![[2602.01289] Gradient-Aligned Calibration for Post-Training Quantization of Diffusion Models](https://arxiv.org/static/browse/0.3.4/images/arxiv-logo-fb.png){kind=logo}
[2602.01289] Gradient-Aligned Calibration for Post-Training Quantization of Diffusion Models
Summary
The paper presents a novel method for post-training quantization (PTQ) of diffusion models, addressing inefficiencies in existing calibration techniques by aligning gradients across timesteps for improved performance.
Why It Matters
This research is significant as it enhances the practical deployment of diffusion models, which are crucial for applications in image synthesis. By improving inference speed and reducing memory usage, this method can facilitate broader adoption of these models in real-world scenarios.
Key Takeaways
- Proposes a new PTQ method that assigns optimal weights to calibration samples.
- Addresses the limitations of uniform quantization in diffusion models.
- Demonstrates superior performance on benchmarks like CIFAR-10 and ImageNet.
Computer Science > Machine Learning arXiv:2602.01289 (cs) [Submitted on 1 Feb 2026 (v1), last revised 23 Feb 2026 (this version, v2)] Title:Gradient-Aligned Calibration for Post-Training Quantization of Diffusion Models Authors:Dung Anh Hoang, Cuong Pham anh Trung Le, Jianfei Cai, Thanh-Toan Do View a PDF of the paper titled Gradient-Aligned Calibration for Post-Training Quantization of Diffusion Models, by Dung Anh Hoang and Cuong Pham anh Trung Le and Jianfei Cai and Thanh-Toan Do View PDF HTML (experimental) Abstract:Diffusion models have shown remarkable performance in image synthesis by progressively estimating a smooth transition from a Gaussian distribution of noise to a real image. Unfortunately, their practical deployment is limited by slow inference speed, high memory usage, and the computational demands of the noise estimation process. Post-training quantization (PTQ) emerges as a promising solution to accelerate sampling and reduce memory overhead for diffusion models. Existing PTQ methods for diffusion models typically apply uniform weights to calibration samples across timesteps, which is sub-optimal since data at different timesteps may contribute differently to the diffusion process. Additionally, due to varying activation distributions and gradients across timesteps, a uniform quantization approach is sub-optimal. Each timestep requires a different gradient direction for optimal quantization, and treating them equally can lead to conflicting gradients that...
