Research map / Architectures and efficiency / Quantization
PTQ4DiT: Post-training Quantization for Diffusion Transformers
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes PTQ4DiT, a post-training quantization method for Diffusion Transformers that balances salient activation/weight channels and calibrates temporal salience via Spearman's rho, enabling W8A8 and W4A8 quantization of DiT-XL/2.
Problem
Post-training quantization of Diffusion Transformers (DiTs) is hard due to salient channels with extreme magnitudes in weights and activations and temporal variation of salient activation channels across denoising timesteps.
Approach
Channel-wise Salience Balancing redistributes complementary extreme magnitudes between activations and weights via diagonal balancing matrices; Spearman's rho-guided calibration weights timesteps by inverse correlation with weight salience; the balancing matrices are folded offline into adaLN MLPs, preceding matmul dequantization, and weights to add no inference cost.
Key results
- At W8A8 on ImageNet 256x256 with 250 steps, PTQ4DiT reaches FID 4.63 vs FP 4.53, versus 5.57 for Q-Diffusion.
- At W4A8 on ImageNet 256x256 with 250 steps, PTQ4DiT achieves FID 7.09 while Q-Diffusion gets 15.31 and PTQ4DM 72.58.
- At W4A8 on ImageNet 512x512 with 250 steps, PTQ4DiT reaches FID 17.55 versus 58.81 for Q-Diffusion.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use PTQ4DiT instead of RepQ* | DiT-XL/2 quantization · ImageNet 256×256, 250 steps, W8A8 | FID: 4.51 -> 4.63 | Hurt |
| Use PTQ4DiT instead of Q-Diffusion | DiT-XL/2 quantization · ImageNet 256×256, 250 steps, W4A8 | FID: 15.31 -> 7.09 | Helped |
| Use PTQ4DiT instead of RepQ* | DiT-XL/2 quantization · ImageNet 256×256, 100 steps, W8A8 | FID: 5.20 -> 4.73 | Helped |
| Use PTQ4DiT instead of RepQ* | DiT-XL/2 quantization · ImageNet 256×256, 100 steps, W4A8 | FID: 26.64 -> 7.75 | Helped |
| Use PTQ4DiT instead of RepQ* | DiT-XL/2 quantization · ImageNet 256×256, 50 steps, W8A8 | FID: 7.17 -> 5.45 | Helped |
| Use PTQ4DiT instead of Q-Diffusion | DiT-XL/2 quantization · ImageNet 512×512, 250 steps, W4A8 | FID: 58.81 -> 17.55 | Helped |
| Add CSB to basic linear quantization (ablation) | DiT-XL/2 · ImageNet 256×256, 250 steps, W4A8 | FID: 22.54 -> 8.17 | Helped |
| Add SSC to Baseline + CSB (ablation) | DiT-XL/2 · ImageNet 256×256, 250 steps, W4A8 | FID: 8.17 -> 7.09 | Helped |
Surprising findings
- At W8A8 with 250 steps, the RepQ* baseline achieves a slightly better FID (4.51) than PTQ4DiT (4.63), though PTQ4DiT has better sFID, IS and Precision.
Limitations
- Research concentrates on class-conditional visual generation; not extended to audio, 3D or other modalities.
- Evaluated only on DiT-XL/2 with ImageNet class-conditional generation.
Experimental setup
Datasets: ImageNet
Benchmarks: ImageNet 256x256 ImageNet 512x512
Metrics: FID sFID IS Precision Size (MB)
Baselines: PTQ4DM Q-Diffusion PTQD RepQ-ViT (RepQ*)
Builds on: Q-Diffusion RepQ-ViT SmoothQuant DDPM
Closest papers by meaning
- Q-DiT: Accurate Post-Training Quantization for Diffusion Transformers (CVPR 2025 · similarity 0.95)
- DiRotQ: Rotation-Aware Quantization for 4-bit Diffusion Transformers (NeurIPS 2026 · similarity 0.93)
- ViDiT-Q: Efficient and Accurate Quantization of Diffusion Transformers for Image and Video Generation (ICLR 2025 · similarity 0.92)
- VQ4DiT: Efficient Post-Training Vector Quantization for Diffusion Transformers (AAAI 2025 · similarity 0.92)
- KroQuant: Kronecker-Structured Block Transforms for Efficient Post-Training Quantization of Diffusion Transformers (NeurIPS 2026 · similarity 0.91)
- EfficientDM: Efficient Quantization-Aware Fine-Tuning of Low-Bit Diffusion Models (ICLR 2024 · similarity 0.91)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.