Resolve v_prediction issue for min-SNR gamma weighted loss function (#5096)

* Resolve v_prediction issue for min-SNR gamma weighted loss function

* Combine MSE loss calculation of epsilon and velocity, with a note about the application of the epsilon code to sample prediction

* style

---------

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

Author: 3 people

examples/text_to_image/train_text_to_image_sdxl.py

@@ -332,15 +332,6 @@ def parse_args(input_args=None):
         help="SNR weighting gamma to be used if rebalancing the loss. Recommended value is 5.0. "
         "More details here: https://arxiv.org/abs/2303.09556.",
     )
-    parser.add_argument(
-        "--force_snr_gamma",
-        action="store_true",
-        help=(
-            "When using SNR gamma with rescaled betas for zero terminal SNR, a divide-by-zero error can cause NaN"
-            " condition when computing the SNR with a sigma value of zero. This parameter overrides the check,"
-            " allowing the use of SNR gamma with a terminal SNR model. Use with caution, and closely monitor results."
-        ),
-    )
     parser.add_argument("--use_ema", action="store_true", help="Whether to use EMA model.")
     parser.add_argument(
         "--allow_tf32",
@@ -554,18 +545,6 @@ def main(args):
     # Load scheduler and models
     noise_scheduler = DDPMScheduler.from_pretrained(args.pretrained_model_name_or_path, subfolder="scheduler")
     # Check for terminal SNR in combination with SNR Gamma
-    if (
-        args.snr_gamma
-        and not args.force_snr_gamma
-        and (
-            hasattr(noise_scheduler.config, "rescale_betas_zero_snr") and noise_scheduler.config.rescale_betas_zero_snr
-        )
-    ):
-        raise ValueError(
-            f"The selected noise scheduler for the model {args.pretrained_model_name_or_path} uses rescaled betas for zero SNR.\n"
-            "When this configuration is present, the parameter --snr_gamma may not be used without parameter --force_snr_gamma.\n"
-            "This is due to a mathematical incompatibility between our current SNR gamma implementation, and a sigma value of zero."
-        )
     text_encoder_one = text_encoder_cls_one.from_pretrained(
         args.pretrained_model_name_or_path, subfolder="text_encoder", revision=args.revision
     )
@@ -1013,9 +992,17 @@ def compute_time_ids(original_size, crops_coords_top_left):
                     # 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(timesteps)
-                    mse_loss_weights = (
+                    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
+                    else:
+                        # Epsilon and sample both use the same loss weights.
+                        mse_loss_weights = base_weight
+
                     # We first calculate the original loss. Then we mean over the non-batch dimensions and
                     # rebalance the sample-wise losses with their respective loss weights.
                     # Finally, we take the mean of the rebalanced loss.