Research map / Video / Video generation
Quant VideoGen: Auto-Regressive Long Video Generation via 2-Bit KV-Cache Quantization
Haocheng Xi, Shuo Yang, Yilong Zhao, Muyang Li, Han Cai, Xingyang Li, Yujun Lin, Zhuoyang Zhang, Jintao Zhang, Xiuyu Li, Zhiying Xu, Jun Wu and 4 more
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 QVG instead of QuaRot at INT2 | KV cache quantization · LongCat-Video-13B INT2 | PSNR: 21.573 -> 28.716 | Improved |
| use QVG instead of QuaRot at INT2 | KV cache quantization · HY-WorldPlay-8B INT2 | PSNR: 25.207 -> 29.174 | Improved |
| use QVG-Pro (4 stages, group 16) instead of QVG (ablation) | QVG · LongCat-Video-13B INT2 | PSNR: 28.716 -> 30.376 | Improved |
| use QVG-Pro instead of QVG (ablation) | QVG · LongCat-Video-13B INT2 | compression ratio: 6.94x -> 4.97x | Hurt |
| use QVG instead of QuaRot at INT4 | KV cache quantization · LongCat-Video-13B INT4 | PSNR: 33.744 -> 37.141 | Improved |
| use QVG instead of QuaRot at INT4 | KV cache quantization · HY-WorldPlay-8B INT4 | PSNR: 33.997 -> 34.454 | Improved |
| use QVG instead of KIVI at 1400 frames | KV cache quantization · Self-Forcing 1400 frames | Image Quality: 35.85 -> 67.28 | Improved |
| use QVG instead of QuaRot at 720p | KV cache quantization · LongCat-Video-13B 720p | PSNR: 20.88 -> 25.99 | Improved |
| add more progressive residual quantization stages (ablation) | QVG · stage sweep | MSE reduction: 5.83x first stage, at least 1.10x each later stage | Improved |
| increase codebook size K from 64 to 512 (ablation) | QVG INT2, 1 stage · codebook sweep | key reconstruction improvement: 4.832 -> 6.428 | Improved |
| increase codebook size K from 64 to 512 (ablation) | QVG INT2, 1 stage · codebook sweep | compression ratio: 7.760 -> 7.307 | Hurt |
| increase chunk size from 9360 to 37440 (ablation) | QVG Self-Forcing INT2 · chunk size sweep | compression ratio: 5.8x -> 7.0x | Improved |
Closest papers by meaning
- Q-VDiT: Towards Accurate Quantization and Distillation of Video-Generation Diffusion Transformers (ICML 2025 · similarity 0.90)
- VideoMAR: Autoregressive Video Generation with Continuous Tokens (NeurIPS 2025 · similarity 0.89)
- ViDiT-Q: Efficient and Accurate Quantization of Diffusion Transformers for Image and Video Generation (ICLR 2025 · similarity 0.89)
- Training-Free Efficient Video Generation via Dynamic Token Carving (NeurIPS 2025 · similarity 0.88)
- VideoMLA: Low-Rank Latent KV Cache for Minute-Scale Autoregressive Video Diffusion (NeurIPS 2026 · similarity 0.88)
- From Slow Bidirectional to Fast Autoregressive Video Diffusion Models (CVPR 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.
