Research map / Architectures and efficiency / Quantization
VQ-Transplant: Efficient VQ-Module Integration for Pre-trained Visual Tokenizers
Xianghong Fang, Yuan Yuan, Dehan Kong, Tim G. J. Rudner
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 decoder adaptation (5 epochs) after VQ substitution (ablation) | MMD VAR (K=4096) · ImageNet-1K | r-FID: 1.52 -> 0.91 | Improved |
| add decoder adaptation (5 epochs) (ablation) | MMD VAR (K=8192) · ImageNet-1K | r-FID: 1.49 -> 0.81 | Improved |
| add decoder adaptation (ablation) | Vanilla VQ fixed-scale (K=16384) · ImageNet-1K | r-FID: 10.61 -> 5.02 | Improved |
| add decoder adaptation (ablation) | EMA VQ fixed-scale (K=16384) · ImageNet-1K | r-FID: 1.86 -> 1.11 | Improved |
| extend decoder adaptation from 5 to 20 epochs (ablation) | MMD VAR · ImageNet-1K | r-FID: 0.91 -> 0.79 (K=4096); 0.81 -> 0.74 (K=8192) | Improved |
| replace native VQ with distribution-aligned VQ (MMD) vs Vanilla VQ (ablation) | VQ-Transplant · ImageNet-1K substitution (K=4096) | codebook utilization: 38.2% -> 100% | Improved |
| use MMD VQ instead of Wasserstein VQ (ablation) | VQ-Transplant · ImageNet-1K adaptation (K=4096) | r-FID: 0.93 -> 0.91 | No clear effect |
| lower quantization error alone (substitution without decoder adaptation) (ablation) | MMD VAR · ImageNet-1K | r-FID: 0.92 (original) -> 1.52 (substituted) | Hurt |
| use VQ-Transplant instead of from-scratch training (ablation) | MMD VAR · ImageNet-1K | r-FID: 1.40 -> 0.91 (K=4096) | Improved |
| apply VQ-Transplant to LDM-16 continuous tokenizer instead of VAR (ablation) | VQ-Transplant · ImageNet-1K | r-FID, r-IS | Hurt |
Closest papers by meaning
- Rethinking the Objectives of Vector-Quantized Tokenizers for Image Synthesis (CVPR 2024 · similarity 0.90)
- Channel-wise Vector Quantization (NeurIPS 2026 · similarity 0.90)
- Finite Scalar Quantization: VQ-VAE Made Simple (ICLR 2024 · similarity 0.89)
- Image Understanding Makes for A Good Tokenizer for Image Generation (NeurIPS 2024 · similarity 0.88)
- TokenFlow: Unified Image Tokenizer for Multimodal Understanding and Generation (CVPR 2025 · similarity 0.88)
- VQ4DiT: Efficient Post-Training Vector Quantization for Diffusion Transformers (AAAI 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.
