Ramkasuru commited on
Commit
8734e27
ยท
verified ยท
1 Parent(s): fcea1f2

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +22 -10
README.md CHANGED
@@ -15,16 +15,28 @@ tags:
15
  - diffusion
16
  ---
17
 
18
- # CubePart INT4 (HQQ W4A16, group size 128)
19
-
20
- INT4 weight-only quantization of the diffusion transformer in Roblox's [CubePart](https://huggingface.co/Roblox/cubepart) (open-vocabulary, part-controllable 3D mesh generation, [arXiv 2605.28763](https://arxiv.org/abs/2605.28763), SIGGRAPH 2026). Companion to [TrNi/efficient-cube3d](https://huggingface.co/TrNi/efficient-cube3d), which does the same for Cube3D.
21
-
22
- ## Summary
23
-
24
- - Checkpoint size: 1.479 GB, versus 8.582 GB for the BF16 diffusion transformer (82.8% smaller).
25
- - Peak VRAM during generation: 27.09 GiB, versus 33.71 GiB for BF16 (19.6% lower).
26
- - Quality: median chamfer distance between BF16 and INT4 part outputs is 0.0114 per part, which is near lossless on the typical part. Errors concentrate in a few small, thin parts.
27
- - Latency: INT4 is slower than BF16 in this release (66.5 to 66.9 s versus 34.7 to 34.8 s per generation). `torch.compile` has to be disabled for CubePart, so INT4 dequantization runs in eager mode.
 
 
 
 
 
 
 
 
 
 
 
 
28
 
29
  ## Efficiency
30
 
 
15
  - diffusion
16
  ---
17
 
18
+ # ๐Ÿš€ First INT4 Quantized Efficient-CubePart3D - 83% Smaller Part Decomposition Model
19
+
20
+ ![](https://img.shields.io/badge/Model_Size-1.48_GB_(82.8%25โ†“)-blueviolet)
21
+ ![](https://img.shields.io/badge/Peak_VRAM-27.09_GiB_(19.6%25โ†“)-brightgreen)
22
+ ![](https://img.shields.io/badge/Part_CD_median-0.0114_(near_lossless)-blue)
23
+ ![](https://img.shields.io/badge/Latency-66.5s_(eager_mode,_1.9x_slower)-orange)
24
+
25
+ Presenting the **first INT4 quantized efficient version** of [CubePart](https://huggingface.co/Roblox/cubepart) ([arXiv 2605.28763](https://arxiv.org/abs/2605.28763), SIGGRAPH 2026), an open-vocabulary, part-controllable 3D mesh generative model.
26
+ Quantized via **HQQ (Half-Quadratic Quantization)**, a calibration-free weight-only method, at **W4A16** (4-bit weights, 16-bit activations, group_size=128) using [torchao](https://github.com/pytorch/ao),
27
+ it cuts the diffusion transformer from **8.58 GB โ†’ 1.48 GB (82.8%โ†“)** and peak VRAM from **33.71 GiB โ†’ 27.09 GiB (19.6%โ†“)** while keeping part decomposition near lossless (median chamfer distance **0.0114** per part over 828 parts).
28
+ Companion to [TrNi/efficient-cube3d](https://huggingface.co/TrNi/efficient-cube3d), which does the same for Cube3D.
29
+
30
+ | | BF16 | **INT4 (HQQ W4A16)** |
31
+ |---|:-:|:-:|
32
+ | ๐Ÿ’พ Model size (diffusion transformer) | 8.58 GB | **1.48 GB (82.8%โ†“)** |
33
+ | ๐ŸŽฎ Peak VRAM | 33.71 GiB | **27.09 GiB (19.6%โ†“)** โœจ |
34
+ | ๐Ÿ“ฆ Setup time | 161.8 s | **166.1 s** (BF16 load + 4.3 s quantize) |
35
+ | โฑ๏ธ Latency | 34.7 s | **66.5 s** (eager mode) |
36
+ | ๐ŸŽฏ Part chamfer distance vs BF16, median | n/a | **0.0114** |
37
+
38
+ <mark>๐Ÿ’ก 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.
39
+ </mark>
40
 
41
  ## Efficiency
42