This is a partial NVFP4 quantization of <a target="_blank" rel="ugc" href="https://huggingface.co/lightx2v/Wan2.1-I2V-14B-480P-StepDistill-CfgDistill-Lightx2v">Wan2.1-I2V-14B-480P-StepDistill-CfgDistill-Lightx2v</a> by lightx2v, produced using <a target="_blank" rel="ugc" href="https://github.com/silveroxides/convert_to_quant">convert_to_quant</a> by <a target="_blank" rel="ugc" href="https://huggingface.co/silveroxides">silveroxides</a>.<a target="_blank" rel="ugc" href="https://huggingface.co/lightx2v/Wan2.1-I2V-14B-480P-StepDistill-CfgDistill-Lightx2v">Wan2.1-I2V-14B-480P-StepDistill-CfgDistill-Lightx2v</a> is an image-to-video generation model built on <a target="_blank" rel="ugc" href="https://huggingface.co/Wan-AI/Wan2.1-I2V-14B-480P">Wan2.1-I2V-14B-480P</a>. It applies step distillation and classifier-free guidance distillation to reduce inference to 4 steps without CFG, cutting generation time substantially while preserving output quality. IMPORTANTSince NVFP4 is only supported on NVIDIA Blackwell architecture GPUs, running this model requires a Blackwell GPU with its corresponding support enabled in torch, along with a recent version of ComfyUI and <a target="_blank" rel="ugc" href="https://github.com/Comfy-Org/comfy-kitchen">comfy-kitchen</a> built against CUDA 13.The model can be used in ComfyUI with the following parameters, based on the distilled model's own recommendations:<ul><li>Shift: 5.0</li><li>Sampler: LCM</li><li>Scheduler normal</li><li>CFG: 1.0</li><li>Steps: 4</li></ul>Other combinations are possible, such as euler/simple and heun/linear_quadratic.

Wan2.1 I2V Lightx2v Step/CFG NVFP4

This version tries to protect the following layers, among others. <ul><li>cross_attn.k</li><li>cross_attn.q</li><li>self_attn.k</li><li>self_attn.q</li></ul>

Files with SHA256: 644857118a5e8f5f8bb13e56c860eca7bdc1a32dd12ffdc97ed9cd9460254945