protecting your timeline from the egregore since 2026
A Chrome extension that detects Milady-style avatars on Twitter/X and lets you hide or fade matching posts. Runs entirely in your browser.
Why
Some people find that a significant percentage of their timeline consists of accounts using aesthetically identical chibi avatars posting aesthetically identical content. This extension addresses that.
How It Works
A bundled ONNX classifier scans avatars as you scroll. When it spots a match, you pick what happens:
- Hide — collapsed behind a click-to-reveal row.
- Fade — visible but at half opacity.
- Debug — borders and confidence scores on every post.
- Off — does nothing.
The popup tracks session stats (posts scanned, match rate, last sighting), keeps a list of detected accounts you can exempt individually, and collects avatar data you can export for offline labeling.
Everything runs locally. No server calls, no telemetry, nothing leaves your browser unless you explicitly export it.
Install
There is no Chrome Web Store release. Install from GitHub instead:
- Download the latest
milady-shrinkifier-*-unpacked.zipfrom the Releases page. - Unzip it somewhere permanent on disk.
- Open
chrome://extensions. - Enable
Developer mode. - Click
Load unpackedand select the unzipped folder, or drag and drop it onto the page.
Accuracy
All scores below come from the full exported avatar corpus.
- Precision — when the extension filters a post, how often it's right.
- Recall — of the Milady-style avatars in the evaluation set, how many it catches.
- Evaluation corpus —
14,869exported avatars (1,293milady,13,576not_milady). - This is broader than the blind split and still fully human-labeled, but it is not a blind benchmark.
| Version | Run | Training mix | Precision | Recall |
|---|---|---|---|---|
v0.2.2 |
20260327T142224Z |
Milady Maker + 2,596 manually reviewed avatars |
0.9960 |
0.5839 |
v0.3.0 |
20260327T212453Z |
+ Remilio, Pixelady + 2,967 manually reviewed avatars |
0.9956 |
0.6930 |
v0.4.0 |
20260328T144735Z |
+ 5,715 manually reviewed avatars |
0.9950 |
0.7695 |
v0.5.0 |
20260328T223931Z |
+ 6,773 manually reviewed avatars |
0.9951 |
0.7827 |
v0.6.0 |
20260329T124912Z |
+ 7,370 manually reviewed avatars |
0.9952 |
0.7974 |
v0.7.0 |
20260329T181946Z |
+ 7,695 manually reviewed avatars |
0.9954 |
0.8306 |
v0.8.0 |
20260329T220050Z |
+ 9,394 manually reviewed avatars |
0.9956 |
0.8801 |
v0.9.0 |
20260330Tlr1e4 |
+ 10,943 manually reviewed avatars |
0.9957 |
0.8894 |
v0.10.0 |
20260401T145142Z |
+ 14,869 manually reviewed avatars |
0.9952 |
0.9582 |
v0.11.0 |
20260402Trefit |
+ 14,869 manually reviewed avatars |
0.9969 |
0.9992 |
Rows through v0.10.0 use the same exported evaluation corpus and are directly comparable.
v0.11.0 is a final refit on train+val with checkpoint selection on test, so its exported-corpus metrics are not a blind comparison to earlier research checkpoints.
Development
See DEVELOPMENT.md for build commands and debugging.
For a user-facing walkthrough of how the offline training loop works end to end, see docs/training-pipeline.md.