Research map / Image generation / Fast diffusion sampling
DGQ: Distribution-Aware Group Quantization for Text-to-Image 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 DGQ, a post-training quantization method for text-to-image diffusion models that uses outlier-preserving group quantization for activations and log quantization with prompt-dependent scales for cross-attention, enabling sub-8-bit quantization without weight fine-tuning.
Problem
Quantizing text-to-image diffusion models to low bit-widths (<8 bits) degrades both image quality and text-image alignment, and prior methods rely on mixed precision or ignore cross-attention distributions.
Approach
Group activations into pixel-wise or channel-wise outlier groups (chosen by a range-variability metric and clustered with K-means) with per-group scales, and apply dynamic log quantization to cross-attention scores while keeping the <start> token's attention separate.
Key results
- On MS-COCO with W8A8 on SD v1.4, DGQ (16 groups) achieves FID 13.15 vs. 14.44 full precision and CLIP 0.297 vs. 0.298.
- At W4A6 on SD v1.4, DGQ (16 groups) reduces FID to 43.66 from over 200 for baselines (TFMQ-DM 229.64, Q-Diffusion 242.75) and raises CLIP to 0.263.
- Reports 93.7% bit-operation reduction (694 to 43.4 TBOPs) relative to full precision in the introduction.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DGQ instead of TFMQ at W8A8 | Stable Diffusion v1.4 · MS-COCO | FID: 18.85 -> 13.15 | Helped |
| Use DGQ instead of TFMQ at W4A6 | Stable Diffusion v1.4 · MS-COCO | FID: 229.64 -> 43.66 | Helped |
| Add outlier-preserving group quantization (ablation) | DGQ · MS-COCO, 8/8 ablation | FID: 18.83 -> 17.00 | Helped |
| Add outlier-preserving group quantization (ablation) | DGQ · MS-COCO, 8/6 ablation | FID: 176.44 -> 50.04 | Helped |
| Add attention-aware quantization (ablation) | DGQ · MS-COCO, 8/8 ablation | FID: 18.83 -> 18.24 | Helped |
| Add attention-aware quantization (ablation) | DGQ · MS-COCO, 8/6 ablation | FID: 176.44 -> 207.46 | Hurt |
| Add both outlier-preserving and attention-aware quantization (ablation) | DGQ · MS-COCO, 8/8 ablation | FID: 18.83 -> 14.40 | Helped |
| Add both outlier-preserving and attention-aware quantization (ablation) | DGQ · MS-COCO, 8/6 ablation | FID: 176.44 -> 40.15 | Helped |
| Use 2 groups with dimension selection instead of 1 group without selection (ablation) | DGQ · MS-COCO, 8/8 grouping ablation | FID: 18.83 -> 16.24 | Helped |
| Increase group count from 1 to 16 with dimension selection (ablation) | DGQ · MS-COCO, 8/6 grouping ablation | FID: 176.44 -> 49.05 | Helped |
| Use logarithmic instead of linear attention quantization, with dynamic scaling and start-token separation (ablation) | DGQ · MS-COCO, 8/6 attention ablation | FID: 49.98 -> 40.15 | Helped |
| Remove dynamic quantization from the logarithmic attention quantizer (ablation) | DGQ · MS-COCO, 8/8 attention ablation | FID: 18.24 -> 28.79 | Hurt |
Surprising findings
- Quantized W8A8 DGQ achieves a lower FID (13.15) than the full-precision SD v1.4 model (14.44).
- Running-minmax calibration of the log quantizer without dynamic scaling performs worse than the linear quantizer on cross-attention.
Limitations
- Dynamic per-prompt scale for cross-attention is a hardware-constrained approximation (max of remaining scores), not a more effective lookup-table or reparameterization approach.
- Weight quantization relies on existing BRECQ/AdaRound rather than the proposed method; combination with advanced weight quantization is left to future work.
- Performance at 6-bit activations still shows notable degradation vs. full precision (FID 43.66 vs. 14.44 at W4A6).
Experimental setup
Datasets: MS-COCO PartiPrompts
Benchmarks: MS-COCO PartiPrompts
Metrics: FID IS CLIP score MANIQA ImageReward BOPs
Baselines: Q-Diffusion TFMQ-DM
Builds on: Q-Diffusion BRECQ AdaRound
Closest papers by meaning
- Outlier-Aware Post-Training Quantization for Discrete Graph Diffusion Models (ICML 2025 · similarity 0.90)
- Data-Free Group-Wise Fully Quantized Winograd Convolution via Learnable Scales (CVPR 2025 · similarity 0.90)
- PermuQuant: Lowering Per-Group Quantization Error by Reordering Channels for Diffusion Models (NeurIPS 2026 · similarity 0.89)
- EfficientDM: Efficient Quantization-Aware Fine-Tuning of Low-Bit Diffusion Models (ICLR 2024 · similarity 0.88)
- SVDQuant: Absorbing Outliers by Low-Rank Component for 4-Bit Diffusion Models (ICLR 2025 · similarity 0.88)
- Würstchen: An Efficient Architecture for Large-Scale Text-to-Image Diffusion Models (ICLR 2024 · similarity 0.88)
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.