Timestep-Aware SVDQuant-GPTQ for W4A4 Quantization of Wan2.2-I2V

Researchers present a new quantization method for large video diffusion models that achieves 59.3% memory reduction while maintaining near-baseline quality. The technique addresses challenges in compressing Wan2.2-I2V's mixture-of-experts architecture by using timestep-aware and expert-specific calibration strategies.

This research tackles a fundamental challenge in deploying large video generation models: reducing memory consumption without sacrificing output quality. Video diffusion transformers represent one of the most computationally demanding AI applications, with their multi-step denoising processes consuming substantial GPU memory. The W4A4 quantization approach (4-bit weights and activations) addresses this by compressing model parameters to their lowest practical precision.

The technical innovation centers on recognizing that video generation models behave differently at different timesteps during the denoising process. Early timesteps handle high-noise predictions while later steps refine details, creating distinct activation patterns. Additionally, mixture-of-experts architectures route different data through specialized expert networks, each with unique quantization sensitivities. Previous approaches using single global calibration policies fail to capture these nuances, leading to accuracy degradation.

The framework combines three complementary techniques: SVDQuant handles sparse activation outliers through low-rank compensation, GPTQ performs reconstruction-aware weight quantization, and timestep-bin-wise per-layer searches optimize clipping ratios independently for each expert. Achieving 59.3% memory reduction with only 0.9% VBench score degradation demonstrates that specialized calibration strategies substantially outperform one-size-fits-all approaches.

For the AI infrastructure sector, this work enables practical deployment of high-quality video generation on consumer-grade GPUs, potentially democratizing video synthesis capabilities. The methodology's emphasis on expert-aware and timestep-aware quantization establishes a template for compressing other complex multi-stage transformer architectures. Developers targeting edge deployment or cost-sensitive inference pipelines gain viable pathways to maintain quality while reducing resource requirements significantly.

• →W4A4 quantization reduces Wan2.2-I2V peak GPU memory by 59.3% with minimal quality loss
• →Timestep-dependent activation distributions require specialized calibration strategies during quantization
• →Mixture-of-experts architectures need per-expert quantization policies for optimal compression
• →SVDQuant-GPTQ framework outperforms single global calibration approaches on video diffusion models
• →Post-training quantization enables practical deployment of large video models on resource-constrained hardware