Problem Description
When calibrating Wan (video diffusion / DiT) quantization with COCO2014 images, the quantized model still “works”, but video-specific behavior degrades:
A prompt like “a cat wearing boxing gloves throwing punches” generates noticeably slower motion after quantization.
Punching / jumping / fast limb movement becomes mushy or delayed.
Single-frame quality may look OK, but temporal dynamics are suppressed.
Compared to FP16/BF16, the quantized clip feels like “slow-motion + lower video quality”.
This suggests COCO2014 is not representative of the activation / motion distribution Wan actually sees during video decoding.
Reproduction Steps
Better calibration source: OpenS2V
Recent Wan quantization work uses OpenS2V / OpenS2V-5M / OpenS2V-Eval as calibration or evaluation data because it is video/subjective-motion/subject-consistency aware.
Relevant papers:
OpenS2V-Nexus: A Detailed Benchmark and Million-Scale Dataset for Subject-to-Video Generation
Shenghai Yuan et al., 2025. OpenS2V-Eval + OpenS2V-5M for subject consistency / natural motion / text relevance.
https://arxiv.org/abs/2505.20292
Collaborative Few-Step Distillation and Low-Bit Quantization for Wan2.2 Dual-Expert Video Diffusion Models
Jinyang Du et al., 2026. Calibrates high-noise / low-noise experts separately; uses OpenS2V-based generation stream for video-quality calibration.
https://arxiv.org/abs/2606.00658
Tail-Aware HiFloat4: W4A4 Post-Training Quantization for Wan2.2
Feng et al., 2026. Wan2.2 W4A4 PTQ; emphasizes activation-tail-aware calibration instead of naive image-domain calibration.
https://arxiv.org/abs/2605.26628
Timestep-Aware SVDQuant-GPTQ for W4A4 Quantization of Wan2.2-I2V
Wu et al., 2026. Shows single global calibration policy is insufficient for MoE video DiTs; timestep/expert-aware calibration matters. Evaluated on OpenS2V-Eval.
Environment Information
No response
Error Logs
Additional Context
No response
Problem Description
When calibrating Wan (video diffusion / DiT) quantization with COCO2014 images, the quantized model still “works”, but video-specific behavior degrades:
A prompt like “a cat wearing boxing gloves throwing punches” generates noticeably slower motion after quantization.
Punching / jumping / fast limb movement becomes mushy or delayed.
Single-frame quality may look OK, but temporal dynamics are suppressed.
Compared to FP16/BF16, the quantized clip feels like “slow-motion + lower video quality”.
This suggests COCO2014 is not representative of the activation / motion distribution Wan actually sees during video decoding.
Reproduction Steps
Better calibration source: OpenS2V
Recent Wan quantization work uses OpenS2V / OpenS2V-5M / OpenS2V-Eval as calibration or evaluation data because it is video/subjective-motion/subject-consistency aware.
Relevant papers:
OpenS2V-Nexus: A Detailed Benchmark and Million-Scale Dataset for Subject-to-Video Generation
Shenghai Yuan et al., 2025. OpenS2V-Eval + OpenS2V-5M for subject consistency / natural motion / text relevance.
https://arxiv.org/abs/2505.20292
Collaborative Few-Step Distillation and Low-Bit Quantization for Wan2.2 Dual-Expert Video Diffusion Models
Jinyang Du et al., 2026. Calibrates high-noise / low-noise experts separately; uses OpenS2V-based generation stream for video-quality calibration.
https://arxiv.org/abs/2606.00658
Tail-Aware HiFloat4: W4A4 Post-Training Quantization for Wan2.2
Feng et al., 2026. Wan2.2 W4A4 PTQ; emphasizes activation-tail-aware calibration instead of naive image-domain calibration.
https://arxiv.org/abs/2605.26628
Timestep-Aware SVDQuant-GPTQ for W4A4 Quantization of Wan2.2-I2V
Wu et al., 2026. Shows single global calibration policy is insufficient for MoE video DiTs; timestep/expert-aware calibration matters. Evaluated on OpenS2V-Eval.
Environment Information
No response
Error Logs
Additional Context
No response