Merge pull request #14 from invoke-ai/ryan/sdxl-lora-2

Add SDXL fine-tuning with LoRA

configs/finetune_lora_sdxl_pokemon_example.yaml

@@ -0,0 +1,30 @@
+# This is a sample config for finetuning a SDXL 1.0 model with LoRA to produce a Pokemon LoRA model.
+
+output:
+  base_output_dir: output/
+
+optimizer:
+  learning_rate: 5.0e-4
+
+dataset:
+  dataset_name: lambdalabs/pokemon-blip-captions
+
+# General
+model: stabilityai/stable-diffusion-xl-base-1.0
+vae_model: madebyollin/sdxl-vae-fp16-fix
+seed: 1
+train_text_encoder: False
+gradient_accumulation_steps: 4
+mixed_precision: fp16
+xformers: True
+gradient_checkpointing: True
+max_train_steps: 4000
+save_every_n_epochs: 1
+save_every_n_steps: null
+max_checkpoints: 100
+validation_prompts:
+  - yoda
+  - astronaut
+  - yoda in a space suit
+validate_every_n_epochs: 1
+train_batch_size: 1

src/invoke_training/lora/injection/stable_diffusion_v1.py → src/invoke_training/lora/injection/stable_diffusion.py

@@ -12,10 +12,10 @@
 from invoke_training.lora.layers import LoRAConv2dLayer, LoRALinearLayer
 
 
-def inject_lora_into_unet_sd1(
+def inject_lora_into_unet(
     unet: UNet2DConditionModel, include_non_attention_blocks: bool = False
 ) -> LoRALayerCollection:
-    """Inject LoRA layers into a Stable Diffusion v1 UNet model.
+    """Inject LoRA layers into a Stable Diffusion UNet model.
 
     Args:
         unet (UNet2DConditionModel): The UNet model to inject LoRA layers into.
@@ -45,7 +45,7 @@ def inject_lora_into_unet_sd1(
     return lora_layers
 
 
-def inject_lora_into_clip_text_encoder(text_encoder: CLIPTextModel):
+def inject_lora_into_clip_text_encoder(text_encoder: CLIPTextModel, prefix: str = "lora_te"):
     lora_layers = inject_lora_layers(
         module=text_encoder,
         lora_map={
@@ -54,17 +54,17 @@ def inject_lora_into_clip_text_encoder(text_encoder: CLIPTextModel):
         },
         include_descendants_of={CLIPAttention, CLIPMLP},
         exclude_descendants_of=None,
-        prefix="lora_te",
+        prefix=prefix,
         dtype=torch.float32,
     )
 
     return lora_layers
 
 
-def convert_lora_state_dict_to_kohya_format_sd1(
+def convert_lora_state_dict_to_kohya_format(
     state_dict: typing.Dict[str, torch.Tensor]
 ) -> typing.Dict[str, torch.Tensor]:
-    """Convert a Stable Diffusion v1 LoRA state_dict from internal invoke-training format to kohya_ss format.
+    """Convert a Stable Diffusion LoRA state_dict from internal invoke-training format to kohya_ss format.
 
     Args:
         state_dict (typing.Dict[str, torch.Tensor]): LoRA layer state_dict in invoke-training format.

src/invoke_training/scripts/invoke_finetune_lora_sdxl.py

@@ -4,7 +4,7 @@
 import yaml
 
 from invoke_training.training.finetune_lora.finetune_lora_config import (
-    FinetuneLoRAConfig,
+    FinetuneLoRASDXLConfig,
 )
 from invoke_training.training.finetune_lora.finetune_lora_sdxl import run_training
 
@@ -27,7 +27,7 @@ def main():
     with open(args.cfg_file, "r") as f:
         cfg = yaml.safe_load(f)
 
-    train_config = FinetuneLoRAConfig(**cfg)
+    train_config = FinetuneLoRASDXLConfig(**cfg)
 
     run_training(train_config)
 

src/invoke_training/training/finetune_lora/finetune_lora_config.py

@@ -167,3 +167,10 @@ class FinetuneLoRAConfig(BaseModel):
 
     # The training batch size.
     train_batch_size: int = 4
+
+
+class FinetuneLoRASDXLConfig(FinetuneLoRAConfig):
+    # The name of the Hugging Face Hub VAE model to train against. This will override the VAE bundled with the base
+    # model (specified by the `model` parameter). This config option is provided for SDXL models, because SDXL shipped
+    # with a VAE that produces NaNs in fp16 mode, so it is common to replace this VAE with a fixed version.
+    vae_model: typing.Optional[str] = None

src/invoke_training/training/finetune_lora/finetune_lora_sd.py

@@ -18,10 +18,10 @@
 from tqdm.auto import tqdm
 from transformers import CLIPTextModel, CLIPTokenizer
 
-from invoke_training.lora.injection.stable_diffusion_v1 import (
-    convert_lora_state_dict_to_kohya_format_sd1,
+from invoke_training.lora.injection.stable_diffusion import (
+    convert_lora_state_dict_to_kohya_format,
     inject_lora_into_clip_text_encoder,
-    inject_lora_into_unet_sd1,
+    inject_lora_into_unet,
 )
 from invoke_training.training.finetune_lora.finetune_lora_config import (
     FinetuneLoRAConfig,
@@ -36,8 +36,8 @@
     check_base_model_version,
 )
 from invoke_training.training.shared.checkpoint_tracker import CheckpointTracker
-from invoke_training.training.shared.datasets.image_caption_dataloader import (
-    build_image_caption_dataloader,
+from invoke_training.training.shared.datasets.image_caption_sd_dataloader import (
+    build_image_caption_sd_dataloader,
 )
 from invoke_training.training.shared.serialization import save_state_dict
 
@@ -122,7 +122,7 @@ def _save_checkpoint(
     state_dict = {}
     for model_lora_layers in lora_layers.values():
         model_state_dict = model_lora_layers.get_lora_state_dict()
-        model_kohya_state_dict = convert_lora_state_dict_to_kohya_format_sd1(model_state_dict)
+        model_kohya_state_dict = convert_lora_state_dict_to_kohya_format(model_state_dict)
         state_dict.update(model_kohya_state_dict)
 
     save_state_dict(state_dict, save_path)
@@ -189,6 +189,8 @@ def _generate_validation_images(
                             prompt,
                             num_inference_steps=30,
                             generator=generator,
+                            height=config.dataset.resolution,
+                            width=config.dataset.resolution,
                         ).images[0]
                     )
 
@@ -312,7 +314,7 @@ def run_training(config: FinetuneLoRAConfig):  # noqa: C901
 
     lora_layers = torch.nn.ModuleDict()
     if config.train_unet:
-        lora_layers["unet"] = inject_lora_into_unet_sd1(unet, config.train_unet_non_attention_blocks)
+        lora_layers["unet"] = inject_lora_into_unet(unet, config.train_unet_non_attention_blocks)
     if config.train_text_encoder:
         lora_layers["text_encoder"] = inject_lora_into_clip_text_encoder(text_encoder)
 
@@ -324,7 +326,7 @@ def run_training(config: FinetuneLoRAConfig):  # noqa: C901
 
     optimizer = _initialize_optimizer(config, lora_layers.parameters())
 
-    data_loader = build_image_caption_dataloader(config.dataset, tokenizer, config.train_batch_size)
+    data_loader = build_image_caption_sd_dataloader(config.dataset, tokenizer, config.train_batch_size)
 
     # TODO(ryand): Test in a distributed training environment and more clearly document the rationale for scaling steps
     # by the number of processes. This scaling logic was copied from the diffusers example training code, but it appears