Research map / Image generation / Fast diffusion sampling
DeltaQuant: 4-bit Video Diffusion Models with Spatiotemporal Delta Smoothing
Xingyang Li, Samuel Tesfai, Zhekai Zhang, Haocheng Xi, Shuo Yang, Lvmin Zhang, Yufei Sun, Kelly Peng, Maneesh Agrawala, Ion Stoica, Kurt Keutzer, Jun-Yan Zhu and 3 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 |
|---|---|---|---|
| quantize delta tokens to 4 bits with FP8 core tokens (DeltaQuant) instead of SVDQuant static smoothing | Wan2.2 I2V W4A4 · Wan2.2 I2V-A14B | PSNR: 21.9 -> 23.2 | Improved |
| replace SVDQuant with DeltaQuant | LTX-Video T2V W4A4 · LTX-Video T2V-13B | LPIPS: 0.193 -> 0.159 | Improved |
| replace SVDQuant with DeltaQuant | Wan2.2 T2V W4A4 · Wan2.2 T2V-A14B | Vision Reward: 0.083 -> 0.083 | No clear effect |
| replace SVDQuant with DeltaQuant | Wan2.2 T2V W4A4 · Wan2.2 T2V-A14B | PSNR: 17.5 -> 17.7 | Improved |
| add DeltaQuant on top of Radial and Sage Attention plus LightX2V | Wan2.2 I2V-A14B · Wan2.2 I2V-A14B | Latency (s): 131.0 -> 43.1 | Improved |
| add DeltaQuant on top of Radial and Sage Attention plus LightX2V | Wan2.2 I2V-A14B · Wan2.2 I2V-A14B | Vision Reward: 0.145 -> 0.148 | No clear effect |
| increase cube size to 64 tokens or a single global cube (ablation) | DeltaQuant · Wan2.2 (cube size ablation) | PSNR | Hurt |
| use DeltaQuant instead of raising the SVDQuant low-rank dimension (ablation) | Wan2.2 low-rank branch · Wan2.2 (rank ablation) | LPIPS vs latency | Improved |
| increase SVDQuant rank too far (ablation) | SVDQuant · Wan2.2 (rank ablation) | accuracy | Hurt |
| keep core tokens in FP8 instead of BF16 (ablation) | DeltaQuant · Wan2.2 (core precision ablation) | PSNR | No clear effect |
| quantize core tokens to NVFP4 instead of FP8 (ablation) | DeltaQuant · Wan2.2 (core precision ablation) | PSNR: drop of 1.3 | Hurt |
| use timestep-aware cube size (16 for 30% of steps, 64 for rest) instead of uniform 32 (ablation) | DeltaQuant · Wan2.2 (cube allocation ablation) | PSNR: +0.5 | Improved |
Closest papers by meaning
- SVDQuant: Absorbing Outliers by Low-Rank Component for 4-Bit Diffusion Models (ICLR 2025 · similarity 0.89)
- $\text{S}^2$Q-VDiT: Accurate Quantized Video Diffusion Transformer with Salient Data and Sparse Token Distillation (NeurIPS 2025 · similarity 0.89)
- Q-VDiT: Towards Accurate Quantization and Distillation of Video-Generation Diffusion Transformers (ICML 2025 · similarity 0.88)
- ViDiT-Q: Efficient and Accurate Quantization of Diffusion Transformers for Image and Video Generation (ICLR 2025 · similarity 0.88)
- EfficientDM: Efficient Quantization-Aware Fine-Tuning of Low-Bit Diffusion Models (ICLR 2024 · similarity 0.88)
- VideoMLA: Low-Rank Latent KV Cache for Minute-Scale Autoregressive Video Diffusion (NeurIPS 2026 · 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.
