Answers are grounded in this paper and cite their sources.
Technical report demonstrates a post-training quantization solution for an image-to-video model, suggesting efficient evaluation methods.
This technical report describes the ReopenAI team's post-training quantization (PTQ) solution for Sub-Challenge 1 of the ICME 2026 Low Bit-width Large Model Quantization Challenge. The target model is Wan2.2-I2V-A14B, an image-to-video generative model evaluated on OpenS2V-5M prompts with VBench-based quality metrics. The proposed method applies HiFloat4 W4A4 quantization to eligible linear layers in both low-noise and high-noise DiT experts, while keeping the first and last Transformer blocks in BF16. The pipeline uses calibration-only activation statistics, SmoothQuant-style equivalent scaling, activation and weight clipping, and reproducible quantization metadata export. The method requires no fine-tuning and preserves the official Wan2.2 generation path for reproducible evaluation with the official HiFloat4 simulator.
No takes yet. Share an insight, caveat, or question.
Jianfeng Wu (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: