Research map / Video / Video generation

DSAQuant: Denoising-Stage-Aligned Quantization-Aware Training for Video Generation

NeurIPS 2026 · Video › Video generation · approach: Efficient video diffusion models · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes DSAQuant, a denoising-stage-aligned QAT framework for video diffusion models that uses an exponential KD-to-target loss schedule during training and disables CFG in the final denoising steps at inference, improving VBench over QVGen under W4A4 and W3A3.

Problem

Conventional quantization-aware training of video diffusion models preserves prompt semantics and layout but severely degrades fine detail, texture and sharpness, because it treats all denoising timesteps uniformly.

Approach

Make QAT and inference stage-aware: transfer supervision from teacher distillation in early steps to the native target loss in late steps via a normalized exponential schedule, and drop classifier-free guidance in the final steps to stop it amplifying cond/uncond quantization mismatch.

Key results

• Under W3A3 with augmented prompts, DSAQuant improves VBench average over QVGen by 2.63 to 6.60 points across Wan backbones.
• On Wan2.1-1.3B W4A4 symmetric, the VBench average rises to 67.21 versus 63.56 for QVGen (full precision 66.28).
• Inference memory for Wan2.1-14B W4A4 drops 45.1% from FP16 and storage drops about 74.8%.

Surprising findings

• CFG-drop also gives a small VBench gain for the full-precision model, though the gains are much larger for quantized models.
• Overly aggressive KD decay (alpha=10) hurts performance, and extending CFG-drop beyond ten steps lowers VBench.

Limitations

• QVGen training data is not released, so the baseline was reproduced by the authors under matched settings.
• Evaluation relies mainly on VBench and automatic metrics; no human evaluation is reported.
• CFG-drop threshold (nine steps) and the schedule alpha are tuned on Wan2.1-1.3B and may not transfer directly.

Experimental setup

Datasets: VidProM TurboDiffusion latent data

Benchmarks: VBench VBench-2.0 VBench augmented prompts Vision Reward Laplacian Variance

Metrics: VBench average imaging quality aesthetic quality motion smoothness dynamic degree consistency scores VBench-2.0 average Laplacian Variance Vision Reward instruction following

Baselines: QVGen ViDiT-Q SVDQuant S2Q-VDiT LSQ Q-DM EfficientDM

Builds on: QVGen classifier-free guidance straight-through estimator flow matching

Closest papers by meaning

• Q-VDiT: Towards Accurate Quantization and Distillation of Video-Generation Diffusion Transformers (ICML 2025 · 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)
• EfficientDM: Efficient Quantization-Aware Fine-Tuning of Low-Bit Diffusion Models (ICLR 2024 · similarity 0.89)
• VQ4DiT: Efficient Post-Training Vector Quantization for Diffusion Transformers (AAAI 2025 · similarity 0.89)
• Q-DiT: Accurate Post-Training Quantization for Diffusion Transformers (CVPR 2025 · similarity 0.89)