Research map / Architectures and efficiency / Quantization
Memory-Efficient Generative Models via Product Quantization
Jie Shao, Hanxiao Zhang, Hao Yu, Jianxin Wu
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 |
|---|---|---|---|
| use product quantization instead of vector quantization at 2-bit | DiT-XL/2 · ImageNet 256x256, 250 steps | FID: 12.34 -> 6.84 | Improved |
| use product quantization instead of vector quantization at 1-bit | DiT-XL/2 · ImageNet 256x256, 250 steps | FID: 91.71 -> 14.03 | Improved |
| use product quantization instead of vector quantization at 4-bit | DiT-XL/2 · ImageNet 256x256, 250 steps | FID: 6.26 -> 6.25 | No clear effect |
| extend VQ calibration from 1 to 10 epochs (ablation) | VQ 2-bit DiT-XL/2 · ImageNet 256x256, 250 steps | FID: 12.33 -> 12.12 | No clear effect |
| apply product quantization with full method vs GPTQ at 2-bit | DiT-XL/2 · ImageNet 256x256, 250 steps | FID: 280.11 -> 6.84 | Improved |
| use our compression instead of Q-Diffusion at 2-bit | LDM-4 · ImageNet 256x256, 20 steps | FID: 43.96 -> 7.98 | Improved |
| use our compression instead of SVDQuant at 2-bit | PixArt-alpha · GenEval overall | GenEval Overall: 0.17 -> 0.46 | Improved |
| use our compression instead of SVDQuant at 2-bit | FLUX.1-dev · GenEval overall | GenEval Overall: 0.32 -> 0.57 | Improved |
| apply our PQ method on top of Compact3D | Compact3D · Tanks and Temples | PSNR: 23.39 -> 23.57 | Improved |
| apply our PQ method on top of Quip# | Quip# LLaMA-2 7B · Wino/PiQA/Hella | accuracy: Wino 62.43 -> 64.87 | Mixed |
| full method (PQ with DDPM loss, assignment adjustment) in ablation (ablation) | Ours 2-bit DiT-XL/2 · ImageNet 256x256 | FID: 7.21 -> 6.84 | Improved |
Closest papers by meaning
- Bridging the Gap between Gaussian Diffusion Models and Universal Quantization for Image Compression (CVPR 2025 · similarity 0.89)
- SqueezeLLM: Dense-and-Sparse Quantization (ICML 2024 · similarity 0.89)
- When Worse is Better: Navigating the Compression Generation Trade-off In Visual Tokenization (NeurIPS 2025 · similarity 0.88)
- HyperVQ: Enabling Hyperprior Entropy Modeling for VQ-Based Generative Image Compression (NeurIPS 2026 · similarity 0.88)
- Compressed Image Generation with Denoising Diffusion Codebook Models (ICML 2025 · similarity 0.88)
- Progressive Compression with Universally Quantized Diffusion Models (ICLR 2025 · 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.
