Add shaded BF16 vs INT4 part renders after the first table

.gitattributes

@@ -33,3 +33,6 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+assets/sh_renders/191_figure.png filter=lfs diff=lfs merge=lfs -text
+assets/sh_renders/305_figure.png filter=lfs diff=lfs merge=lfs -text
+assets/sh_renders/020_figure.png filter=lfs diff=lfs merge=lfs -text

README.md

@@ -38,6 +38,16 @@ Companion to [TrNi/efficient-cube3d](https://huggingface.co/TrNi/efficient-cube3
 <mark>💡 The diffusion transformer is 82.8% smaller on disk and part decomposition stays near lossless on the typical part. Peak VRAM drops less (19.6%) because the Qwen3-VL text encoder and the shape VAE stay in BF16. INT4 is slower here because CubePart needs <code>torch.compile</code> disabled, so dequantization runs in eager mode.
 </mark>
 
+## BF16 vs INT4 renders
+
+Full mesh on top, CubePart parts below. Rendered with spherical harmonic lighting, Blinn-Phong specular and ambient occlusion: gray only, no color, so the geometry itself shows.
+
+![A pineapple with a crown of 8 spiky leaves and a hexagonal surface pattern](assets/sh_renders/191_figure.png)
+
+![steampunk boombox with 2 circular speakers and 6 brass dials](assets/sh_renders/305_figure.png)
+
+![A sitting cat with pointed ears and a long tail](assets/sh_renders/020_figure.png)
+
 ## Efficiency
 
 Measured on one NVIDIA L40S (44 GiB usable), PyTorch 2.8.0+cu128, torchao 0.11.0, eager mode, 50 diffusion steps, guidance scale 7.5, over 10 meshes, one from each of 10 categories. Two independent runs; ranges below are run 1 to run 2.