Question about training loss in train_dreambooth_sd3.py

[ningpingwang]

Here is the code snippet of computing training loss diffusers/examples/dreambooth/train_dreambooth_sd3.py

noisy_model_input = (1.0 - sigmas) * model_input + sigmas * noise
model_pred = transformer(noisy_model_input, ...)
model_pred = model_pred * (-sigmas) + noisy_model_input
target = model_input
loss = torch.mean(
    (weighting.float() * (model_pred.float() - target.float()) ** 2).reshape(target.shape[0], -1),
    1,
)

In the final step, model_pred is sigmas * (noise - vanilla_model_pred) + (1.0 - sigmas) * model_input,
so model_pred - target is sigmas * (noise - vanilla_model_pred) - sigmas * model_input.
But from my understanding of flow match loss, it should be sigmas * (noise - vanilla_model_pred) without the term - sigmas * model_input.

Am I missing something?