Research map / Image generation / Fast diffusion sampling
QVGen: Pushing the Limit of Quantized Video Generative Models
Yushi Huang, Ruihao Gong, Jing Liu, Yifu Ding, Chengtao Lv, Haotong Qin, Jun Zhang
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 auxiliary module Phi to naive KD-based QAT (ablation) | Naive QAT, W4A4 Wan 1.3B · VBench | Imaging Quality: 60.40 -> 63.41 | Improved |
| add rank-decay to remove Phi (ablation) | Naive + Phi, W4A4 Wan 1.3B · VBench | Imaging Quality: 63.41 -> 63.08 | No clear effect |
| add rank-decay to remove Phi (ablation) | Naive + Phi, W4A4 Wan 1.3B · VBench | Overall Consistency: 22.98 -> 23.01 | Improved |
| use shrinking ratio 1/2 instead of 1 (decaying all Phi at once) (ablation) | QVGen W4A4 Wan 1.3B · VBench | Aesthetic Quality: 52.48 -> 54.67 | Improved |
| use initial rank r=32 instead of 8 (ablation) | QVGen W4A4 Wan 1.3B · VBench | Dynamic Degree: 74.62 -> 77.78 | Improved |
| increase initial rank from 32 to 64 (ablation) | QVGen W4A4 Wan 1.3B · VBench | Dynamic Degree: 77.78 -> 76.74 | Hurt |
| use rank-decay instead of linear decay (ablation) | Phi decay, W4A4 Wan 1.3B · VBench | Imaging Quality: 60.82 -> 63.08 | Improved |
| use rank-decay instead of Sparse with MaskLLM (ablation) | Phi decay, W4A4 Wan 1.3B · VBench | Imaging Quality: 61.36 -> 63.08 | Improved |
| use QVGen instead of EfficientDM on 3-bit CogVideoX-2B | W3A3 CogVideoX-2B · VBench | Dynamic Degree: 28.61 -> 53.89 | Improved |
| use QVGen instead of Q-DM on 4-bit Wan 1.3B | W4A4 Wan 1.3B · VBench | Aesthetic Quality: 52.50 -> 54.67 | Improved |
| mix half high-motion and half low-motion training videos instead of high-motion only (ablation) | W4A4 QVGen CogVideoX-2B · VBench | Dynamic Degree: 68.21 -> 67.08 | Mixed |
| use QVGen on larger Wan 14B at 3 bits instead of 4 bits | QVGen Wan 14B · VBench | Aesthetic Quality: 59.41 -> 29.73 | Hurt |
Closest papers by meaning
- Q-VDiT: Towards Accurate Quantization and Distillation of Video-Generation Diffusion Transformers (ICML 2025 · similarity 0.93)
- DSAQuant: Denoising-Stage-Aligned Quantization-Aware Training for Video Generation (NeurIPS 2026 · similarity 0.92)
- $\text{S}^2$Q-VDiT: Accurate Quantized Video Diffusion Transformer with Salient Data and Sparse Token Distillation (NeurIPS 2025 · similarity 0.91)
- ViDiT-Q: Efficient and Accurate Quantization of Diffusion Transformers for Image and Video Generation (ICLR 2025 · similarity 0.91)
- Q-ARVD: Quantizing Autoregressive Video Diffusion Models (NeurIPS 2026 · similarity 0.90)
- VQ4DiT: Efficient Post-Training Vector Quantization for Diffusion Transformers (AAAI 2025 · similarity 0.89)
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.
