Update README and clean-up the code for SD3 timesteps

README.md

@@ -14,6 +14,13 @@ The command to install PyTorch is as follows:
 
 ### Recent Updates
 
+Nov 7, 2024:
+
+- The distribution of timesteps during SD3/3.5 training has been adjusted. This applies to both fine-tuning and LoRA training. PR [#1768](https://github.com/kohya-ss/sd-scripts/pull/1768) Thanks to Dango233!
+  - Previously, the side closer to noise was more sampled, but now it is uniform by default. This may improve the problem of difficulty in learning details.
+  - Specifically, the problem of double shifting has been fixed. The default for `--weighting_scheme` has been changed to `uniform` (the previous default was `logit_normal`).
+  - A new option `--training_shift` has been added. The default is 1.0, and all timesteps are sampled uniformly. If less than 1.0, the side closer to the image is more sampled, and if more than 1.0, the side closer to noise is more sampled.
+
 Oct 31, 2024:
 
 - Added support for SD3.5L/M training. See [SD3 training](#sd3-training) for details.
@@ -641,6 +648,7 @@ Here are the arguments. The arguments and sample settings are still experimental
 - `--clip_l_dropout_rate`, `--clip_g_dropout_rate` and `--t5_dropout_rate` are the dropout rates for the embeddings of CLIP-L, CLIP-G, and T5XXL, described in [SAI research papre](http://arxiv.org/pdf/2403.03206). The default is 0.0. For LoRA training, it is seems to be better to set 0.0.
 - `--pos_emb_random_crop_rate` is the rate of random cropping of positional embeddings, described in [SD3.5M model card](https://huggingface.co/stabilityai/stable-diffusion-3.5-medium). The default is 0. It is seems to be better to set 0.0 for LoRA training.
 - `--enable_scaled_pos_embed` is to enable the scaled positional embeddings. The default is False. This option is an experimental feature for SD3.5M. Details are described below.
+- `--training_shift` is the shift value for the training distribution of timesteps. The default is 1.0 (uniform distribution, no shift).  If less than 1.0, the side closer to the image is more sampled, and if more than 1.0, the side closer to noise is more sampled. 
 
 Other options are described below.
 
@@ -681,7 +689,10 @@ Other options are described below.
    - Same as FLUX.1 for data preparation.
    - If you train with multiple resolutions, you can enable the scaled positional embeddings with `--enable_scaled_pos_embed`. The default is False. __This option is an experimental feature.__
 
-
+6. Weighting scheme and training shift:
+   - The weighting scheme is described in the section 3.1 of the [SD3 paper](https://arxiv.org/abs/2403.03206v1). 
+   - The uniform distribution is the default. If you want to change the distribution, see `--help` for options. 
+   - `--training_shift` is the shift value for the training distribution of timesteps.
 
 
 Technical details of multi-resolution training for SD3.5M:

library/config_util.py

@@ -526,7 +526,7 @@ def generate_dataset_group_by_blueprint(dataset_group_blueprint: DatasetGroupBlu
           secondary_separator: {subset.secondary_separator}
           enable_wildcard: {subset.enable_wildcard}
           caption_dropout_rate: {subset.caption_dropout_rate}
-          caption_dropout_every_n_epoches: {subset.caption_dropout_every_n_epochs}
+          caption_dropout_every_n_epochs: {subset.caption_dropout_every_n_epochs}
           caption_tag_dropout_rate: {subset.caption_tag_dropout_rate}
           caption_prefix: {subset.caption_prefix}
           caption_suffix: {subset.caption_suffix}

library/sd3_models.py

@@ -871,7 +871,7 @@ class MMDiT(nn.Module):
             # remove pos_embed to free up memory up to 0.4 GB
             self.pos_embed = None
 
-            # remove duplcates and sort latent sizes in ascending order
+            # remove duplicates and sort latent sizes in ascending order
             latent_sizes = list(set(latent_sizes))
             latent_sizes = sorted(latent_sizes)
 

library/sd3_train_utils.py

@@ -253,7 +253,7 @@ def add_sd3_training_arguments(parser: argparse.ArgumentParser):
         " / 複数解像度学習時に解像度ごとに位置埋め込みをスケーリングする。SD3.5M以外では予期しない動作になります",
     )
 
-    # Dependencies of Diffusers noise sampler has been removed for clearity.
+    # Dependencies of Diffusers noise sampler has been removed for clarity.
     parser.add_argument(
         "--weighting_scheme",
         type=str,
@@ -286,6 +286,7 @@ def add_sd3_training_arguments(parser: argparse.ArgumentParser):
         help="Discrete flow shift for training timestep distribution adjustment, applied in addition to the weighting scheme, default is 1.0. /タイムステップ分布のための離散フローシフト、重み付けスキームの上に適用される、デフォルトは1.0。",
     )
 
+
 def verify_sdxl_training_args(args: argparse.Namespace, supportTextEncoderCaching: bool = True):
     assert not args.v2, "v2 cannot be enabled in SDXL training / SDXL学習ではv2を有効にすることはできません"
     if args.v_parameterization:
@@ -956,9 +957,10 @@ def compute_loss_weighting_for_sd3(weighting_scheme: str, sigmas=None):
     return weighting
 
 
-def get_noisy_model_input_and_timesteps(
-    args, noise_scheduler, latents, noise, device, dtype
-) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+# endregion
+
+
+def get_noisy_model_input_and_timesteps(args, latents, noise, device, dtype) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
     bsz = latents.shape[0]
 
     # Sample a random timestep for each image
@@ -986,4 +988,3 @@ def get_noisy_model_input_and_timesteps(
     noisy_model_input = sigmas * noise + (1.0 - sigmas) * latents
 
     return noisy_model_input, timesteps, sigmas
-

sd3_train.py

@@ -811,8 +811,8 @@ def train(args):
     progress_bar = tqdm(range(args.max_train_steps), smoothing=0, disable=not accelerator.is_local_main_process, desc="steps")
     global_step = 0
 
-    noise_scheduler = sd3_train_utils.FlowMatchEulerDiscreteScheduler(num_train_timesteps=1000, shift=3.0)
-    noise_scheduler_copy = copy.deepcopy(noise_scheduler)
+    # noise_scheduler = sd3_train_utils.FlowMatchEulerDiscreteScheduler(num_train_timesteps=1000, shift=3.0)
+    # noise_scheduler_copy = copy.deepcopy(noise_scheduler)
 
     if accelerator.is_main_process:
         init_kwargs = {}
@@ -940,11 +940,11 @@ def train(args):
 
                 # Sample noise that we'll add to the latents
                 noise = torch.randn_like(latents)
-                bsz = latents.shape[0]
+                # bsz = latents.shape[0]
 
                 # get noisy model input and timesteps
                 noisy_model_input, timesteps, sigmas = sd3_train_utils.get_noisy_model_input_and_timesteps(
-                    args, noise_scheduler_copy, latents, noise, accelerator.device, weight_dtype
+                    args, latents, noise, accelerator.device, weight_dtype
                 )
 
                 # debug: NaN check for all inputs

sd3_train_network.py

@@ -275,9 +275,8 @@ class Sd3NetworkTrainer(train_network.NetworkTrainer):
         )
 
     def get_noise_scheduler(self, args: argparse.Namespace, device: torch.device) -> Any:
-        # shift 3.0 is the default value
-        noise_scheduler = sd3_train_utils.FlowMatchEulerDiscreteScheduler(num_train_timesteps=1000, shift=3.0)
-        self.noise_scheduler_copy = copy.deepcopy(noise_scheduler)
+        # this scheduler is not used in training, but used  to get num_train_timesteps etc.
+        noise_scheduler = sd3_train_utils.FlowMatchEulerDiscreteScheduler(num_train_timesteps=1000, shift=args.training_shift)
         return noise_scheduler
 
     def encode_images_to_latents(self, args, accelerator, vae, images):
@@ -304,7 +303,7 @@ class Sd3NetworkTrainer(train_network.NetworkTrainer):
 
         # get noisy model input and timesteps
         noisy_model_input, timesteps, sigmas = sd3_train_utils.get_noisy_model_input_and_timesteps(
-            args, self.noise_scheduler_copy, latents, noise, accelerator.device, weight_dtype
+            args, latents, noise, accelerator.device, weight_dtype
         )
 
         # ensure the hidden state will require grad