Research map / Image generation / Autoregressive and latent generation
ArcVQ-VAE: A Spherical Vector Quantization Framework with ArcCosine Additive Margin
Jaeyung Kim, YoungJoon Yoo
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 |
|---|---|---|---|
| add Ball-Bounded Norm Regularization to VQ-VAE (ablation) | VQ-VAE · MNIST | rFID: 3.43 -> 2.71 | Improved |
| add Ball-Bounded Norm Regularization to VQ-VAE (ablation) | VQ-VAE · CIFAR10 | rFID: 39.67 -> 35.18 | Improved |
| add ArcCosine Additive Margin Loss on top of BBNR (ablation) | VQ-VAE + BBNR · MNIST | rFID: 2.71 -> 1.68 | Improved |
| add ArcCosine Additive Margin Loss on top of BBNR (ablation) | VQ-VAE + BBNR · CIFAR10 | rFID: 35.18 -> 26.91 | Improved |
| vary top-k of nearest latent tokens (1, 3, 8) (ablation) | ArcVQ-VAE · MNIST and CIFAR10 | PSNR / rFID: CIFAR10 rFID 27.08, 26.91, 27.02 | No clear effect |
| increase additive angular margin m from 0.1 to 1.0 (ablation) | ArcVQ-VAE · CIFAR10 | rFID: 26.91 -> 28.52 | Hurt |
| increase scaling factor s from 10 to 20 (ablation) | ArcVQ-VAE · CIFAR10 | PSNR / rFID: PSNR 24.78 -> 23.95; rFID 26.91 -> 28.90 | Hurt |
| increase codebook dimension from 4 to 16 (ablation) | ArcVQ-VAE · FFHQ | codebook usage vs rFID: usage 98% -> 78%; rFID 2.71 -> 2.14 | Mixed |
| apply SAMP (BBNR + ArcLoss) to vanilla VQGAN tokenizer | VQGAN · ImageNet-1K, 16x16 tokens, K=1024 | rFID: 7.94 -> 1.63 | Improved |
| use ArcVQ-VAE instead of CVQ-VAE at 16x16 tokens | ImageNet-1K tokenizer · ImageNet-1K, 16x16 tokens | rFID: 1.57 -> 1.63 | Hurt |
| use ArcVQ-VAE instead of CVQ-VAE on CIFAR10 | small-scale tokenizer · CIFAR10 | rFID: 24.73 -> 26.91 | Hurt |
| use ArcVQ-VAE tokenizer instead of CVQ-VAE for LDM generation | LDM on ImageNet-1K · ImageNet-1K generation | FID: 6.87 -> 6.79 | Improved |
Closest papers by meaning
- Restructuring Vector Quantization with the Rotation Trick (ICLR 2025 · similarity 0.90)
- Finite Scalar Quantization: VQ-VAE Made Simple (ICLR 2024 · similarity 0.89)
- Unveiling And Addressing Dimensional Collapse In Vector Quantization Models Via Codebook Regularization (ICML 2026 · similarity 0.89)
- Scalable Training for Vector-Quantized Networks with 100% Codebook Utilization (ICLR 2026 · similarity 0.88)
- Online Clustered Codebook (ICCV 2023 · similarity 0.88)
- Channel-wise Vector Quantization (NeurIPS 2026 · similarity 0.86)
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.
