Update README.md
#2
by Ramkasuru - opened
No description provided.
README: Cube3D-style header, define HQQ before first use
Restyles the top of the README to match the efficient-cube3d card, as requested.
Changes (README.md only, nothing below "Efficiency" touched):
- Emoji title, 4 shields.io badges, bold intro paragraph, emoji metrics table, highlighted note.
- HQQ is now defined at first mention: Half-Quadratic Quantization, a calibration-free weight-only method, W4A16 (4-bit weights, 16-bit activations), group size 128.
- Keeps your front matter edits (base_model_relation, pipeline_tag, extra tags) and the OpenRAIL license.
All numbers are unchanged from the merged README:
- Diffusion transformer 8.58 GB -> 1.48 GB (82.8% smaller)
- Peak VRAM 33.71 GiB -> 27.09 GiB (19.6% lower)
- Median part chamfer distance vs BF16: 0.0114 (828 parts, 250 prompts)
- Latency 34.7 s -> 66.5 s (eager mode, torch.compile disabled)
Differences from the Cube3D card, on purpose:
- Latency badge says slower, not "on par", because that is what we measured.
- No GitHub or Colab badges yet. Happy to add them if there is a CubePart repo or notebook to link.
Ramkasuru changed pull request status to closed