Merge pull request #14170 from MrCheeze/sd-turbo

Add support for SD 2.1 Turbo

modules/sd_hijack.py

@@ -38,9 +38,6 @@ ldm.models.diffusion.ddpm.print = shared.ldm_print
 optimizers = []
 current_optimizer: sd_hijack_optimizations.SdOptimization = None
 
-ldm_original_forward = patches.patch(__file__, ldm.modules.diffusionmodules.openaimodel.UNetModel, "forward", sd_unet.UNetModel_forward)
-sgm_original_forward = patches.patch(__file__, sgm.modules.diffusionmodules.openaimodel.UNetModel, "forward", sd_unet.UNetModel_forward)
-
 def list_optimizers():
     new_optimizers = script_callbacks.list_optimizers_callback()
 
@@ -258,6 +255,9 @@ class StableDiffusionModelHijack:
 
         import modules.models.diffusion.ddpm_edit
 
+        ldm_original_forward = patches.patch(__file__, ldm.modules.diffusionmodules.openaimodel.UNetModel, "forward", sd_unet.UNetModel_forward)
+        sgm_original_forward = patches.patch(__file__, sgm.modules.diffusionmodules.openaimodel.UNetModel, "forward", sd_unet.UNetModel_forward)
+
         if isinstance(m, ldm.models.diffusion.ddpm.LatentDiffusion):
             sd_unet.original_forward = ldm_original_forward
         elif isinstance(m, modules.models.diffusion.ddpm_edit.LatentDiffusion):
@@ -303,6 +303,9 @@ class StableDiffusionModelHijack:
         self.layers = None
         self.clip = None
 
+        patches.undo(__file__, ldm.modules.diffusionmodules.openaimodel.UNetModel, "forward")
+        patches.undo(__file__, sgm.modules.diffusionmodules.openaimodel.UNetModel, "forward")
+
         sd_unet.original_forward = None
 
 

modules/sd_models.py

@@ -230,15 +230,19 @@ def select_checkpoint():
     return checkpoint_info
 
 
-checkpoint_dict_replacements = {
+checkpoint_dict_replacements_sd1 = {
     'cond_stage_model.transformer.embeddings.': 'cond_stage_model.transformer.text_model.embeddings.',
     'cond_stage_model.transformer.encoder.': 'cond_stage_model.transformer.text_model.encoder.',
     'cond_stage_model.transformer.final_layer_norm.': 'cond_stage_model.transformer.text_model.final_layer_norm.',
 }
 
+checkpoint_dict_replacements_sd2_turbo = { # Converts SD 2.1 Turbo from SGM to LDM format.
+    'conditioner.embedders.0.': 'cond_stage_model.',
+}
+
 
-def transform_checkpoint_dict_key(k):
-    for text, replacement in checkpoint_dict_replacements.items():
+def transform_checkpoint_dict_key(k, replacements):
+    for text, replacement in replacements.items():
         if k.startswith(text):
             k = replacement + k[len(text):]
 
@@ -249,9 +253,14 @@ def get_state_dict_from_checkpoint(pl_sd):
     pl_sd = pl_sd.pop("state_dict", pl_sd)
     pl_sd.pop("state_dict", None)
 
+    is_sd2_turbo = 'conditioner.embedders.0.model.ln_final.weight' in pl_sd and pl_sd['conditioner.embedders.0.model.ln_final.weight'].size()[0] == 1024
+
     sd = {}
     for k, v in pl_sd.items():
-        new_key = transform_checkpoint_dict_key(k)
+        if is_sd2_turbo:
+            new_key = transform_checkpoint_dict_key(k, checkpoint_dict_replacements_sd2_turbo)
+        else:
+            new_key = transform_checkpoint_dict_key(k, checkpoint_dict_replacements_sd1)
 
         if new_key is not None:
             sd[new_key] = v