fix: add eps to v-objective to hopefully fix NaNs

Author: flavioschneider

audio_diffusion_pytorch/diffusion.py

@@ -161,7 +161,7 @@ def forward(self, x: Tensor, noise: Tensor = None, **kwargs) -> Tensor:
         x_pred = self.net(c_in * x_noisy, self.sigma_to_t(sigmas), **kwargs)
 
         # Compute v-objective target
-        v_target = (x - c_skip * x_noisy) / c_out
+        v_target = (x - c_skip * x_noisy) / (c_out + 1e-7)
 
         # Compute loss
         loss = F.mse_loss(x_pred, v_target)