Embed hero image

README.md

@@ -17,6 +17,8 @@ convolutional halves run on the `CompiledModel` **GPU** delegate (`LITERT_CL`);
 Transformers and the split RVQ run on **CPU**. Device-verified on a Pixel 8a (Tensor G3): full
 round-trip at **RTF ≈ 0.35** (faster than real-time), reconstruction at the codec's quality floor.
 
+![Mimi — original vs 32-codebook RVQ round-trip on-device (LiteRT)](samples/sample.png)
+
 ## Files
 
 | File | Size | Delegate | In → Out |