Memelith is a local-first desktop library for collecting, organizing, and finding memes. It keeps media in a storage directory chosen by the user and stores the searchable index in SQLite, so the library remains usable without a hosted service.
- Collect images or text into an inbox before deciding how to organize them.
- Detect exact duplicates with content hashes and flag visually similar images with local CLIP embeddings. Similarity suggestions can be dismissed when an item is intentionally different.
- Save memes with a name, description, tags, and one or more image, animated media, or text contents.
- Organize the library into editable MemePacks and browse all saved memes from one searchable view.
- Search meme names, descriptions, tags, content types, and text contents with field-aware expressions.
- Analyze images with Waifu Sensor to suggest anime-character tags while adding a meme.
- Receive images and text through a Telegram Bot and synchronize Telegram sticker packs into MemePacks. Telegram integration is optional and disabled until configured in Settings.
- Keep the meme library's media and SQLite index under a user-selected local storage root.
Release workflows produce a macOS DMG, Linux packages, and Windows installers or a portable executable. Download the artifact for your platform from the project's release page, then launch Memelith and choose a storage directory in Settings.
The macOS application bundle does not include the ONNX model binaries. On the
first launch, Memelith downloads about 992 MiB of pinned model files, verifies
their byte sizes and SHA-256 digests, and stores them under
~/Library/Application Support/memelith/models. A network connection is
required only for this initial setup (or when a cached model is missing or
invalid). Completed installations use a lightweight metadata receipt on later
launches instead of hashing every model again, and interrupted downloads resume
from the saved partial file. The exact sources and checksums are documented in
docs/onnx-model-sources.md.
- Open Collector and paste text or select one or more images.
- Review duplicate warnings, dismiss false similarity matches, and select the items to keep.
- Promote the selection to a MemePack, add a name or description, and attach tags. The Add page can also create a meme directly from new content.
- Use All to search and browse the library, or open MemePack to manage collections. Configure the optional Telegram Bot and storage root in Settings.
This is a Rust workspace. When flake.nix is available, enter the development
shell before running the commands below:
nix developRun the GUI locally:
cargo run -p memelithRun all workspace checks and tests:
cargo test --workspaceBuild the native macOS application bundle from the workspace root:
nix develop --command scripts/package-macos.shThe application is written to
target/macos-bundle/release/bundle/osx/Memelith.app. The script bundles
non-system dynamic libraries and applies an ad-hoc signature by default. Set
CODESIGN_IDENTITY to use an installed signing identity instead. The generated
arm64 bundle requires macOS 14 or later.
The workspace enables GPUI's runtime_shaders feature so macOS development does
not require the Metal command-line compiler. Release packaging can disable this
feature when a full Xcode installation is available.
apps/gui- GPUI desktop application, model setup, search, and Telegram integration.crates/core- SQLite-backed meme, MemePack, tag, Collector, and duplicate detection logic.crates/clip- local FP32 ONNX text and image encoders in a shared 512-dimensional embedding space.crates/waifu-sensor- explainable anime-character retrieval and its optional CLI.crates/character-segmentation- AnimeInsSeg and YOLO11 ONNX segmentation tools with cross-model duplicate suppression.crates/liquid-glass- reusable GPUI visual components.
Build the independent Waifu Sensor CLI:
cargo build -p waifu-sensor --features cli --bin waifu-sensorRun local YOLO11n segmentation after exporting a model to ONNX:
cargo run -p character-segmentation --bin yolo11-seg -- \
--model /path/to/yolo11n-seg.onnx \
--image input.png \
--output-dir output \
--class-label personRun both local character-segmentation backends on the same image:
cargo run -p character-segmentation --release -- \
--anime-model /path/to/rtmdetl_e60.raw.640.onnx \
--yolo-model /path/to/yolo11n-seg.onnx \
--image input.png \
--output-dir output \
--merge-iou 0.6This writes the untouched animeinsseg/ and yolo11n-seg/ results plus a
merged/ directory where overlapping predictions retain the
higher-confidence instance.
Workspace code is licensed under AGPL-3.0-or-later. The bundled and
downloaded model assets have separate upstream licensing and redistribution
terms; see docs/onnx-model-sources.md and the
model metadata before redistributing them.