file structure

.gitmodules

@@ -0,0 +1,6 @@
+[submodule "repos/diffusers"]
+	path = repos/diffusers
+	url = https://github.com/CCRcmcpe/diffusers
+[submodule "repos/xformers"]
+	path = repos/xformers
+	url = https://github.com/facebookresearch/xformers

back.ps1

@@ -4,7 +4,7 @@ param(
     [string]$step
  )
 
-$BACK_CONVERTER = "repos/diffusers/scripts/convert_diffusers_to_original_stable_diffusion.py"
+$BackConverter = "repos/diffusers/scripts/convert_diffusers_to_original_stable_diffusion.py"
 
 $AutoDLTmp = "/root/autodl-tmp"
 $OutPath = "$AutoDLTmp/output" # no slash at the end
@@ -14,7 +14,7 @@ $ModelPath = Join-Path $OutPath $step
 $CheckpointPath = Join-Path $ModelPath "model.ckpt"
 
 # unet_half could reduce the size of the model
-python $BACK_CONVERTER  --model_path $ModelPath `
+python $BackConverter  --model_path $ModelPath `
                         --checkpoint_path $CheckpointPath `
                         --unet_half
 

convert.ps1

@@ -1,17 +1,10 @@
-$TRAINER = "repos/diffusers/examples/dreambooth/train_dreambooth.py"
-$CONVERTER = "repos/diffusers/scripts/convert_original_stable_diffusion_to_diffusers.py"
-$BACK_CONVERTER = "repos/diffusers/scripts/convert_diffusers_to_original_stable_diffusion.py"
+$Converter = "repos/diffusers/scripts/convert_original_stable_diffusion_to_diffusers.py"
 
 # Download the model
 # https://pub-2fdef7a2969f43289c42ac5ae3412fd4.r2.dev/animefull-pruned.tar
 # https://pub-2fdef7a2969f43289c42ac5ae3412fd4.r2.dev/animevae.pt
 # tar -cf animefull-pruned.tar
 
-# models/
-# |-- animevae.pt
-# |-- config.yaml // the config file comes with the model
-# `-- model.ckpt
-
 $ModelDir = "models"
 $CheckpointPath = Join-Path $ModelDir "model.ckpt"
 $VaePath = Join-Path $ModelDir "animevae.pt"
@@ -21,7 +14,7 @@ $ModelName = "animefull-pruned"
 $AutoDLTmp = "/root/autodl-tmp"
 $DumpPath = Join-Path $AutoDLTmp $ModelName
 
-python $CONVERTER --checkpoint_path $CheckpointPath `
+python $Converter --checkpoint_path $CheckpointPath `
                   --original_config_file $ConfPath `
                   --vae_path $VaePath `
                   --dump_path $DumpPath `

train.ps1

@@ -1,17 +1,4 @@
-$TRAINER = "repos/diffusers/examples/dreambooth/train_dreambooth.py"
-$CONVERTER = "repos/diffusers/scripts/convert_original_stable_diffusion_to_diffusers.py"
-$BACK_CONVERTER = "repos/diffusers/scripts/convert_diffusers_to_original_stable_diffusion.py"
-
-# Download the model
-# https://pub-2fdef7a2969f43289c42ac5ae3412fd4.r2.dev/animefull-pruned.tar
-# https://pub-2fdef7a2969f43289c42ac5ae3412fd4.r2.dev/animevae.pt
-
-# models/
-# |-- animevae.pt
-# |-- config.yaml // the config file comes with the model
-# `-- model.ckpt
-# See `convert.ps1` for more details
-
+$Trainer = "repos/diffusers/examples/dreambooth/train_dreambooth.py"
 $AutoDLTmp = "/root/autodl-tmp"
 
 # Prompt describing the subject, like subject's name.
@@ -28,25 +15,29 @@ $TrainBatchSize = 1
 # Regularization set
 # A more general prompt describing the subject, for generating regularization
 # image set.
-$ClassPrompt = "masterpiece, best quality, 1girl"
+$ClassPrompt = "1girl"
 $ClassNegativePrompt = "lowres, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry" 
 # Regularization set path.
 $ClassDir = "${AutoDLTmp}/content/class-images"
 
 # Previewing
 # Prompt for saving samples.
-$SaveSamplePrompt = "masterpiece, best quality, sks 1girl, looking at viewer" 
+$SaveSamplePrompt = "sks 1girl standing looking at viewer, cowboy shot" 
 $SaveSampleNegative = "lowres, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry" 
 
+
+# TODO: detect if the model exists
+# Remind the user to excecute `convert.ps1` if the model is not existing.
+
 # Model path
 $ModelName = "animefull-pruned"
 # I will dump to auto-tmp to save space of system disk
-$ModelPath = "${AutoDLTmp}/${ModelName}"
-$VaePath = "${ModelPath}/vae"
-$OutPath = "${AutoDLTmp}/output"
+$ModelPath = Join-Path $AutoDLTmp $ModelName
+$VaePath = Join-Path $ModelPath "vae"
+$OutPath = Join-Path $AutoDLTmp "output"
 mkdir -p $OutPath
 
-accelerate launch $TRAINER `
+accelerate launch $Trainer `
   --instance_data_dir $InstanceDir `
   --instance_prompt $InstancePrompt `
   --pretrained_model_name_or_path $ModelPath `
@@ -75,6 +66,6 @@ accelerate launch $TRAINER `
   --save_unet_half `
   --mixed_precision="fp16"
 
-# TODO: write a script to inference images with parameters
+# TODO: write a script to inference images with parameters like json?
 # don't care about inference. I would do it some where else.
 # see `back.ps1`