Research map / Image generation / Fast diffusion sampling
Q&C: When Quantization Meets Cache in Efficient Generation
Xin Ding, Xin Li, Haotong Qin, Zhibo Chen
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| naively stack quantization (PTQ4DiT) and cache (Learn-to-Cache) (ablation) | PTQ4DiT · ImageNet 256 W8A8 50 steps | FID: 5.45 -> 13.67 | Hurt |
| add Temporal-Aware Parallel Clustering for calibration data (ablation) | Quantization+cache baseline · ImageNet 256 W8A8 50 steps | FID: 13.67 -> 8.34 | Improved |
| add Variance Compensation (ablation) | Quantization+cache baseline · ImageNet 256 W8A8 50 steps | FID: 13.67 -> 9.65 | Improved |
| combine TAP and VC (ablation) | Quantization+cache baseline · ImageNet 256 W8A8 50 steps | FID: 13.67 -> 5.43 | Improved |
| use TAP instead of K-Means for calibration selection (ablation) | Q&C · ImageNet 256 W8A8 50 steps | FID: 10.31 -> 8.34 | Improved |
| use TAP instead of Agglomerative clustering (ablation) | Q&C · ImageNet 256 W8A8 50 steps | FID: 9.56 -> 8.34 | Improved |
| vary spatial/temporal weight alpha in TAP from 0.3 to 0.7 (ablation) | TAP · ImageNet 256 W8A8 50 steps | FID: 5.57 / 5.46 / 5.43 / 5.36 / 5.45 | No clear effect |
| combine quantization and cache (Q&C) vs PTQ4DiT alone at 50 steps | PTQ4DiT · ImageNet 256 W8A8 50 steps | speedup: 10x -> 12.7x | Improved |
| apply Q&C to 512x512 W4A8 | PTQ4DiT · ImageNet 512 W4A8 100 steps | FID: 19.00 -> 19.05; speed 2.5x -> 4x | No clear effect |
| apply Q&C to latent diffusion on LSUN-Church | APQ-DM · LSUN-Church W8A8 | FID: 9.04 -> 9.10 | No clear effect |
| combined quantization and cache raises calibration sample similarity | Quantization+cache · DiT calibration samples | cosine similarity | Hurt |
Closest papers by meaning
- QuantCache: Adaptive Importance-Guided Quantization with Hierarchical Latent and Layer Caching for Video Generation (ICCV 2025 · similarity 0.86)
- Diffusion on Demand: Selective Caching and Modulation for Efficient Generation (NeurIPS 2025 · similarity 0.85)
- KIVI: A Tuning-Free Asymmetric 2bit Quantization for KV Cache (ICML 2024 · similarity 0.85)
- Predicting Quantization Price for Selecting PTQ Configurations Before Deployment (NeurIPS 2026 · similarity 0.84)
- ZipCache: Accurate and Efficient KV Cache Quantization with Salient Token Identification (NeurIPS 2024 · similarity 0.84)
- QJL: 1-Bit Quantized JL Transform for KV Cache Quantization with Zero Overhead (AAAI 2025 · similarity 0.83)
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.
