Function logsnr_schedule_cosine_interpolated

k_diffusion/utils.py:367–370

(t, image_d, noise_d_low, noise_d_high, logsnr_min, logsnr_max)

        return logsnr_schedule_cosine(t, logsnr_min - shift, logsnr_max - shift) + shift

    def logsnr_schedule_cosine_interpolated(t, image_d, noise_d_low, noise_d_high, logsnr_min, logsnr_max):
        logsnr_low = logsnr_schedule_cosine_shifted(t, image_d, noise_d_low, logsnr_min, logsnr_max)
        logsnr_high = logsnr_schedule_cosine_shifted(t, image_d, noise_d_high, logsnr_min, logsnr_max)
        return torch.lerp(logsnr_low, logsnr_high, t)

    logsnr_min = -2 * math.log(min_value / sigma_data)
    logsnr_max = -2 * math.log(max_value / sigma_data)