use this to replace https://github.com/real-stanford/diffusion_policy/blob/main/diffusion_policy/policy/diffusion_unet_hybrid_image_policy.py we find beta(0.5, 0.5) sampling and loss reweighting(more noisy, more weight on gt, verse visa) can beat trigflow/flow matching/diffusion in only one step.