Research map / Image generation / Fast diffusion sampling
TCAQ-DM: Timestep-Channel Adaptive Quantization for Diffusion Models
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 TCAQ-DM, a post-training quantization method for diffusion models that combines timestep-channel joint reparameterization, a dynamically adaptive log2/uniform quantizer for post-Softmax activations, and progressively aligned reconstruction.
Problem
Post-training quantization of diffusion models suffers large errors because activation ranges fluctuate across channels and timesteps, post-Softmax distributions shift over timesteps, and reconstruction inputs mismatch the iterative inference process.
Approach
Jointly rescales activations across timesteps and channels by folding scaling factors into weights, selects per-timestep between log2 and uniform quantizers using a power-law likelihood-ratio test, and iteratively re-samples calibration data from the quantized model to align reconstruction inputs with inference.
Key results
- On CIFAR-10 DDIM W6A6, FID 4.40 versus 9.59 for TFMQ-DM (full precision 4.14).
- On CIFAR-10 W4A4, FID 6.38 where PTQ4DM, Q-Diffusion and TFMQ-DM collapse to 236-384 FID.
- Ablation on CIFAR-10 W6A6: TCR reduces FID from 26.60 to 4.59 over the baseline.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add TCR to the BRECQ baseline (ablation) | TCAQ-DM · CIFAR-10, DDIM 100 steps, W8A8 | FID: 4.78 -> 4.12 | Helped |
| Add TCR to the BRECQ baseline (ablation) | TCAQ-DM · CIFAR-10, DDIM 100 steps, W6A6 | FID: 26.60 -> 4.59 | Helped |
| Add TCR to the BRECQ baseline (ablation) | TCAQ-DM · CIFAR-10, DDIM 100 steps, W4A4 | FID: 371.61 -> 9.54 | Helped |
| Add DAQ after TCR (ablation) | TCAQ-DM · CIFAR-10, DDIM 100 steps, W4A8 | FID: 4.76 -> 4.59 | Helped |
| Add DAQ after TCR (ablation) | TCAQ-DM · CIFAR-10, DDIM 100 steps, W4A4 | FID: 9.54 -> 9.09 | Helped |
| Add PAR after TCR and DAQ (ablation) | TCAQ-DM · CIFAR-10, DDIM 100 steps, W4A4 | FID: 9.09 -> 6.38 | Helped |
| Use DAQ instead of a log2 quantizer (ablation) | CIFAR-10 DDIM post-Softmax quantization · CIFAR-10, DDIM 100 steps, W6A6 S8 | FID: 4.97 -> 4.42 | Helped |
| Use DAQ instead of a uniform quantizer (ablation) | CIFAR-10 DDIM post-Softmax quantization · CIFAR-10, DDIM 100 steps, W6A6 S8 | FID: 4.66 -> 4.42 | Helped |
| Use DAQ instead of a log2 quantizer (ablation) | CIFAR-10 DDIM post-Softmax quantization · CIFAR-10, DDIM 100 steps, W6A6 S6 | FID: 4.77 -> 4.61 | Helped |
| Use DAQ instead of a uniform quantizer (ablation) | CIFAR-10 DDIM post-Softmax quantization · CIFAR-10, DDIM 100 steps, W6A6 S6 | FID: 4.87 -> 4.61 | Helped |
| Use DAQ instead of a uniform quantizer (ablation) | CIFAR-10 DDIM post-Softmax quantization · CIFAR-10, DDIM 100 steps, W6A6 S4 | FID: 14.20 -> 4.66 | Helped |
Surprising findings
- A pure log2 quantizer can perform worse than a uniform quantizer at high bit-widths for post-Softmax activations in diffusion models, motivating the adaptive choice.
Limitations
- Scaling vector can enlarge weight values, so truncation hyper-parameter R_tru is needed for low-bit weight settings like W4A8.
- Reconstruction needs multiple progressive rounds, adding calibration cost; W4A4 results still show a gap to full precision (e.g., LSUN-Churches 29.17 vs 4.08 FID; ImageNet 30.69 FID).
Experimental setup
Datasets: CIFAR-10 ImageNet LSUN-Bedrooms LSUN-Churches
Benchmarks: CIFAR-10 ImageNet LSUN-Bedrooms LSUN-Churches
Metrics: FID IS sFID
Baselines: PTQ4DM Q-Diffusion PTQD APQ-DM TFMQ-DM TAC-Diffusion
Builds on: BRECQ Q-Diffusion RepQ-ViT TFMQ-DM
Closest papers by meaning
- Q-DiT: Accurate Post-Training Quantization for Diffusion Transformers (CVPR 2025 · similarity 0.93)
- Optimizing Quantized Diffusion Models via Distillation with Cross-Timestep Error Correction (AAAI 2025 · similarity 0.92)
- EfficientDM: Efficient Quantization-Aware Fine-Tuning of Low-Bit Diffusion Models (ICLR 2024 · similarity 0.92)
- Towards Accurate Post-training Quantization for Diffusion Models (CVPR 2024 · similarity 0.92)
- TFMQ-DM: Temporal Feature Maintenance Quantization for Diffusion Models (CVPR 2024 · similarity 0.91)
- MPQ-DM: Mixed Precision Quantization for Extremely Low Bit Diffusion Models (AAAI 2025 · 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.