Untitled

In stable_audio_tools/training/diffusion.py, change:

        elif self.diffusion_objective == "rectified_flow":
            targets = noise - diffusion_input

to

        elif self.diffusion_objective == "rectified_flow":
            t, noised_inputs, targets = self.flow_matcher.sample_location_and_conditional_flow(diffusion_input, noise, t)

The flow_matcher has to be initialized like this:

        import torchcfm as cfm
        self.flow_matcher = cfm.ExactOptimalTransportConditionalFlowMatcher(sigma=1e-4)