fix min-snr implementation (huggingface#8466)

* fix min-snr implementation

https://github.com/kohya-ss/sd-scripts/blob/main/library/custom_train_functions.py#L66

* Update train_dreambooth.py

fix variable name mse_loss_weights

* fix divisor

* make style

---------

Co-authored-by: Sayak Paul <[email protected]>
Co-authored-by: YiYi Xu <[email protected]>

Author: 3 people

examples/dreambooth/train_dreambooth.py

@@ -1300,16 +1300,17 @@ def compute_text_embeddings(prompt):
                     # Since we predict the noise instead of x_0, the original formulation is slightly changed.
                     # This is discussed in Section 4.2 of the same paper.
                     snr = compute_snr(noise_scheduler, timesteps)
-                    base_weight = (
-                        torch.stack([snr, args.snr_gamma * torch.ones_like(timesteps)], dim=1).min(dim=1)[0] / snr
-                    )
 
                     if noise_scheduler.config.prediction_type == "v_prediction":
                         # Velocity objective needs to be floored to an SNR weight of one.
-                        mse_loss_weights = base_weight + 1
+                        divisor = snr + 1
                     else:
-                        # Epsilon and sample both use the same loss weights.
-                        mse_loss_weights = base_weight
+                        divisor = snr
+
+                    mse_loss_weights = (
+                        torch.stack([snr, args.snr_gamma * torch.ones_like(timesteps)], dim=1).min(dim=1)[0] / divisor
+                    )
+
                     loss = F.mse_loss(model_pred.float(), target.float(), reduction="none")
                     loss = loss.mean(dim=list(range(1, len(loss.shape)))) * mse_loss_weights
                     loss = loss.mean()