add configs?

configs/ae/video.yaml

@@ -0,0 +1,35 @@
+target: sgm.models.autoencoder.AutoencodingEngine
+params:
+  loss_config:
+    target: torch.nn.Identity
+  regularizer_config:
+    target: sgm.modules.autoencoding.regularizers.DiagonalGaussianRegularizer
+  encoder_config: 
+    target: sgm.modules.diffusionmodules.model.Encoder
+    params:
+      attn_type: vanilla
+      double_z: True
+      z_channels: 4
+      resolution: 256
+      in_channels: 3
+      out_ch: 3
+      ch: 128
+      ch_mult: [1, 2, 4, 4]
+      num_res_blocks: 2
+      attn_resolutions: []
+      dropout: 0.0
+  decoder_config:
+    target: sgm.modules.autoencoding.temporal_ae.VideoDecoder
+    params:
+      attn_type: vanilla
+      double_z: True
+      z_channels: 4
+      resolution: 256
+      in_channels: 3
+      out_ch: 3
+      ch: 128
+      ch_mult: [1, 2, 4, 4]
+      num_res_blocks: 2
+      attn_resolutions: []
+      dropout: 0.0
+      video_kernel_size: [3, 1, 1]

configs/embedder/clip_image.yaml

@@ -0,0 +1,8 @@
+target: sgm.modules.encoders.modules.FrozenOpenCLIPImagePredictionEmbedder
+params:
+  n_cond_frames: 1
+  n_copies: 1
+  open_clip_embedding_config:
+    target: sgm.modules.encoders.modules.FrozenOpenCLIPImageEmbedder
+    params:
+      freeze: True

configs/example_training/autoencoder/kl-f4/imagenet-attnfree-logvar.yaml

@@ -0,0 +1,104 @@
+model:
+  base_learning_rate: 4.5e-6
+  target: sgm.models.autoencoder.AutoencodingEngine
+  params:
+    input_key: jpg
+    monitor: val/rec_loss
+
+    loss_config:
+      target: sgm.modules.autoencoding.losses.GeneralLPIPSWithDiscriminator
+      params:
+        perceptual_weight: 0.25
+        disc_start: 20001
+        disc_weight: 0.5
+        learn_logvar: True
+
+        regularization_weights:
+          kl_loss: 1.0
+
+    regularizer_config:
+      target: sgm.modules.autoencoding.regularizers.DiagonalGaussianRegularizer
+
+    encoder_config:
+      target: sgm.modules.diffusionmodules.model.Encoder
+      params:
+        attn_type: none
+        double_z: True
+        z_channels: 4
+        resolution: 256
+        in_channels: 3
+        out_ch: 3
+        ch: 128
+        ch_mult: [1, 2, 4]
+        num_res_blocks: 4
+        attn_resolutions: []
+        dropout: 0.0
+
+    decoder_config:
+      target: sgm.modules.diffusionmodules.model.Decoder
+      params: ${model.params.encoder_config.params}
+
+data:
+  target: sgm.data.dataset.StableDataModuleFromConfig
+  params:
+    train:
+      datapipeline:
+        urls:
+          - DATA-PATH
+        pipeline_config:
+          shardshuffle: 10000
+          sample_shuffle: 10000
+
+        decoders:
+          - pil
+
+        postprocessors:
+          - target: sdata.mappers.TorchVisionImageTransforms
+            params:
+              key: jpg
+              transforms:
+                - target: torchvision.transforms.Resize
+                  params:
+                    size: 256
+                    interpolation: 3
+                - target: torchvision.transforms.ToTensor
+          - target: sdata.mappers.Rescaler
+          - target: sdata.mappers.AddOriginalImageSizeAsTupleAndCropToSquare
+            params:
+              h_key: height
+              w_key: width
+
+      loader:
+        batch_size: 8
+        num_workers: 4
+
+
+lightning:
+  strategy:
+    target: pytorch_lightning.strategies.DDPStrategy
+    params:
+      find_unused_parameters: True
+
+  modelcheckpoint:
+    params:
+      every_n_train_steps: 5000
+
+  callbacks:
+    metrics_over_trainsteps_checkpoint:
+      params:
+        every_n_train_steps: 50000
+
+    image_logger:
+      target: main.ImageLogger
+      params:
+        enable_autocast: False
+        batch_frequency: 1000
+        max_images: 8
+        increase_log_steps: True
+
+  trainer:
+    devices: 0,
+    limit_val_batches: 50
+    benchmark: True
+    accumulate_grad_batches: 1
+    val_check_interval: 10000

configs/example_training/autoencoder/kl-f4/imagenet-kl_f8_8chn.yaml

@@ -0,0 +1,105 @@
+model:
+  base_learning_rate: 4.5e-6
+  target: sgm.models.autoencoder.AutoencodingEngine
+  params:
+    input_key: jpg
+    monitor: val/loss/rec
+    disc_start_iter: 0
+
+    encoder_config:
+      target: sgm.modules.diffusionmodules.model.Encoder
+      params:
+        attn_type: vanilla-xformers
+        double_z: true
+        z_channels: 8
+        resolution: 256
+        in_channels: 3
+        out_ch: 3
+        ch: 128
+        ch_mult: [1, 2, 4, 4]
+        num_res_blocks: 2
+        attn_resolutions: []
+        dropout: 0.0
+
+    decoder_config:
+      target: sgm.modules.diffusionmodules.model.Decoder
+      params: ${model.params.encoder_config.params}
+
+    regularizer_config:
+      target: sgm.modules.autoencoding.regularizers.DiagonalGaussianRegularizer
+
+    loss_config:
+      target: sgm.modules.autoencoding.losses.GeneralLPIPSWithDiscriminator
+      params:
+        perceptual_weight: 0.25
+        disc_start: 20001
+        disc_weight: 0.5
+        learn_logvar: True
+
+        regularization_weights:
+          kl_loss: 1.0
+
+data:
+  target: sgm.data.dataset.StableDataModuleFromConfig
+  params:
+    train:
+      datapipeline:
+        urls:
+          - DATA-PATH
+        pipeline_config:
+          shardshuffle: 10000
+          sample_shuffle: 10000
+
+        decoders:
+          - pil
+
+        postprocessors:
+          - target: sdata.mappers.TorchVisionImageTransforms
+            params:
+              key: jpg
+              transforms:
+                - target: torchvision.transforms.Resize
+                  params:
+                    size: 256
+                    interpolation: 3
+                - target: torchvision.transforms.ToTensor
+          - target: sdata.mappers.Rescaler
+          - target: sdata.mappers.AddOriginalImageSizeAsTupleAndCropToSquare
+            params:
+              h_key: height
+              w_key: width
+
+      loader:
+        batch_size: 8
+        num_workers: 4
+
+
+lightning:
+  strategy:
+    target: pytorch_lightning.strategies.DDPStrategy
+    params:
+      find_unused_parameters: True
+
+  modelcheckpoint:
+    params:
+      every_n_train_steps: 5000
+
+  callbacks:
+    metrics_over_trainsteps_checkpoint:
+      params:
+        every_n_train_steps: 50000
+
+    image_logger:
+      target: main.ImageLogger
+      params:
+        enable_autocast: False
+        batch_frequency: 1000
+        max_images: 8
+        increase_log_steps: True
+
+  trainer:
+    devices: 0,
+    limit_val_batches: 50
+    benchmark: True
+    accumulate_grad_batches: 1
+    val_check_interval: 10000

configs/example_training/imagenet-f8_cond.yaml

@@ -0,0 +1,185 @@
+model:
+  base_learning_rate: 1.0e-4
+  target: sgm.models.diffusion.DiffusionEngine
+  params:
+    scale_factor: 0.13025
+    disable_first_stage_autocast: True
+    log_keys:
+      - cls
+
+    scheduler_config:
+      target: sgm.lr_scheduler.LambdaLinearScheduler
+      params:
+        warm_up_steps: [10000]
+        cycle_lengths: [10000000000000]
+        f_start: [1.e-6]
+        f_max: [1.]
+        f_min: [1.]
+
+    denoiser_config:
+      target: sgm.modules.diffusionmodules.denoiser.DiscreteDenoiser
+      params:
+        num_idx: 1000
+
+        scaling_config:
+          target: sgm.modules.diffusionmodules.denoiser_scaling.EpsScaling
+        discretization_config:
+          target: sgm.modules.diffusionmodules.discretizer.LegacyDDPMDiscretization
+
+    network_config:
+      target: sgm.modules.diffusionmodules.openaimodel.UNetModel
+      params:
+        use_checkpoint: True
+        in_channels: 4
+        out_channels: 4
+        model_channels: 256
+        attention_resolutions: [1, 2, 4]
+        num_res_blocks: 2
+        channel_mult: [1, 2, 4]
+        num_head_channels: 64
+        num_classes: sequential
+        adm_in_channels: 1024
+        transformer_depth: 1
+        context_dim: 1024
+        spatial_transformer_attn_type: softmax-xformers
+
+    conditioner_config:
+      target: sgm.modules.GeneralConditioner
+      params:
+        emb_models:
+          - is_trainable: True
+            input_key: cls
+            ucg_rate: 0.2
+            target: sgm.modules.encoders.modules.ClassEmbedder
+            params:
+              add_sequence_dim: True
+              embed_dim: 1024
+              n_classes: 1000
+
+          - is_trainable: False
+            ucg_rate: 0.2
+            input_key: original_size_as_tuple
+            target: sgm.modules.encoders.modules.ConcatTimestepEmbedderND
+            params:
+              outdim: 256
+
+          - is_trainable: False
+            input_key: crop_coords_top_left
+            ucg_rate: 0.2
+            target: sgm.modules.encoders.modules.ConcatTimestepEmbedderND
+            params:
+              outdim: 256
+
+    first_stage_config:
+      target: sgm.models.autoencoder.AutoencoderKL
+      params:
+        ckpt_path: CKPT_PATH
+        embed_dim: 4
+        monitor: val/rec_loss
+        ddconfig:
+          attn_type: vanilla-xformers
+          double_z: true
+          z_channels: 4
+          resolution: 256
+          in_channels: 3
+          out_ch: 3
+          ch: 128
+          ch_mult: [1, 2, 4, 4]
+          num_res_blocks: 2
+          attn_resolutions: []
+          dropout: 0.0
+        lossconfig:
+          target: torch.nn.Identity
+
+    loss_fn_config:
+      target: sgm.modules.diffusionmodules.loss.StandardDiffusionLoss
+      params:        
+        loss_weighting_config:
+          target: sgm.modules.diffusionmodules.loss_weighting.EpsWeighting
+        sigma_sampler_config:
+          target: sgm.modules.diffusionmodules.sigma_sampling.DiscreteSampling
+          params:
+            num_idx: 1000
+
+            discretization_config:
+              target: sgm.modules.diffusionmodules.discretizer.LegacyDDPMDiscretization
+
+    sampler_config:
+      target: sgm.modules.diffusionmodules.sampling.EulerEDMSampler
+      params:
+        num_steps: 50
+
+        discretization_config:
+          target: sgm.modules.diffusionmodules.discretizer.LegacyDDPMDiscretization
+
+        guider_config:
+          target: sgm.modules.diffusionmodules.guiders.VanillaCFG
+          params:
+            scale: 5.0
+
+data:
+  target: sgm.data.dataset.StableDataModuleFromConfig
+  params:
+    train:
+      datapipeline:
+        urls:
+          # USER: adapt this path the root of your custom dataset
+          - DATA_PATH
+        pipeline_config:
+          shardshuffle: 10000
+          sample_shuffle: 10000 # USER: you might wanna adapt depending on your available RAM
+
+        decoders:
+          - pil
+
+        postprocessors:
+          - target: sdata.mappers.TorchVisionImageTransforms
+            params:
+              key: jpg # USER: you might wanna adapt this for your custom dataset
+              transforms:
+                - target: torchvision.transforms.Resize
+                  params:
+                    size: 256
+                    interpolation: 3
+                - target: torchvision.transforms.ToTensor
+          - target: sdata.mappers.Rescaler
+
+          - target: sdata.mappers.AddOriginalImageSizeAsTupleAndCropToSquare
+            params:
+              h_key: height # USER: you might wanna adapt this for your custom dataset
+              w_key: width # USER: you might wanna adapt this for your custom dataset
+
+      loader:
+        batch_size: 64
+        num_workers: 6
+
+lightning:
+  modelcheckpoint:
+    params:
+      every_n_train_steps: 5000
+
+  callbacks:
+    metrics_over_trainsteps_checkpoint:
+      params:
+        every_n_train_steps: 25000
+
+    image_logger:
+      target: main.ImageLogger
+      params:
+        disabled: False
+        enable_autocast: False
+        batch_frequency: 1000
+        max_images: 8
+        increase_log_steps: True
+        log_first_step: False
+        log_images_kwargs:
+          use_ema_scope: False
+          N: 8
+          n_rows: 2
+
+  trainer:
+    devices: 0,
+    benchmark: True
+    num_sanity_val_steps: 0
+    accumulate_grad_batches: 1
+    max_epochs: 1000

configs/example_training/toy/cifar10_cond.yaml

@@ -0,0 +1,98 @@
+model:
+  base_learning_rate: 1.0e-4
+  target: sgm.models.diffusion.DiffusionEngine
+  params:
+    denoiser_config:
+      target: sgm.modules.diffusionmodules.denoiser.Denoiser
+      params:
+        scaling_config:
+          target: sgm.modules.diffusionmodules.denoiser_scaling.EDMScaling
+          params:
+            sigma_data: 1.0
+
+    network_config:
+      target: sgm.modules.diffusionmodules.openaimodel.UNetModel
+      params:
+        in_channels: 3
+        out_channels: 3
+        model_channels: 32
+        attention_resolutions: []
+        num_res_blocks: 4
+        channel_mult: [1, 2, 2]
+        num_head_channels: 32
+        num_classes: sequential
+        adm_in_channels: 128
+
+    conditioner_config:
+      target: sgm.modules.GeneralConditioner
+      params:
+        emb_models:
+          - is_trainable: True
+            input_key: cls
+            ucg_rate: 0.2
+            target: sgm.modules.encoders.modules.ClassEmbedder
+            params:
+              embed_dim: 128
+              n_classes: 10
+
+    first_stage_config:
+      target: sgm.models.autoencoder.IdentityFirstStage
+
+    loss_fn_config:
+      target: sgm.modules.diffusionmodules.loss.StandardDiffusionLoss
+      params:
+        loss_weighting_config:
+          target: sgm.modules.diffusionmodules.loss_weighting.EDMWeighting
+          params:
+            sigma_data: 1.0
+        sigma_sampler_config:
+          target: sgm.modules.diffusionmodules.sigma_sampling.EDMSampling
+
+    sampler_config:
+      target: sgm.modules.diffusionmodules.sampling.EulerEDMSampler
+      params:
+        num_steps: 50
+
+        discretization_config:
+          target: sgm.modules.diffusionmodules.discretizer.EDMDiscretization
+
+        guider_config:
+          target: sgm.modules.diffusionmodules.guiders.VanillaCFG
+          params:
+            scale: 3.0
+
+data:
+  target: sgm.data.cifar10.CIFAR10Loader
+  params:
+    batch_size: 512
+    num_workers: 1
+
+lightning:
+  modelcheckpoint:
+    params:
+      every_n_train_steps: 5000
+
+  callbacks:
+    metrics_over_trainsteps_checkpoint:
+      params:
+        every_n_train_steps: 25000
+
+    image_logger:
+      target: main.ImageLogger
+      params:
+        disabled: False
+        batch_frequency: 1000
+        max_images: 64
+        increase_log_steps: True
+        log_first_step: False
+        log_images_kwargs:
+          use_ema_scope: False
+          N: 64
+          n_rows: 8
+
+  trainer:
+    devices: 0,
+    benchmark: True
+    num_sanity_val_steps: 0
+    accumulate_grad_batches: 1
+    max_epochs: 20

configs/example_training/toy/mnist.yaml

@@ -0,0 +1,79 @@
+model:
+  base_learning_rate: 1.0e-4
+  target: sgm.models.diffusion.DiffusionEngine
+  params:
+    denoiser_config:
+      target: sgm.modules.diffusionmodules.denoiser.Denoiser
+      params:
+        scaling_config:
+          target: sgm.modules.diffusionmodules.denoiser_scaling.EDMScaling
+          params:
+            sigma_data: 1.0
+
+    network_config:
+      target: sgm.modules.diffusionmodules.openaimodel.UNetModel
+      params:
+        in_channels: 1
+        out_channels: 1
+        model_channels: 32
+        attention_resolutions: []
+        num_res_blocks: 4
+        channel_mult: [1, 2, 2]
+        num_head_channels: 32
+
+    first_stage_config:
+      target: sgm.models.autoencoder.IdentityFirstStage
+
+    loss_fn_config:
+      target: sgm.modules.diffusionmodules.loss.StandardDiffusionLoss
+      params:
+        loss_weighting_config:
+          target: sgm.modules.diffusionmodules.loss_weighting.EDMWeighting
+          params:
+            sigma_data: 1.0
+        sigma_sampler_config:
+          target: sgm.modules.diffusionmodules.sigma_sampling.EDMSampling
+
+    sampler_config:
+      target: sgm.modules.diffusionmodules.sampling.EulerEDMSampler
+      params:
+        num_steps: 50
+
+        discretization_config:
+          target: sgm.modules.diffusionmodules.discretizer.EDMDiscretization
+
+data:
+  target: sgm.data.mnist.MNISTLoader
+  params:
+    batch_size: 512
+    num_workers: 1
+
+lightning:
+  modelcheckpoint:
+    params:
+      every_n_train_steps: 5000
+
+  callbacks:
+    metrics_over_trainsteps_checkpoint:
+      params:
+        every_n_train_steps: 25000
+
+    image_logger:
+      target: main.ImageLogger
+      params:
+        disabled: False
+        batch_frequency: 1000
+        max_images: 64
+        increase_log_steps: False
+        log_first_step: False
+        log_images_kwargs:
+          use_ema_scope: False
+          N: 64
+          n_rows: 8
+
+  trainer:
+    devices: 0,
+    benchmark: True
+    num_sanity_val_steps: 0
+    accumulate_grad_batches: 1
+    max_epochs: 10

configs/example_training/toy/mnist_cond.yaml

@@ -0,0 +1,98 @@
+model:
+  base_learning_rate: 1.0e-4
+  target: sgm.models.diffusion.DiffusionEngine
+  params:
+    denoiser_config:
+      target: sgm.modules.diffusionmodules.denoiser.Denoiser
+      params:
+        scaling_config:
+          target: sgm.modules.diffusionmodules.denoiser_scaling.EDMScaling
+          params:
+            sigma_data: 1.0
+
+    network_config:
+      target: sgm.modules.diffusionmodules.openaimodel.UNetModel
+      params:
+        in_channels: 1
+        out_channels: 1
+        model_channels: 32
+        attention_resolutions: []
+        num_res_blocks: 4
+        channel_mult: [1, 2, 2]
+        num_head_channels: 32
+        num_classes: sequential
+        adm_in_channels: 128
+
+    conditioner_config:
+      target: sgm.modules.GeneralConditioner
+      params:
+        emb_models:
+          - is_trainable: True
+            input_key: cls
+            ucg_rate: 0.2
+            target: sgm.modules.encoders.modules.ClassEmbedder
+            params:
+              embed_dim: 128
+              n_classes: 10
+
+    first_stage_config:
+      target: sgm.models.autoencoder.IdentityFirstStage
+
+    loss_fn_config:
+      target: sgm.modules.diffusionmodules.loss.StandardDiffusionLoss
+      params:
+        loss_weighting_config:
+          target: sgm.modules.diffusionmodules.loss_weighting.EDMWeighting
+          params:
+            sigma_data: 1.0
+        sigma_sampler_config:
+          target: sgm.modules.diffusionmodules.sigma_sampling.EDMSampling
+
+    sampler_config:
+      target: sgm.modules.diffusionmodules.sampling.EulerEDMSampler
+      params:
+        num_steps: 50
+
+        discretization_config:
+          target: sgm.modules.diffusionmodules.discretizer.EDMDiscretization
+
+        guider_config:
+          target: sgm.modules.diffusionmodules.guiders.VanillaCFG
+          params:
+            scale: 3.0
+
+data:
+  target: sgm.data.mnist.MNISTLoader
+  params:
+    batch_size: 512
+    num_workers: 1
+
+lightning:
+  modelcheckpoint:
+    params:
+      every_n_train_steps: 5000
+
+  callbacks:
+    metrics_over_trainsteps_checkpoint:
+      params:
+        every_n_train_steps: 25000
+
+    image_logger:
+      target: main.ImageLogger
+      params:
+        disabled: False
+        batch_frequency: 1000
+        max_images: 16
+        increase_log_steps: True
+        log_first_step: False
+        log_images_kwargs:
+          use_ema_scope: False
+          N: 16
+          n_rows: 4
+
+  trainer:
+    devices: 0,
+    benchmark: True
+    num_sanity_val_steps: 0
+    accumulate_grad_batches: 1
+    max_epochs: 20

configs/example_training/toy/mnist_cond_discrete_eps.yaml

@@ -0,0 +1,103 @@
+model:
+  base_learning_rate: 1.0e-4
+  target: sgm.models.diffusion.DiffusionEngine
+  params:
+    denoiser_config:
+      target: sgm.modules.diffusionmodules.denoiser.DiscreteDenoiser
+      params:
+        num_idx: 1000
+
+        scaling_config:
+          target: sgm.modules.diffusionmodules.denoiser_scaling.EDMScaling
+        discretization_config:
+          target: sgm.modules.diffusionmodules.discretizer.LegacyDDPMDiscretization
+
+    network_config:
+      target: sgm.modules.diffusionmodules.openaimodel.UNetModel
+      params:
+        in_channels: 1
+        out_channels: 1
+        model_channels: 32
+        attention_resolutions: []
+        num_res_blocks: 4
+        channel_mult: [1, 2, 2]
+        num_head_channels: 32
+        num_classes: sequential
+        adm_in_channels: 128
+
+    conditioner_config:
+      target: sgm.modules.GeneralConditioner
+      params:
+        emb_models:
+          - is_trainable: True
+            input_key: cls
+            ucg_rate: 0.2
+            target: sgm.modules.encoders.modules.ClassEmbedder
+            params:
+              embed_dim: 128
+              n_classes: 10
+
+    first_stage_config:
+      target: sgm.models.autoencoder.IdentityFirstStage
+
+    loss_fn_config:
+      target: sgm.modules.diffusionmodules.loss.StandardDiffusionLoss
+      params:
+        loss_weighting_config:
+          target: sgm.modules.diffusionmodules.loss_weighting.EDMWeighting
+        sigma_sampler_config:
+          target: sgm.modules.diffusionmodules.sigma_sampling.DiscreteSampling
+          params:
+            num_idx: 1000
+
+            discretization_config:
+              target: sgm.modules.diffusionmodules.discretizer.LegacyDDPMDiscretization
+
+    sampler_config:
+      target: sgm.modules.diffusionmodules.sampling.EulerEDMSampler
+      params:
+        num_steps: 50
+
+        discretization_config:
+          target: sgm.modules.diffusionmodules.discretizer.LegacyDDPMDiscretization
+
+        guider_config:
+          target: sgm.modules.diffusionmodules.guiders.VanillaCFG
+          params:
+            scale: 5.0
+
+data:
+  target: sgm.data.mnist.MNISTLoader
+  params:
+    batch_size: 512
+    num_workers: 1
+
+lightning:
+  modelcheckpoint:
+    params:
+      every_n_train_steps: 5000
+
+  callbacks:
+    metrics_over_trainsteps_checkpoint:
+      params:
+        every_n_train_steps: 25000
+
+    image_logger:
+      target: main.ImageLogger
+      params:
+        disabled: False
+        batch_frequency: 1000
+        max_images: 16
+        increase_log_steps: True
+        log_first_step: False
+        log_images_kwargs:
+          use_ema_scope: False
+          N: 16
+          n_rows: 4
+
+  trainer:
+    devices: 0,
+    benchmark: True
+    num_sanity_val_steps: 0
+    accumulate_grad_batches: 1
+    max_epochs: 20

configs/example_training/toy/mnist_cond_l1_loss.yaml

@@ -0,0 +1,99 @@
+model:
+  base_learning_rate: 1.0e-4
+  target: sgm.models.diffusion.DiffusionEngine
+  params:
+    denoiser_config:
+      target: sgm.modules.diffusionmodules.denoiser.Denoiser
+      params:
+        scaling_config:
+          target: sgm.modules.diffusionmodules.denoiser_scaling.EDMScaling
+          params:
+            sigma_data: 1.0
+
+    network_config:
+      target: sgm.modules.diffusionmodules.openaimodel.UNetModel
+      params:
+        in_channels: 1
+        out_channels: 1
+        model_channels: 32
+        attention_resolutions: []
+        num_res_blocks: 4
+        channel_mult: [1, 2, 2]
+        num_head_channels: 32
+        num_classes: sequential
+        adm_in_channels: 128
+
+    conditioner_config:
+      target: sgm.modules.GeneralConditioner
+      params:
+        emb_models:
+          - is_trainable: True
+            input_key: cls
+            ucg_rate: 0.2
+            target: sgm.modules.encoders.modules.ClassEmbedder
+            params:
+              embed_dim: 128
+              n_classes: 10
+
+    first_stage_config:
+      target: sgm.models.autoencoder.IdentityFirstStage
+
+    loss_fn_config:
+      target: sgm.modules.diffusionmodules.loss.StandardDiffusionLoss
+      params:
+        loss_type: l1
+        loss_weighting_config:
+          target: sgm.modules.diffusionmodules.loss_weighting.EDMWeighting
+          params:
+            sigma_data: 1.0
+        sigma_sampler_config:
+          target: sgm.modules.diffusionmodules.sigma_sampling.EDMSampling
+
+    sampler_config:
+      target: sgm.modules.diffusionmodules.sampling.EulerEDMSampler
+      params:
+        num_steps: 50
+
+        discretization_config:
+          target: sgm.modules.diffusionmodules.discretizer.EDMDiscretization
+
+        guider_config:
+          target: sgm.modules.diffusionmodules.guiders.VanillaCFG
+          params:
+            scale: 3.0
+
+data:
+  target: sgm.data.mnist.MNISTLoader
+  params:
+    batch_size: 512
+    num_workers: 1
+
+lightning:
+  modelcheckpoint:
+    params:
+      every_n_train_steps: 5000
+
+  callbacks:
+    metrics_over_trainsteps_checkpoint:
+      params:
+        every_n_train_steps: 25000
+
+    image_logger:
+      target: main.ImageLogger
+      params:
+        disabled: False
+        batch_frequency: 1000
+        max_images: 64
+        increase_log_steps: True
+        log_first_step: False
+        log_images_kwargs:
+          use_ema_scope: False
+          N: 64
+          n_rows: 8
+
+  trainer:
+    devices: 0,
+    benchmark: True
+    num_sanity_val_steps: 0
+    accumulate_grad_batches: 1
+    max_epochs: 20

configs/example_training/toy/mnist_cond_with_ema.yaml

@@ -0,0 +1,100 @@
+model:
+  base_learning_rate: 1.0e-4
+  target: sgm.models.diffusion.DiffusionEngine
+  params:
+    use_ema: True
+
+    denoiser_config:
+      target: sgm.modules.diffusionmodules.denoiser.Denoiser
+      params:
+        scaling_config:
+          target: sgm.modules.diffusionmodules.denoiser_scaling.EDMScaling
+          params:
+            sigma_data: 1.0
+
+    network_config:
+      target: sgm.modules.diffusionmodules.openaimodel.UNetModel
+      params:
+        in_channels: 1
+        out_channels: 1
+        model_channels: 32
+        attention_resolutions: []
+        num_res_blocks: 4
+        channel_mult: [1, 2, 2]
+        num_head_channels: 32
+        num_classes: sequential
+        adm_in_channels: 128
+
+    conditioner_config:
+      target: sgm.modules.GeneralConditioner
+      params:
+        emb_models:
+          - is_trainable: True
+            input_key: cls
+            ucg_rate: 0.2
+            target: sgm.modules.encoders.modules.ClassEmbedder
+            params:
+              embed_dim: 128
+              n_classes: 10
+
+    first_stage_config:
+      target: sgm.models.autoencoder.IdentityFirstStage
+
+    loss_fn_config:
+      target: sgm.modules.diffusionmodules.loss.StandardDiffusionLoss
+      params:
+        loss_weighting_config:
+          target: sgm.modules.diffusionmodules.loss_weighting.EDMWeighting
+          params:
+            sigma_data: 1.0
+        sigma_sampler_config:
+          target: sgm.modules.diffusionmodules.sigma_sampling.EDMSampling
+
+    sampler_config:
+      target: sgm.modules.diffusionmodules.sampling.EulerEDMSampler
+      params:
+        num_steps: 50
+
+        discretization_config:
+          target: sgm.modules.diffusionmodules.discretizer.EDMDiscretization
+
+        guider_config:
+          target: sgm.modules.diffusionmodules.guiders.VanillaCFG
+          params:
+            scale: 3.0
+
+data:
+  target: sgm.data.mnist.MNISTLoader
+  params:
+    batch_size: 512
+    num_workers: 1
+
+lightning:
+  modelcheckpoint:
+    params:
+      every_n_train_steps: 5000
+
+  callbacks:
+    metrics_over_trainsteps_checkpoint:
+      params:
+        every_n_train_steps: 25000
+
+    image_logger:
+      target: main.ImageLogger
+      params:
+        disabled: False
+        batch_frequency: 1000
+        max_images: 64
+        increase_log_steps: True
+        log_first_step: False
+        log_images_kwargs:
+          use_ema_scope: False
+          N: 64
+          n_rows: 8
+
+  trainer:
+    devices: 0,
+    benchmark: True
+    num_sanity_val_steps: 0
+    accumulate_grad_batches: 1
+    max_epochs: 20

configs/example_training/txt2img-clipl-legacy-ucg-training.yaml

@@ -0,0 +1,182 @@
+model:
+  base_learning_rate: 1.0e-4
+  target: sgm.models.diffusion.DiffusionEngine
+  params:
+    scale_factor: 0.13025
+    disable_first_stage_autocast: True
+    log_keys:
+      - txt
+
+    scheduler_config:
+      target: sgm.lr_scheduler.LambdaLinearScheduler
+      params:
+        warm_up_steps: [10000]
+        cycle_lengths: [10000000000000]
+        f_start: [1.e-6]
+        f_max: [1.]
+        f_min: [1.]
+
+    denoiser_config:
+      target: sgm.modules.diffusionmodules.denoiser.DiscreteDenoiser
+      params:
+        num_idx: 1000
+
+        scaling_config:
+          target: sgm.modules.diffusionmodules.denoiser_scaling.EpsScaling
+        discretization_config:
+          target: sgm.modules.diffusionmodules.discretizer.LegacyDDPMDiscretization
+
+    network_config:
+      target: sgm.modules.diffusionmodules.openaimodel.UNetModel
+      params:
+        use_checkpoint: True
+        in_channels: 4
+        out_channels: 4
+        model_channels: 320
+        attention_resolutions: [1, 2, 4]
+        num_res_blocks: 2
+        channel_mult: [1, 2, 4, 4]
+        num_head_channels: 64
+        num_classes: sequential
+        adm_in_channels: 1792
+        num_heads: 1
+        transformer_depth: 1
+        context_dim: 768
+        spatial_transformer_attn_type: softmax-xformers
+
+    conditioner_config:
+      target: sgm.modules.GeneralConditioner
+      params:
+        emb_models:
+          - is_trainable: True
+            input_key: txt
+            ucg_rate: 0.1
+            legacy_ucg_value: ""
+            target: sgm.modules.encoders.modules.FrozenCLIPEmbedder
+            params:
+              always_return_pooled: True
+
+          - is_trainable: False
+            ucg_rate: 0.1
+            input_key: original_size_as_tuple
+            target: sgm.modules.encoders.modules.ConcatTimestepEmbedderND
+            params:
+              outdim: 256
+
+          - is_trainable: False
+            input_key: crop_coords_top_left
+            ucg_rate: 0.1
+            target: sgm.modules.encoders.modules.ConcatTimestepEmbedderND
+            params:
+              outdim: 256
+
+    first_stage_config:
+      target: sgm.models.autoencoder.AutoencoderKL
+      params:
+        ckpt_path: CKPT_PATH
+        embed_dim: 4
+        monitor: val/rec_loss
+        ddconfig:
+          attn_type: vanilla-xformers
+          double_z: true
+          z_channels: 4
+          resolution: 256
+          in_channels: 3
+          out_ch: 3
+          ch: 128
+          ch_mult: [ 1, 2, 4, 4 ]
+          num_res_blocks: 2
+          attn_resolutions: [ ]
+          dropout: 0.0
+        lossconfig:
+          target: torch.nn.Identity
+
+    loss_fn_config:
+      target: sgm.modules.diffusionmodules.loss.StandardDiffusionLoss
+      params:
+        loss_weighting_config:
+          target: sgm.modules.diffusionmodules.loss_weighting.EpsWeighting
+        sigma_sampler_config:
+          target: sgm.modules.diffusionmodules.sigma_sampling.DiscreteSampling
+          params:
+            num_idx: 1000
+
+            discretization_config:
+              target: sgm.modules.diffusionmodules.discretizer.LegacyDDPMDiscretization
+
+    sampler_config:
+      target: sgm.modules.diffusionmodules.sampling.EulerEDMSampler
+      params:
+        num_steps: 50
+
+        discretization_config:
+          target: sgm.modules.diffusionmodules.discretizer.LegacyDDPMDiscretization
+
+        guider_config:
+          target: sgm.modules.diffusionmodules.guiders.VanillaCFG
+          params:
+            scale: 7.5
+
+data:
+  target: sgm.data.dataset.StableDataModuleFromConfig
+  params:
+    train:
+      datapipeline:
+        urls:
+          # USER: adapt this path the root of your custom dataset
+          - DATA_PATH
+        pipeline_config:
+          shardshuffle: 10000
+          sample_shuffle: 10000 # USER: you might wanna adapt depending on your available RAM
+
+        decoders:
+          - pil
+
+        postprocessors:
+          - target: sdata.mappers.TorchVisionImageTransforms
+            params:
+              key: jpg # USER: you might wanna adapt this for your custom dataset
+              transforms:
+                - target: torchvision.transforms.Resize
+                  params:
+                    size: 256
+                    interpolation: 3
+                - target: torchvision.transforms.ToTensor
+          - target: sdata.mappers.Rescaler
+          - target: sdata.mappers.AddOriginalImageSizeAsTupleAndCropToSquare
+            # USER: you might wanna use non-default parameters due to your custom dataset
+
+      loader:
+        batch_size: 64
+        num_workers: 6
+
+lightning:
+  modelcheckpoint:
+    params:
+      every_n_train_steps: 5000
+
+  callbacks:
+    metrics_over_trainsteps_checkpoint:
+      params:
+        every_n_train_steps: 25000
+
+    image_logger:
+      target: main.ImageLogger
+      params:
+        disabled: False
+        enable_autocast: False
+        batch_frequency: 1000
+        max_images: 8
+        increase_log_steps: True
+        log_first_step: False
+        log_images_kwargs:
+          use_ema_scope: False
+          N: 8
+          n_rows: 2
+
+  trainer:
+    devices: 0,
+    benchmark: True
+    num_sanity_val_steps: 0
+    accumulate_grad_batches: 1
+    max_epochs: 1000

configs/example_training/txt2img-clipl.yaml

@@ -0,0 +1,184 @@
+model:
+  base_learning_rate: 1.0e-4
+  target: sgm.models.diffusion.DiffusionEngine
+  params:
+    scale_factor: 0.13025
+    disable_first_stage_autocast: True
+    log_keys:
+      - txt
+
+    scheduler_config:
+      target: sgm.lr_scheduler.LambdaLinearScheduler
+      params:
+        warm_up_steps: [10000]
+        cycle_lengths: [10000000000000]
+        f_start: [1.e-6]
+        f_max: [1.]
+        f_min: [1.]
+
+    denoiser_config:
+      target: sgm.modules.diffusionmodules.denoiser.DiscreteDenoiser
+      params:
+        num_idx: 1000
+
+        scaling_config:
+          target: sgm.modules.diffusionmodules.denoiser_scaling.EpsScaling
+        discretization_config:
+          target: sgm.modules.diffusionmodules.discretizer.LegacyDDPMDiscretization
+
+    network_config:
+      target: sgm.modules.diffusionmodules.openaimodel.UNetModel
+      params:
+        use_checkpoint: True
+        in_channels: 4
+        out_channels: 4
+        model_channels: 320
+        attention_resolutions: [1, 2, 4]
+        num_res_blocks: 2
+        channel_mult: [1, 2, 4, 4]
+        num_head_channels: 64
+        num_classes: sequential
+        adm_in_channels: 1792
+        num_heads: 1
+        transformer_depth: 1
+        context_dim: 768
+        spatial_transformer_attn_type: softmax-xformers
+
+    conditioner_config:
+      target: sgm.modules.GeneralConditioner
+      params:
+        emb_models:
+          - is_trainable: True
+            input_key: txt
+            ucg_rate: 0.1
+            legacy_ucg_value: ""
+            target: sgm.modules.encoders.modules.FrozenCLIPEmbedder
+            params:
+              always_return_pooled: True
+
+          - is_trainable: False
+            ucg_rate: 0.1
+            input_key: original_size_as_tuple
+            target: sgm.modules.encoders.modules.ConcatTimestepEmbedderND
+            params:
+              outdim: 256
+
+          - is_trainable: False
+            input_key: crop_coords_top_left
+            ucg_rate: 0.1
+            target: sgm.modules.encoders.modules.ConcatTimestepEmbedderND
+            params:
+              outdim: 256
+
+    first_stage_config:
+      target: sgm.models.autoencoder.AutoencoderKL
+      params:
+        ckpt_path: CKPT_PATH
+        embed_dim: 4
+        monitor: val/rec_loss
+        ddconfig:
+          attn_type: vanilla-xformers
+          double_z: true
+          z_channels: 4
+          resolution: 256
+          in_channels: 3
+          out_ch: 3
+          ch: 128
+          ch_mult: [1, 2, 4, 4]
+          num_res_blocks: 2
+          attn_resolutions: []
+          dropout: 0.0
+        lossconfig:
+          target: torch.nn.Identity
+
+    loss_fn_config:
+      target: sgm.modules.diffusionmodules.loss.StandardDiffusionLoss
+      params:
+        loss_weighting_config:
+          target: sgm.modules.diffusionmodules.loss_weighting.EpsWeighting
+        sigma_sampler_config:
+          target: sgm.modules.diffusionmodules.sigma_sampling.DiscreteSampling
+          params:
+            num_idx: 1000
+
+            discretization_config:
+              target: sgm.modules.diffusionmodules.discretizer.LegacyDDPMDiscretization
+
+    sampler_config:
+      target: sgm.modules.diffusionmodules.sampling.EulerEDMSampler
+      params:
+        num_steps: 50
+
+        discretization_config:
+          target: sgm.modules.diffusionmodules.discretizer.LegacyDDPMDiscretization
+
+        guider_config:
+          target: sgm.modules.diffusionmodules.guiders.VanillaCFG
+          params:
+            scale: 7.5
+
+data:
+  target: sgm.data.dataset.StableDataModuleFromConfig
+  params:
+    train:
+      datapipeline:
+        urls:
+          # USER: adapt this path the root of your custom dataset
+          - DATA_PATH
+        pipeline_config:
+          shardshuffle: 10000
+          sample_shuffle: 10000
+
+
+        decoders:
+          - pil
+
+        postprocessors:
+          - target: sdata.mappers.TorchVisionImageTransforms
+            params:
+              key: jpg # USER: you might wanna adapt this for your custom dataset
+              transforms:
+                - target: torchvision.transforms.Resize
+                  params:
+                    size: 256
+                    interpolation: 3
+                - target: torchvision.transforms.ToTensor
+          - target: sdata.mappers.Rescaler
+            # USER: you might wanna use non-default parameters due to your custom dataset
+          - target: sdata.mappers.AddOriginalImageSizeAsTupleAndCropToSquare
+            # USER: you might wanna use non-default parameters due to your custom dataset
+
+      loader:
+        batch_size: 64
+        num_workers: 6
+
+lightning:
+  modelcheckpoint:
+    params:
+      every_n_train_steps: 5000
+
+  callbacks:
+    metrics_over_trainsteps_checkpoint:
+      params:
+        every_n_train_steps: 25000
+
+    image_logger:
+      target: main.ImageLogger
+      params:
+        disabled: False
+        enable_autocast: False
+        batch_frequency: 1000
+        max_images: 8
+        increase_log_steps: True
+        log_first_step: False
+        log_images_kwargs:
+          use_ema_scope: False
+          N: 8
+          n_rows: 2
+
+  trainer:
+    devices: 0,
+    benchmark: True
+    num_sanity_val_steps: 0
+    accumulate_grad_batches: 1
+    max_epochs: 1000

configs/inference/sd_2_1.yaml

@@ -0,0 +1,60 @@
+model:
+  target: sgm.models.diffusion.DiffusionEngine
+  params:
+    scale_factor: 0.18215
+    disable_first_stage_autocast: True
+
+    denoiser_config:
+      target: sgm.modules.diffusionmodules.denoiser.DiscreteDenoiser
+      params:
+        num_idx: 1000
+
+        scaling_config:
+          target: sgm.modules.diffusionmodules.denoiser_scaling.EpsScaling
+        discretization_config:
+          target: sgm.modules.diffusionmodules.discretizer.LegacyDDPMDiscretization
+
+    network_config:
+      target: sgm.modules.diffusionmodules.openaimodel.UNetModel
+      params:
+        use_checkpoint: True
+        in_channels: 4
+        out_channels: 4
+        model_channels: 320
+        attention_resolutions: [4, 2, 1]
+        num_res_blocks: 2
+        channel_mult: [1, 2, 4, 4]
+        num_head_channels: 64
+        use_linear_in_transformer: True
+        transformer_depth: 1
+        context_dim: 1024
+
+    conditioner_config:
+      target: sgm.modules.GeneralConditioner
+      params:
+        emb_models:
+          - is_trainable: False
+            input_key: txt
+            target: sgm.modules.encoders.modules.FrozenOpenCLIPEmbedder
+            params:
+              freeze: true
+              layer: penultimate
+
+    first_stage_config:
+      target: sgm.models.autoencoder.AutoencoderKL
+      params:
+        embed_dim: 4
+        monitor: val/rec_loss
+        ddconfig:
+          double_z: true
+          z_channels: 4
+          resolution: 256
+          in_channels: 3
+          out_ch: 3
+          ch: 128
+          ch_mult: [1, 2, 4, 4]
+          num_res_blocks: 2
+          attn_resolutions: []
+          dropout: 0.0
+        lossconfig:
+          target: torch.nn.Identity

configs/inference/sd_2_1_768.yaml

@@ -0,0 +1,60 @@
+model:
+  target: sgm.models.diffusion.DiffusionEngine
+  params:
+    scale_factor: 0.18215
+    disable_first_stage_autocast: True
+
+    denoiser_config:
+      target: sgm.modules.diffusionmodules.denoiser.DiscreteDenoiser
+      params:
+        num_idx: 1000
+
+        scaling_config:
+          target: sgm.modules.diffusionmodules.denoiser_scaling.VScaling
+        discretization_config:
+          target: sgm.modules.diffusionmodules.discretizer.LegacyDDPMDiscretization
+
+    network_config:
+      target: sgm.modules.diffusionmodules.openaimodel.UNetModel
+      params:
+        use_checkpoint: True
+        in_channels: 4
+        out_channels: 4
+        model_channels: 320
+        attention_resolutions: [4, 2, 1]
+        num_res_blocks: 2
+        channel_mult: [1, 2, 4, 4]
+        num_head_channels: 64
+        use_linear_in_transformer: True
+        transformer_depth: 1
+        context_dim: 1024
+
+    conditioner_config:
+      target: sgm.modules.GeneralConditioner
+      params:
+        emb_models:
+          - is_trainable: False
+            input_key: txt
+            target: sgm.modules.encoders.modules.FrozenOpenCLIPEmbedder
+            params:
+              freeze: true
+              layer: penultimate
+
+    first_stage_config:
+      target: sgm.models.autoencoder.AutoencoderKL
+      params:
+        embed_dim: 4
+        monitor: val/rec_loss
+        ddconfig:
+          double_z: true
+          z_channels: 4
+          resolution: 256
+          in_channels: 3
+          out_ch: 3
+          ch: 128
+          ch_mult: [1, 2, 4, 4]
+          num_res_blocks: 2
+          attn_resolutions: []
+          dropout: 0.0
+        lossconfig:
+          target: torch.nn.Identity

configs/inference/sd_xl_base.yaml

@@ -0,0 +1,93 @@
+model:
+  target: sgm.models.diffusion.DiffusionEngine
+  params:
+    scale_factor: 0.13025
+    disable_first_stage_autocast: True
+
+    denoiser_config:
+      target: sgm.modules.diffusionmodules.denoiser.DiscreteDenoiser
+      params:
+        num_idx: 1000
+
+        scaling_config:
+          target: sgm.modules.diffusionmodules.denoiser_scaling.EpsScaling
+        discretization_config:
+          target: sgm.modules.diffusionmodules.discretizer.LegacyDDPMDiscretization
+
+    network_config:
+      target: sgm.modules.diffusionmodules.openaimodel.UNetModel
+      params:
+        adm_in_channels: 2816
+        num_classes: sequential
+        use_checkpoint: True
+        in_channels: 4
+        out_channels: 4
+        model_channels: 320
+        attention_resolutions: [4, 2]
+        num_res_blocks: 2
+        channel_mult: [1, 2, 4]
+        num_head_channels: 64
+        use_linear_in_transformer: True
+        transformer_depth: [1, 2, 10]
+        context_dim: 2048
+        spatial_transformer_attn_type: softmax-xformers
+
+    conditioner_config:
+      target: sgm.modules.GeneralConditioner
+      params:
+        emb_models:
+          - is_trainable: False
+            input_key: txt
+            target: sgm.modules.encoders.modules.FrozenCLIPEmbedder
+            params:
+              layer: hidden
+              layer_idx: 11
+
+          - is_trainable: False
+            input_key: txt
+            target: sgm.modules.encoders.modules.FrozenOpenCLIPEmbedder2
+            params:
+              arch: ViT-bigG-14
+              version: laion2b_s39b_b160k
+              freeze: True
+              layer: penultimate
+              always_return_pooled: True
+              legacy: False
+
+          - is_trainable: False
+            input_key: original_size_as_tuple
+            target: sgm.modules.encoders.modules.ConcatTimestepEmbedderND
+            params:
+              outdim: 256
+
+          - is_trainable: False
+            input_key: crop_coords_top_left
+            target: sgm.modules.encoders.modules.ConcatTimestepEmbedderND
+            params:
+              outdim: 256
+
+          - is_trainable: False
+            input_key: target_size_as_tuple
+            target: sgm.modules.encoders.modules.ConcatTimestepEmbedderND
+            params:
+              outdim: 256
+
+    first_stage_config:
+      target: sgm.models.autoencoder.AutoencoderKL
+      params:
+        embed_dim: 4
+        monitor: val/rec_loss
+        ddconfig:
+          attn_type: vanilla-xformers
+          double_z: true
+          z_channels: 4
+          resolution: 256
+          in_channels: 3
+          out_ch: 3
+          ch: 128
+          ch_mult: [1, 2, 4, 4]
+          num_res_blocks: 2
+          attn_resolutions: []
+          dropout: 0.0
+        lossconfig:
+          target: torch.nn.Identity

configs/inference/sd_xl_refiner.yaml

@@ -0,0 +1,86 @@
+model:
+  target: sgm.models.diffusion.DiffusionEngine
+  params:
+    scale_factor: 0.13025
+    disable_first_stage_autocast: True
+
+    denoiser_config:
+      target: sgm.modules.diffusionmodules.denoiser.DiscreteDenoiser
+      params:
+        num_idx: 1000
+
+        scaling_config:
+          target: sgm.modules.diffusionmodules.denoiser_scaling.EpsScaling
+        discretization_config:
+          target: sgm.modules.diffusionmodules.discretizer.LegacyDDPMDiscretization
+
+    network_config:
+      target: sgm.modules.diffusionmodules.openaimodel.UNetModel
+      params:
+        adm_in_channels: 2560
+        num_classes: sequential
+        use_checkpoint: True
+        in_channels: 4
+        out_channels: 4
+        model_channels: 384
+        attention_resolutions: [4, 2]
+        num_res_blocks: 2
+        channel_mult: [1, 2, 4, 4]
+        num_head_channels: 64
+        use_linear_in_transformer: True
+        transformer_depth: 4
+        context_dim: [1280, 1280, 1280, 1280]
+        spatial_transformer_attn_type: softmax-xformers
+
+    conditioner_config:
+      target: sgm.modules.GeneralConditioner
+      params:
+        emb_models:
+          - is_trainable: False
+            input_key: txt
+            target: sgm.modules.encoders.modules.FrozenOpenCLIPEmbedder2
+            params:
+              arch: ViT-bigG-14
+              version: laion2b_s39b_b160k
+              legacy: False
+              freeze: True
+              layer: penultimate
+              always_return_pooled: True
+
+          - is_trainable: False
+            input_key: original_size_as_tuple
+            target: sgm.modules.encoders.modules.ConcatTimestepEmbedderND
+            params:
+              outdim: 256
+
+          - is_trainable: False
+            input_key: crop_coords_top_left
+            target: sgm.modules.encoders.modules.ConcatTimestepEmbedderND
+            params:
+              outdim: 256
+
+          - is_trainable: False
+            input_key: aesthetic_score
+            target: sgm.modules.encoders.modules.ConcatTimestepEmbedderND
+            params:
+              outdim: 256
+
+    first_stage_config:
+      target: sgm.models.autoencoder.AutoencoderKL
+      params:
+        embed_dim: 4
+        monitor: val/rec_loss
+        ddconfig:
+          attn_type: vanilla-xformers
+          double_z: true
+          z_channels: 4
+          resolution: 256
+          in_channels: 3
+          out_ch: 3
+          ch: 128
+          ch_mult: [1, 2, 4, 4]
+          num_res_blocks: 2
+          attn_resolutions: []
+          dropout: 0.0
+        lossconfig:
+          target: torch.nn.Identity

configs/inference/svd.yaml

@@ -0,0 +1,131 @@
+model:
+  target: sgm.models.diffusion.DiffusionEngine
+  params:
+    scale_factor: 0.18215
+    disable_first_stage_autocast: True
+
+    denoiser_config:
+      target: sgm.modules.diffusionmodules.denoiser.Denoiser
+      params:
+        scaling_config:
+          target: sgm.modules.diffusionmodules.denoiser_scaling.VScalingWithEDMcNoise
+
+    network_config:
+      target: sgm.modules.diffusionmodules.video_model.VideoUNet
+      params:
+        adm_in_channels: 768
+        num_classes: sequential
+        use_checkpoint: True
+        in_channels: 8
+        out_channels: 4
+        model_channels: 320
+        attention_resolutions: [4, 2, 1]
+        num_res_blocks: 2
+        channel_mult: [1, 2, 4, 4]
+        num_head_channels: 64
+        use_linear_in_transformer: True
+        transformer_depth: 1
+        context_dim: 1024
+        spatial_transformer_attn_type: softmax-xformers
+        extra_ff_mix_layer: True
+        use_spatial_context: True
+        merge_strategy: learned_with_images
+        video_kernel_size: [3, 1, 1]
+
+    conditioner_config:
+      target: sgm.modules.GeneralConditioner
+      params:
+        emb_models:
+        - is_trainable: False
+          input_key: cond_frames_without_noise
+          target: sgm.modules.encoders.modules.FrozenOpenCLIPImagePredictionEmbedder
+          params:
+            n_cond_frames: 1
+            n_copies: 1
+            open_clip_embedding_config:
+              target: sgm.modules.encoders.modules.FrozenOpenCLIPImageEmbedder
+              params:
+                freeze: True
+
+        - input_key: fps_id
+          is_trainable: False
+          target: sgm.modules.encoders.modules.ConcatTimestepEmbedderND
+          params:
+            outdim: 256
+
+        - input_key: motion_bucket_id
+          is_trainable: False
+          target: sgm.modules.encoders.modules.ConcatTimestepEmbedderND
+          params:
+            outdim: 256
+
+        - input_key: cond_frames
+          is_trainable: False
+          target: sgm.modules.encoders.modules.VideoPredictionEmbedderWithEncoder
+          params:
+            disable_encoder_autocast: True
+            n_cond_frames: 1
+            n_copies: 1
+            is_ae: True
+            encoder_config:
+              target: sgm.models.autoencoder.AutoencoderKLModeOnly
+              params:
+                embed_dim: 4
+                monitor: val/rec_loss
+                ddconfig:
+                  attn_type: vanilla-xformers
+                  double_z: True
+                  z_channels: 4
+                  resolution: 256
+                  in_channels: 3
+                  out_ch: 3
+                  ch: 128
+                  ch_mult: [1, 2, 4, 4]
+                  num_res_blocks: 2
+                  attn_resolutions: []
+                  dropout: 0.0
+                lossconfig:
+                  target: torch.nn.Identity
+
+        - input_key: cond_aug
+          is_trainable: False
+          target: sgm.modules.encoders.modules.ConcatTimestepEmbedderND
+          params:
+            outdim: 256
+
+    first_stage_config:
+      target: sgm.models.autoencoder.AutoencodingEngine
+      params:
+        loss_config:
+          target: torch.nn.Identity
+        regularizer_config:
+          target: sgm.modules.autoencoding.regularizers.DiagonalGaussianRegularizer
+        encoder_config: 
+          target: sgm.modules.diffusionmodules.model.Encoder
+          params:
+            attn_type: vanilla
+            double_z: True
+            z_channels: 4
+            resolution: 256
+            in_channels: 3
+            out_ch: 3
+            ch: 128
+            ch_mult: [1, 2, 4, 4]
+            num_res_blocks: 2
+            attn_resolutions: []
+            dropout: 0.0
+        decoder_config:
+          target: sgm.modules.autoencoding.temporal_ae.VideoDecoder
+          params:
+            attn_type: vanilla
+            double_z: True
+            z_channels: 4
+            resolution: 256
+            in_channels: 3
+            out_ch: 3
+            ch: 128
+            ch_mult: [1, 2, 4, 4]
+            num_res_blocks: 2
+            attn_resolutions: []
+            dropout: 0.0
+            video_kernel_size: [3, 1, 1]

configs/inference/svd_image_decoder.yaml

@@ -0,0 +1,114 @@
+model:
+  target: sgm.models.diffusion.DiffusionEngine
+  params:
+    scale_factor: 0.18215
+    disable_first_stage_autocast: True
+
+    denoiser_config:
+      target: sgm.modules.diffusionmodules.denoiser.Denoiser
+      params:
+        scaling_config:
+          target: sgm.modules.diffusionmodules.denoiser_scaling.VScalingWithEDMcNoise
+
+    network_config:
+      target: sgm.modules.diffusionmodules.video_model.VideoUNet
+      params:
+        adm_in_channels: 768
+        num_classes: sequential
+        use_checkpoint: True
+        in_channels: 8
+        out_channels: 4
+        model_channels: 320
+        attention_resolutions: [4, 2, 1]
+        num_res_blocks: 2
+        channel_mult: [1, 2, 4, 4]
+        num_head_channels: 64
+        use_linear_in_transformer: True
+        transformer_depth: 1
+        context_dim: 1024
+        spatial_transformer_attn_type: softmax-xformers
+        extra_ff_mix_layer: True
+        use_spatial_context: True
+        merge_strategy: learned_with_images
+        video_kernel_size: [3, 1, 1]
+
+    conditioner_config:
+      target: sgm.modules.GeneralConditioner
+      params:
+        emb_models:
+        - is_trainable: False
+          input_key: cond_frames_without_noise
+          target: sgm.modules.encoders.modules.FrozenOpenCLIPImagePredictionEmbedder
+          params:
+            n_cond_frames: 1
+            n_copies: 1
+            open_clip_embedding_config:
+              target: sgm.modules.encoders.modules.FrozenOpenCLIPImageEmbedder
+              params:
+                freeze: True
+
+        - input_key: fps_id
+          is_trainable: False
+          target: sgm.modules.encoders.modules.ConcatTimestepEmbedderND
+          params:
+            outdim: 256
+
+        - input_key: motion_bucket_id
+          is_trainable: False
+          target: sgm.modules.encoders.modules.ConcatTimestepEmbedderND
+          params:
+            outdim: 256
+
+        - input_key: cond_frames
+          is_trainable: False
+          target: sgm.modules.encoders.modules.VideoPredictionEmbedderWithEncoder
+          params:
+            disable_encoder_autocast: True
+            n_cond_frames: 1
+            n_copies: 1
+            is_ae: True
+            encoder_config:
+              target: sgm.models.autoencoder.AutoencoderKLModeOnly
+              params:
+                embed_dim: 4
+                monitor: val/rec_loss
+                ddconfig:
+                  attn_type: vanilla-xformers
+                  double_z: True
+                  z_channels: 4
+                  resolution: 256
+                  in_channels: 3
+                  out_ch: 3
+                  ch: 128
+                  ch_mult: [1, 2, 4, 4]
+                  num_res_blocks: 2
+                  attn_resolutions: []
+                  dropout: 0.0
+                lossconfig:
+                  target: torch.nn.Identity
+
+        - input_key: cond_aug
+          is_trainable: False
+          target: sgm.modules.encoders.modules.ConcatTimestepEmbedderND
+          params:
+            outdim: 256
+
+    first_stage_config:
+      target: sgm.models.autoencoder.AutoencoderKL
+      params:
+        embed_dim: 4
+        monitor: val/rec_loss
+        ddconfig:
+          attn_type: vanilla-xformers
+          double_z: True
+          z_channels: 4
+          resolution: 256
+          in_channels: 3
+          out_ch: 3
+          ch: 128
+          ch_mult: [1, 2, 4, 4]
+          num_res_blocks: 2
+          attn_resolutions: []
+          dropout: 0.0
+        lossconfig:
+          target: torch.nn.Identity

configs/inference/svd_mv.yaml

@@ -0,0 +1,202 @@
+model:
+  base_learning_rate: 1.0e-05
+  target: sgm.models.video_diffusion.DiffusionEngine
+  params:
+    ckpt_path: ckpts/svd_xt.safetensors
+    scale_factor: 0.18215
+    disable_first_stage_autocast: true
+    scheduler_config:
+      target: sgm.lr_scheduler.LambdaLinearScheduler
+      params:
+        warm_up_steps:
+        - 1
+        cycle_lengths:
+        - 10000000000000
+        f_start:
+        - 1.0e-06
+        f_max:
+        - 1.0
+        f_min:
+        - 1.0
+    denoiser_config:
+      target: sgm.modules.diffusionmodules.denoiser.Denoiser
+      params:
+        scaling_config:
+          target: sgm.modules.diffusionmodules.denoiser_scaling.VScalingWithEDMcNoise
+    network_config:
+      target: sgm.modules.diffusionmodules.video_model.VideoUNet
+      params:
+        adm_in_channels: 768
+        num_classes: sequential
+        use_checkpoint: true
+        in_channels: 8
+        out_channels: 4
+        model_channels: 320
+        attention_resolutions:
+        - 4
+        - 2
+        - 1
+        num_res_blocks: 2
+        channel_mult:
+        - 1
+        - 2
+        - 4
+        - 4
+        num_head_channels: 64
+        use_linear_in_transformer: true
+        transformer_depth: 1
+        context_dim: 1024
+        spatial_transformer_attn_type: softmax-xformers
+        extra_ff_mix_layer: true
+        use_spatial_context: true
+        merge_strategy: learned_with_images
+        video_kernel_size:
+        - 3
+        - 1
+        - 1
+    conditioner_config:
+      target: sgm.modules.GeneralConditioner
+      params:
+        emb_models:
+        - is_trainable: false
+          ucg_rate: 0.2
+          input_key: cond_frames_without_noise
+          target: sgm.modules.encoders.modules.FrozenOpenCLIPImagePredictionEmbedder
+          params:
+            n_cond_frames: 1
+            n_copies: 1
+            open_clip_embedding_config:
+              target: sgm.modules.encoders.modules.FrozenOpenCLIPImageEmbedder
+              params:
+                freeze: true
+        - input_key: fps_id
+          is_trainable: true
+          target: sgm.modules.encoders.modules.ConcatTimestepEmbedderND
+          params:
+            outdim: 256
+        - input_key: motion_bucket_id
+          is_trainable: true
+          target: sgm.modules.encoders.modules.ConcatTimestepEmbedderND
+          params:
+            outdim: 256
+        - input_key: cond_frames
+          is_trainable: false
+          ucg_rate: 0.2
+          target: sgm.modules.encoders.modules.VideoPredictionEmbedderWithEncoder
+          params:
+            disable_encoder_autocast: true
+            n_cond_frames: 1
+            n_copies: 1
+            is_ae: true
+            encoder_config:
+              target: sgm.models.autoencoder.AutoencoderKLModeOnly
+              params:
+                embed_dim: 4
+                monitor: val/rec_loss
+                ddconfig:
+                  attn_type: vanilla-xformers
+                  double_z: true
+                  z_channels: 4
+                  resolution: 256
+                  in_channels: 3
+                  out_ch: 3
+                  ch: 128
+                  ch_mult:
+                  - 1
+                  - 2
+                  - 4
+                  - 4
+                  num_res_blocks: 2
+                  attn_resolutions: []
+                  dropout: 0.0
+                lossconfig:
+                  target: torch.nn.Identity
+        - input_key: cond_aug
+          is_trainable: true
+          target: sgm.modules.encoders.modules.ConcatTimestepEmbedderND
+          params:
+            outdim: 256
+    first_stage_config:
+      target: sgm.models.autoencoder.AutoencodingEngine
+      params:
+        loss_config:
+          target: torch.nn.Identity
+        regularizer_config:
+          target: sgm.modules.autoencoding.regularizers.DiagonalGaussianRegularizer
+        encoder_config:
+          target: sgm.modules.diffusionmodules.model.Encoder
+          params:
+            attn_type: vanilla
+            double_z: true
+            z_channels: 4
+            resolution: 256
+            in_channels: 3
+            out_ch: 3
+            ch: 128
+            ch_mult:
+            - 1
+            - 2
+            - 4
+            - 4
+            num_res_blocks: 2
+            attn_resolutions: []
+            dropout: 0.0
+        decoder_config:
+          target: sgm.modules.autoencoding.temporal_ae.VideoDecoder
+          params:
+            attn_type: vanilla
+            double_z: true
+            z_channels: 4
+            resolution: 256
+            in_channels: 3
+            out_ch: 3
+            ch: 128
+            ch_mult:
+            - 1
+            - 2
+            - 4
+            - 4
+            num_res_blocks: 2
+            attn_resolutions: []
+            dropout: 0.0
+            video_kernel_size:
+            - 3
+            - 1
+            - 1
+    sampler_config:
+      target: sgm.modules.diffusionmodules.sampling.EulerEDMSampler
+      params:
+        num_steps: 30
+        discretization_config:
+          target: sgm.modules.diffusionmodules.discretizer.EDMDiscretization
+          params:
+            sigma_max: 700.0
+        guider_config:
+          target: sgm.modules.diffusionmodules.guiders.LinearPredictionGuider
+          params:
+            max_scale: 2.5
+            min_scale: 1.0
+            num_frames: 24
+    loss_fn_config:
+      target: sgm.modules.diffusionmodules.loss.StandardDiffusionLoss
+      params:
+        batch2model_keys:
+        - num_video_frames
+        - image_only_indicator
+        loss_weighting_config:
+          target: sgm.modules.diffusionmodules.loss_weighting.EDMWeighting
+          params:
+            sigma_data: 1.0
+        sigma_sampler_config:
+          target: sgm.modules.diffusionmodules.sigma_sampling.EDMSampling
+          params:
+            p_mean: 0.3
+            p_std: 1.2
+data:
+  target: sgm.data.objaverse.ObjaverseSpiralDataset
+  params:
+    root_dir: /mnt/mfs/zilong.chen/Downloads/objaverse-ndd-samples
+    random_front: true
+    batch_size: 2
+    num_workers: 16
+    cond_aug_mean: -0.0