Update README.md

README.md

@@ -3,8 +3,17 @@ license: agpl-3.0
 pipeline_tag: image-classification
 ---
 
-A tiny 230K param, opinionated image compression classifier. CNN. Trained on JPEG, JXL, AVIF, WEBP and PNG. Hits 94.3% accuracy on the task.
+A small 2M param CNN classifier to predict the quality factor of lossy image codecs. Covers JPEG, JXL, AVIF and WEBP. Dataset (45k) included RAW photography and PNG illustrations.
 
-Attempts to flags images below q=80 for JPEG and below 65 for the rest. This was driven by research showing image models trained on many q=75 JPEGS had massive performance penalties.
+Overall accuracy was 95.3% on train, 96.6% on val (7% split). Take the predictions with a large grain of salt, its not yet ready for production.
 
-Training was 128x128 patches + multi-scale high freq residuals. DCT aware 8x8 stem and dual heads (binary acceptable/not + codec classifier).
+This model will *not* work properly in the following scenarios:
+
+- Really small images (sub 512px)
+- Mutiple resize and/or compression loops (would have ballooned training time for first iteration)
+- <insert lossy format> renamed to PNG's
+- Mixed media; low quality JPEG background with an overlaid illustration.
+
+Motivated by research showing image models trained on many <= q=75 JPEGS had negative outcomes.
+
+Credit: Rimuru original model and original code