flow_loss implementation in HuggingFace may have an issue related to velocity computation.

[liq22]

Hi, thanks for your work.
Should the variable target be equal to target - noise in your flow loss?

This version seems to predict the target sample directly, rather than the velocity.

 def forward(self, target, z, mask=None, mask_y=None):

    noise = torch.randn_like(target)

    t = torch.rand(target.shape[0], device=target.device)

    noised_target = t[:, None] * target + (1 - t[:, None]) * noise

    predict_v = self.net(noised_target, t * 1000, z)

    weights = 1.0 / \
        torch.arange(1, self.in_channels + 1, dtype=torch.float32, device=target.device)
    if mask_y is not None:
        loss = (mask_y * weights * (predict_v - target) ** 2).sum(dim=-1)
    else:
        loss = (weights * (predict_v - target) ** 2).sum(dim=-1)

    if mask is not None:
        loss = (loss * mask).sum() / mask.sum()
    return loss.mean()

[WenWeiTHU]

Hi，we have a discussion here： https://huggingface.co/thuml/sundial-base-128m/discussions/2

[hvlgo]

You treat predict_v + noise as the new predict_v, effectively moving the noise addition step inside the model. As a result, during inference, you need to use pred - noise. Essentially, it's the same in principle—just two different implementations. The approach used in the code tends to make training more stable.

The main difference between Flow and standard diffusion lies in the way noise is added linearly; predicting the velocity field is not the fundamental distinction.