually goes to review, since a RAW can recover it. |
| Bursts and duplicates | Groups similar shots taken seconds apart by the same camera and keeps the best one, or the best few. Exact copies are found whatever their date. |
| Three sensitivities | Strict, Balanced or Lenient. Switching re-sorts the shoot instantly, without analysing it again. |
| Your criteria | Before the first analysis of a shoot, turn each check on or off: blur, motion blur, subject focus, closed eyes, under- and overexposure, similar shots, exact copies. Respect artistic blur is there for bokeh and blur on purpose. Each shoot keeps its own criteria; the defaults are in Settings → Analysis. |
| Standard or Maximum precision | Standard comes with the app and is fast. Maximum is an optional download of bigger models (about 777 MB), from Settings → Analysis: finer choices on group photos, bursts and subjects without a face, for a slower analysis. Until it is installed, analyses run in Standard. |
| Review at the keyboard | Grid, loupe with zoom at 100 %, and side-by-side compare of a burst. All photos or one decision at a time, filtered by reason (Blurry, Eyes closed…). In the loupe, ↑ keeps and ↓ rejects, ⌥← / ⌥→ step through a burst; the Lightroom keys P, X and U work everywhere. |
| AI search | Type "bride bouquet" or "first dance" and the matching photos come first. Common wedding, event and portrait words, colours and clothing ("robe bleue", "haut rouge") also work in French, Spanish, German and Italian. |
| Ratings for Lightroom | Writes stars and colour labels to an .xmp file next to each RAW, for Lightroom, Capture One, Bridge or darktable. You choose the stars and colour of each decision in the export (Selected ★★★ green, To review yellow, Rejected by default). Stars you gave in the camera count as yours, an existing .xmp keeps all its settings, and stars and labels from another app stay unless you change them in Pelli Instant. |
| Copy the selection | RAW, JPEG or both, to the folder of your choice, sub-folders kept. Pelli Instant tells you how much disk space the copy needs. Each file is copied whole or not at all, and exporting again doesn't copy a photo twice. |
| Rejects filed away | Moves the rejects to a _Rejets Pelli sub-folder next to the originals: a RAW, its JPEG and its .xmp move together. Nothing is deleted. |
| Gives way to your other apps | A switch in Settings → General runs the analysis in the background, on fewer cores at a low priority: Lightroom and your browser stay smooth, sorting takes 3 to 5 times longer. Off by default, for the fastest sort. |
| Works while you're away | An analysis, an export or the Maximum download keeps the computer from going to sleep, and on a Mac keeps full speed with the window hidden. The screen can still turn off. |
| Most cameras | Canon CR2 and CR3, Nikon NEF, Sony ARW, Fujifilm RAF, DNG, Olympus and OM System ORF, Panasonic RW2, Pentax PEF, Samsung SRW, plus JPEG, PNG and TIFF. |
| Picks up where you left off | Your projects stay on the home screen. Opening a folder again only analyses new or changed photos. |
| A guided start | A short tour on first launch, with a voice-over you can mute: how it judges a photo, the keys, search, the exports, the precision and what your Pelli plan includes. Replay it from Settings → About. |
| Made for judging photos | A dark, neutral interface, like Lightroom, so you judge colours on a neutral surround. In 9 languages. |
| Key | Action |
|---|---|
| ← → | Previous / next photo |
| ⌥ ← → | Previous / next photo of the burst |
| ↑ / ↓ (loupe) | Keep / reject |
| P / X / U | Keep / reject / remove the flag, everywhere |
| 1–5 | Stars |
| 6–9 | Red, yellow, green, blue label |
| Space | Loupe |
| C | Compare |
| Z | Zoom 100 % |
| ⌘ F | AI search |
| ⌘ E | Export |
| ? | All shortcuts |
On Windows, Ctrl replaces ⌘.
Pelli Instant is free. The official builds ask you to sign in with a Pelli account, free to create: that is the only requirement. How many photos you can sort each month depends on your Pelli plan.
| Pelli plan | Photos sorted per month |
|---|---|
| Free | 200 |
| Lite | 1,000 |
| Pro | Unlimited |
| Ultra | Unlimited |
| Free trial | Unlimited |
- A RAW and its JPEG twin count as one photo.
- Sorting photos again doesn't count: changing the sensitivity or the precision, reopening a project or analysing a folder a second time only counts the photos it has never seen.
- The count starts again on the first day of each month (UTC).
Prices and what each plan includes come from Pelli, in the app as on pelli.io. Pelli plans also cover what Pelli does on the web: client galleries, online sales, quotes and invoices, bookings.
Note
Building it yourself? Pelli Instant is MIT-licensed, and you can switch the account and the quota off when you compile it: PELLI_INSTANT_NO_ACCOUNT=1 pnpm tauri dev, or the same variable in front of pnpm tauri build (see Building an installer yourself). The official builds, signed and with the account, are how Pelli funds the project.
| Platform | Latest version |
|---|---|
| macOS 12 or later, Apple Silicon (M1 or newer) | PelliInstant-macos-arm64.dmg |
| Windows 10 and 11 (x64) | PelliInstant-windows-x64-setup.exe |
The Standard models come inside the installer, so Pelli Instant works as soon as it is installed; the Maximum pack is downloaded from the app, only if you choose it. Each download has its SHA-256 fingerprint next to it (.sha256), and every release built by the release workflow comes with a SHA256SUMS file and a build attestation you can check with gh attestation verify <file> --repo Synapsr/Pelli-Instant.
Intel Macs
There is no build for Intel Macs: ONNX Runtime, which runs the AI models, ships no prebuilt library for them. Building ONNX Runtime yourself is possible; see crates/cull-ml/README.md.
Windows SmartScreen warning
The Windows installer isn't signed with a code-signing certificate yet, so SmartScreen may warn about an unknown publisher. Click More info → Run anyway. You can check the installer against its .sha256 first. It installs for your user only and doesn't ask for administrator rights. Its updates are signed with Pelli Instant's update key, which the app checks before installing them.
Linux
There is no Linux build. The engine compiles and its tests run on Linux in CI, and pnpm tauri dev may start, but the app is neither packaged nor tested there. On Linux, the sign-in token would be kept in the app's local database rather than in a system keychain.
macOS folder access
If macOS refuses access to a folder (Desktop, Documents, an external drive), allow Pelli Instant in System Settings → Privacy & Security → Files and Folders, then open the folder again.
Pelli Instant updates itself. A few seconds after launch, then every six hours, it checks the latest release; when there is a newer version, the window offers it. Install it in one click: the update is downloaded, its signature is checked against Pelli Instant's own key, and the app restarts on the new version once no analysis, export or download is under way. Settings → About checks by hand. Prereleases are never offered.
Requires rustup (it installs the Rust version pinned in rust-toolchain.toml), Node.js 22 and pnpm 10, on an Apple Silicon Mac or a Windows x64 PC.
git clone https://github.com/Synapsr/Pelli-Instant.git
cd Pelli-Instant
pnpm install
pnpm models # download the Standard models once (about 223 MB, SHA-256 checked)
pnpm tauri dev # run the app in developmentPELLI_INSTANT_NO_ACCOUNT=1 pnpm tauri dev runs it without a Pelli account and without the quota. To work on the interface alone, run pnpm dev and open http://localhost:1420: a simulated engine in the browser, no Rust build needed. Building an installer with the models inside, and everything else, is in docs/contributing.md.
- Your photos are sorted on your computer. Pelli Instant reads them where they are; sorting never sends them anywhere. Its measures, decisions and your marks are kept in a local database, its previews in a cache folder, both in your user account (where).
- Your folders stay untouched until you export. Exporting writes
.xmpfiles next to your RAWs, copies the selection where you say, or moves the rejects to_Rejets Pelli. Your photo files are never modified or deleted. - What Pelli receives. With an account, the app talks to Pelli (
app.pelli.io) for three things. Signing you in, in your browser: Pelli receives your computer's name, shown on the consent page; the token is then kept in the macOS Keychain or the Windows Credential Manager. Reading your plan and Pelli's current plans. Counting the photos you sort: before an analysis, the number of photos it sorts for the first time; after it, the number it couldn't sort, given back. Each request also carries the app version, the system (macOS or Windows) and the interface language. For the account and the quota, Pelli never receives your photos, their file names or their content. - What the app downloads. The Maximum pack, only when you choose it: from Hugging Face (
huggingface.co) and Pelli's storage (pelli.s3.fr-par.scw.cloud), each file checked against its SHA-256. Updates: the app readslatest.jsonfrom the latest release on GitHub, and downloads an update only when you install it. - The AI models are fixed files at pinned versions, and their SHA-256 fingerprints are checked when they load. Built from source,
pnpm modelsdownloads them once from Hugging Face and the other sources listed inmodels/manifest.json. - No analytics, no telemetry, no crash reporter. ONNX Runtime's own telemetry is switched off.
- Without an account (
PELLI_INSTANT_NO_ACCOUNT=1builds), Pelli Instant never talks toapp.pelli.io. The Maximum pack, if you download it, still comes partly from Pelli's storage. - Signing out deletes the token from your computer; your projects stay. Removing a project from the home screen deletes its previews and analysis, never your photos. Deleting the Maximum pack (Settings → Analysis) frees its 777 MB.
Interface: English, French, Spanish, German, Italian, Portuguese, Dutch, Polish and Swedish, with the welcome tour narrated in each of them. Add yours.
AI search: the model understands English. A built-in glossary translates common wedding, event and portrait words, colours, clothing and places from French, Spanish, German and Italian; other words pass through as typed.
Is Pelli Instant free?
Yes. The app is free and open source under the MIT licence. The official builds need a Pelli account, free to create, and sort 200 photos a month on the Free plan, 1,000 on Lite, and as many as you like on Pro, Ultra or during the free trial. If you build it yourself, you can turn the account off.
Why do the official builds need a Pelli account?
Because that is how Pelli pays for the work on Pelli Instant: the monthly quota is tied to Pelli's plans. The account only serves to sign you in and count photos. The code stays open, and anyone can build it without the account.
Standard or Maximum precision?
Standard comes with the app, is fast, and suits most shoots. Maximum adds bigger models, about 777 MB to download once: a second face detector for group photos, a model that tells bursts and near-duplicates apart from different scenes more clearly, a bigger CLIP model for search and the aesthetic score, and the outline of people and animals, so a subject without a visible face (seen from behind, back-lit) is judged on its own focus and exposure. The analysis is slower and needs more memory.
Choose it in Settings → Analysis → Precision (or in the welcome tour): the download starts in the background, resumes where it stopped if you pause it or quit, and every file is checked before use. Until it is ready, analyses run in Standard. A shoot already sorted can be analysed again in Maximum; photos sorted again never count on your quota.
Does it work offline?
The analysis runs entirely on your computer, without the internet. Signing in needs a connection. On Pro, Ultra or a free trial, you can sort offline. On the Free and Lite plans, Pelli Instant needs a connection to count new photos before it sorts them; reviewing and exporting a shoot already sorted works offline. Downloading the Maximum pack and updates needs a connection too.
Does it change my RAW files?
No. Stars and colour labels go into an .xmp sidecar next to each RAW (IMG_0001.xmp, or IMG_0001.CR3.xmp when that one already exists), and every other setting in an existing sidecar is kept: develop settings, keywords, history. Stars and labels you set in Lightroom or Capture One stay too, unless you set them in Pelli Instant: the ratings that come from the AI's decisions only replace the ones Pelli Instant wrote itself. Stars you gave in the camera become your own stars in Pelli Instant (unless the RAW already has a sidecar), so the export keeps them.
In Lightroom Classic, select the photos, then Metadata → Read Metadata from Files. Lightroom has no XMP field for its Pick and Reject flags, so a rejected photo without stars gets −1 star, which Lightroom and Bridge show as Rejected. JPEG-only shoots get no sidecar, since Lightroom reads a JPEG's ratings inside the file.
What happens to my files when I copy the selection or file away the rejects?
Copies go through a hidden temporary file, renamed once complete: a full disk or an unplugged drive never leaves a half-copied photo. A file already copied by an earlier export is left as it is, and when a name is taken by another file, the photo gets one suffix for all its files (IMG_0001-1.CR3, IMG_0001-1.JPG, IMG_0001-1.xmp).
Rejects move with their JPEG twin and their sidecar, all together or not at all: if one file can't move, the others go back. Across disks, a file is copied and checked before the original is removed.
The folders an export fills (the rejects folder, a new or empty copy folder) get a small hidden file, .pelli-instant-export, so that a copy made inside the shoot is never analysed again as new photos. Delete that file to analyse them.
Which cameras are supported?
Pelli Instant reads the JPEG preview that the camera embeds in each RAW: Canon CR2 and CR3, Nikon NEF and NRW, Sony ARW, SR2 and SRF, Fujifilm RAF, DNG, Olympus and OM System ORF, Panasonic RW2, Pentax PEF, Samsung SRW, and other TIFF-based RAWs, plus JPEG, PNG and TIFF files. Not supported: HEIF/HEIC, and the CR3 files of recent Canon bodies shot in HDR PQ mode, whose previews are HEVC. The full list, with the cameras tested, is in crates/cull-decode/README.md.
How does it decide?
Every photo gets a score from 0 to 100, which mixes focus, an aesthetic score made relative to your shoot, eyes, exposure and uniqueness. A clear technical problem rejects a photo; a mild one sends it to review. The thresholds are relative to your shoot as well as absolute, so a dim reception at high ISO is not mass-rejected. The details, written for photographers, are in crates/cull-analysis/README.md.
Are my photos used to train AI models?
No. Sorting runs on your computer and never sends your photos, so neither Pelli nor anyone else receives them to train anything. The models are fixed files; they don't learn from your shoots.
Why no Intel Mac or Linux version?
Intel Macs: ONNX Runtime has no prebuilt library for them. Linux: the app isn't packaged or tested there yet. Both are open to contributions.
Bug reports, translations and pull requests are welcome. Start with docs/contributing.md (setup, checks, conventions), then read the architecture and the models. Pelli Instant aims to stay quick and trustworthy: every rejection gives its reason, and before adding a setting, ask whether a better default would do.
- Report a bug
- Suggest an improvement
- Found a security issue? See SECURITY.md.
- Anything else: contact@pelli.io
If Pelli Instant saves you an evening of culling, a star helps other photographers find it.
Pelli Instant runs YuNet (OpenCV Zoo) to find faces, MediaPipe Face Mesh and Face Blendshapes (Google; Face Mesh converted to ONNX by PINTO0309) to read eyes and expressions, CLIP (OpenAI, ONNX by Xenova) for similar shots and search, and the LAION aesthetic predictor. Maximum precision adds CLIP ViT-L/14, DINOv2 (Meta, ONNX by onnx-community), RF-DETR (Roboflow), RetinaFace (biubug6) and the improved aesthetic predictor (Christoph Schuhmann). All run through ONNX Runtime and ort. The app is built with Tauri, React and HeroUI. The welcome tour's photos and voice-over are credited in src/assets/onboarding/CREDITS.md. See THIRD_PARTY_NOTICES.md.
MIT © Synapsr