Research map / Image generation / Autoregressive and latent generation
Product-Quantised Image Representation for High-Quality Image Synthesis
Denis Zavadski, Nikita Philip Tatsch, Carsten Rother
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 |
|---|---|---|---|
| replace vector quantisation with product quantisation (S>1) as latent dim grows (ablation) | VQGAN autoencoder · ImageNet 256 | FID | Improved |
| increase number of subspaces S up to d/2 (ablation) | PQGAN · ImageNet 256 | FID | Improved |
| increase codebook size K (ablation) | PQGAN · ImageNet 256 | FID | No clear effect |
| PQGAN (F=8, K=512) vs best baseline VQGAN-LC (K=100000) | latent autoencoder · ImageNet 256 val | PSNR (dB): 27.0 -> 37.4 | Improved |
| PQGAN F=16 vs continuous SDv2.1 VAE F=8 | latent autoencoder · ImageNet 256 val | rFID: 0.75 -> 0.41 | Improved |
| train latent diffusion from scratch on PQGAN latents instead of SDv2.1 VAE | LDM 400M · ImageNet 256 | gFID: 27.21 -> 22.66 | Improved |
| adapt SD to PQ latent with F=16 and 48x48 latent resolution | SDv2.1 (PQSD-Quick) · 768x768 generation | samples/s: 0.116 -> 0.465 | Improved |
| PQGAN F=16 on 1536px generation vs SDv2.1 768px | SDv2.1 (PQSD-HR) · generation cost | samples/s: 0.116 -> 0.112 | No clear effect |
Closest papers by meaning
- Scaling the Codebook Size of VQ-GAN to 100,000 with a Utilization Rate of 99% (NeurIPS 2024 · similarity 0.88)
- Finite Scalar Quantization: VQ-VAE Made Simple (ICLR 2024 · similarity 0.88)
- Rethinking the Objectives of Vector-Quantized Tokenizers for Image Synthesis (CVPR 2024 · similarity 0.88)
- VQ4DiT: Efficient Post-Training Vector Quantization for Diffusion Transformers (AAAI 2025 · similarity 0.87)
- ViDiT-Q: Efficient and Accurate Quantization of Diffusion Transformers for Image and Video Generation (ICLR 2025 · similarity 0.87)
- Learning to Quantize for Training Vector-Quantized Networks (ICML 2025 · similarity 0.87)
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.
