MASQ: Accelerating Masked Diffusion via Stage-Wise Multi-Precision Quantization
Fuente:
arXiv
Saved in:
| Main Authors: | , , , |
|---|---|
| Format: | Preprint |
| Published: |
2026
|
| Subjects: | |
| Online Access: | |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| _version_ | 1866913154870542336 |
|---|---|
| author | Kim, Seeyeon Lee, Jaehun Yoo, Sungyeob Kim, Joo-Young |
| author_facet | Kim, Seeyeon Lee, Jaehun Yoo, Sungyeob Kim, Joo-Young |
| contents | Masked diffusion enables region-specific image synthesis but suffers from computational redundancy, since the entire image is processed each timestep even though only the masked region requires generation. To address this, we introduce MASQ, a hardware-software co-designed accelerator for masked diffusion. Our approach performs stage-wise MXINT8/4/2 precision assignment that dynamically reflects spatial and semantic importance, complemented by timestep-aware scheduling and optimized non-matrix operations. MASQ features a block-wise multi-precision compute engine and mask management unit, efficiently handling our approach. It achieves up to 16.06x and 5.39x speedup and 4.18x and 4.93x energy-efficiency gain over A100 and Orin NX, respectively, while preserving quality. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2605_23226 |
| institution | arXiv |
| publishDate | 2026 |
| record_format | arxiv |
| spellingShingle | MASQ: Accelerating Masked Diffusion via Stage-Wise Multi-Precision Quantization Kim, Seeyeon Lee, Jaehun Yoo, Sungyeob Kim, Joo-Young Hardware Architecture Masked diffusion enables region-specific image synthesis but suffers from computational redundancy, since the entire image is processed each timestep even though only the masked region requires generation. To address this, we introduce MASQ, a hardware-software co-designed accelerator for masked diffusion. Our approach performs stage-wise MXINT8/4/2 precision assignment that dynamically reflects spatial and semantic importance, complemented by timestep-aware scheduling and optimized non-matrix operations. MASQ features a block-wise multi-precision compute engine and mask management unit, efficiently handling our approach. It achieves up to 16.06x and 5.39x speedup and 4.18x and 4.93x energy-efficiency gain over A100 and Orin NX, respectively, while preserving quality. |
| title | MASQ: Accelerating Masked Diffusion via Stage-Wise Multi-Precision Quantization |
| topic | Hardware Architecture |
| url | https://arxiv.org/abs/2605.23226 |