Quantization of Diffusion Models

News!

[April, 2025] We have released a new survey paper： Diffusion Model Quantization: A Review (https://arxiv.org/abs/2505.05215) based on this repository and summarized the current mainstream quantization methods for diffusion models! We are looking forward to any comments or discussions on this topic :)

Selected papers, corresponding codes and pre-trained models in our review paper.
If I missed your paper in this review, please email me or just pull a request here. I am more than happy to add it. Thanks!

A Taxonomy of Contemporary Approaches

、

Diffusion Backbone Types	Quantization Methods	Paper&Code
Unet-based Diffusion Quantization	Calibration Strategy Customization	✅[Post-training Quantization on Diffusion Models] (Paper || Code: Torch-based) (CVPR2023, First Diffusion Quantization Paper)
		✅[EDA-DM: Enhanced Distribution Alignment for Post-Training Quantization of Diffusion Models] (Paper || Code: Torch-based) (CVPR2023, First Diffusion Quantization Paper)
	Bi-modal Distribution Elimination	✅[Q-Diffusion: Quantizing Diffusion Models] (Paper || Code: Torch-based) (ICCV 2023)
	Dynamic Quantization	✅[Temporal Dynamic Quantization for Diffusion Models] (Paper || Code: No public) (NeurIPS 2023)
	Time Information Align	✅[TFMQ-DM: Temporal Feature Maintenance Quantization for Diffusion Models] (Paper || Code: Torch-based) (CVPR 2024, Highlight)
		✅[QVD: Post-training Quantization for Video Diffusion Models] (Paper || Code: No public) (ACM MM 2024)
	Quantization Error Correction	✅[QNCD: Quantization Noise Correction for Diffusion Models] (Paper || Code: Torch-based) (ACM MM 2024)
		✅[PTQD: Accurate Post-Training Quantization for Diffusion Models] (Paper || Code: Torch-based) (NeurIPS 2024)
		✅[Softmax Bias Correction for Quantized Generative Models] (Paper || Code: No public) (ICCV 2023, Workshop)
		✅[D$^2$-DPM: Dual Denoising for Quantized Diffusion Probabilistic Models] (Paper || Code: Torch-based) (AAAI 2025)
		✅[Timestep-Aware Correction for Quantized Diffusion Models] (Paper || Code: No public) (ECCV 2024)
	Text-Image Consistency Preservation	✅[DGQ: Distribution-Aware Group Quantization for Text-to-Image Diffusion Models] (Paper || Code: Torch-based) (ICLR 2025)
	Holistic QAT Optimization	✅[Q-DM: An Efficient Low-bit Quantized Diffusion Model] (Paper || Code: No public) (NeurIPS 2023)
		✅[QuEST: Low-bit Diffusion Model Quantization via Efficient Selective Finetuning] (Paper || Code: Torch-based)
		✅[Memory-Efficient Fine-Tuning for Quantized Diffusion Model] (Paper || Code: No public) (ECCV2024)
	Ultra-Low-Bit DMs	✅[Binary Latent Diffusion] (Paper || Code: No public)
		✅[BiDM: Pushing the Limit of Quantization for Diffusion Models] (Paper || Code: Torch-based)
		✅[Binarydm: Towards accurate binarization of diffusion model] (Paper || Code: Torch-based)
		✅[BitsFusion: 1.99 bits Weight Quantization of Diffusion Model] (Paper || Code: No public) (NeurIPS 2024)
		✅[Binarized Diffusion Model for Image Super-Resolution] (Paper || Code: Torch-based)
	LoRA-Based Enhancements	✅[EfficientDM: Efficient Quantization-Aware Fine-Tuning of Low-Bit Diffusion Models] (Paper || Code: Torch-based) (ICLR 2024)
		✅[IntLoRA: Integral Low-rank Adaptation of Quantized Diffusion Models] (Paper || Code: Torch-based)
Diffusion-Transformer(DiT) Quantization	Group-wise Quantization	✅[An Analysis on Quantizing Diffusion Transformers] (Paper || Code: No public) (CVPR workshop)
		✅[Q-DIT: ACCURATE POST-TRAINING QUANTIZATION FOR DIFFUSION TRANSFORMERS] (Paper || Code: Torch-based) (CVPR 2025)
	Channel Equalization	✅[PTQ4DiT: Post-training Quantization for Diffusion Transformers] (Paper || Code: Torch-based) (NerIPS 2024)
		✅[DiTAS: Quantizing Diffusion Transformers via Enhanced Activation Smoothing] (Paper || Code: Torch-based) (WACV 2025)
		✅[ViDiT-Q: Efficient and Accurate Quantization of Diffusion Transformers for Image and Video Generation] (Paper || Code: Torch-based) (ICLR 2025)
		✅[HQ-DiT: Efficient Diffusion Transformer with FP4 Hybrid Quantization] (Paper || Code: No public)

BenchMarking Experiments

We comprehensively evaluate the open-source PTQ-based and QAT-based methods in three tasks: unconditional image generation, class-conditional image generation, and text-conditional guided image generation. The details of the benchmarks for each task are outlined as follows.

Unconditional Image Generation

• We coducted a performance evaluation of various approaches for unconditional image generation, using LDM-4 (η = 1.0, steps = 200) on the LSUN-Bedrooms 256×256 dataset and LDM-8 (η = 0.0, steps = 200) on the LSUN-Churches 256×256 dataset.

Class-conditional Image Generation

• We also conducted a performance evaluation of various approaches for class-conditional image generation, using LDM-4 (scale = 3.0, η = 0.0, steps = 20) on the ImageNet 256×256 dataset. The dagger (†) symbol indicates QAT-based methods.

Text-conditional Guided Image Generation

• At last, we conducted a performance comparison of text-guided image generation with Stable-Diffusion v1-4 on MSCOCO captions.