feat: add general synthetic data iterator and examples for WAN and FLUX.

Author: root

src/maxdiffusion/configs/base_flux_dev.yml

@@ -177,7 +177,20 @@ allow_split_physical_axes: False
 # Replace with dataset path or train_data_dir. One has to be set.
 dataset_name: 'diffusers/pokemon-gpt4-captions'
 train_split: 'train'
-dataset_type: 'tf'
+dataset_type: 'tfrecord'  # Options: 'tfrecord', 'hf', 'tf', 'grain', 'synthetic'
+# ==============================================================================
+# Synthetic Data Configuration (only used when dataset_type='synthetic')
+# ==============================================================================
+# To use synthetic data for testing/debugging without real datasets:
+# 1. Set dataset_type: 'synthetic' above
+# 2. Optionally set synthetic_num_samples (null=infinite, or a number like 10000)
+# 3. Optionally override dimensions
+#
+# synthetic_num_samples: null  # null for infinite, or set a number
+#
+# Optional dimension overrides:
+# resolution: 512
+# ==============================================================================
 cache_latents_text_encoder_outputs: True
 # cache_latents_text_encoder_outputs only apply to dataset_type="tf",
 # only apply to small dataset that fits in memory

src/maxdiffusion/configs/base_flux_dev_multi_res.yml

@@ -199,7 +199,28 @@ allow_split_physical_axes: False
 # Replace with dataset path or train_data_dir. One has to be set.
 dataset_name: 'diffusers/pokemon-gpt4-captions'
 train_split: 'train'
-dataset_type: 'tfrecord'
+dataset_type: 'tfrecord'  # Options: 'tfrecord', 'hf', 'tf', 'grain', 'synthetic'
+# ==============================================================================
+# Synthetic Data Configuration (only used when dataset_type='synthetic')
+# ==============================================================================
+# To use synthetic data for testing/debugging without real datasets:
+# 1. Set dataset_type: 'synthetic' above
+# 2. Optionally set synthetic_num_samples (null=infinite, or a number like 10000)
+# 3. Optionally override dimensions with synthetic_override_* flags below
+#
+# synthetic_num_samples: null  # null for infinite, or set a number
+#
+# Optional dimension overrides (comment out to use pipeline/config values):
+# synthetic_override_height: 720
+# synthetic_override_width: 1280
+# synthetic_override_num_frames: 121
+# synthetic_override_max_sequence_length: 512
+# synthetic_override_text_embed_dim: 4096
+# synthetic_override_num_channels_latents: 16
+# synthetic_override_vae_scale_factor_spatial: 8
+# synthetic_override_vae_scale_factor_temporal: 4
+# ==============================================================================
+
 cache_latents_text_encoder_outputs: True
 # cache_latents_text_encoder_outputs only apply to dataset_type="tf",
 # only apply to small dataset that fits in memory