diff --git a/pipelines/model_cogview.py b/pipelines/model_cogview.py
index 673c41565..146177063 100644
--- a/pipelines/model_cogview.py
+++ b/pipelines/model_cogview.py
@@ -3,6 +3,11 @@ import diffusers
 from modules import shared, devices, sd_models, model_quant, sd_hijack_te
 from modules.logger import log
 from pipelines import generic
+from pipelines.native_transformer import TransformerSpec
+
+
+COGVIEW3_SPEC = TransformerSpec(cls=diffusers.CogView3PlusTransformer2DModel)
+COGVIEW4_SPEC = TransformerSpec(cls=diffusers.CogView4Transformer2DModel)
 
 
 def load_cogview3(checkpoint_info, diffusers_load_config=None):
@@ -14,7 +19,7 @@ def load_cogview3(checkpoint_info, diffusers_load_config=None):
     load_args, _quant_args = model_quant.get_dit_args(diffusers_load_config)
     log.debug(f'Load model: type=CogView3 repo="{repo_id}" config={diffusers_load_config} offload={shared.opts.diffusers_offload_mode} dtype={devices.dtype} args={load_args}')
 
-    transformer = generic.load_transformer(repo_id, cls_name=diffusers.CogView3PlusTransformer2DModel, load_config=diffusers_load_config, subfolder="transformer")
+    transformer = generic.load_transformer(repo_id, cls_name=diffusers.CogView3PlusTransformer2DModel, load_config=diffusers_load_config, subfolder="transformer", native_spec=COGVIEW3_SPEC)
     text_encoder = generic.load_text_encoder(repo_id, cls_name=transformers.T5EncoderModel, load_config=diffusers_load_config, subfolder="text_encoder")
 
     pipe = diffusers.CogView3PlusPipeline.from_pretrained(
@@ -40,7 +45,7 @@ def load_cogview4(checkpoint_info, diffusers_load_config=None):
     load_args, _quant_args = model_quant.get_dit_args(diffusers_load_config)
     log.debug(f'Load model: type=CogView4 repo="{repo_id}" config={diffusers_load_config} offload={shared.opts.diffusers_offload_mode} dtype={devices.dtype} args={load_args}')
 
-    transformer = generic.load_transformer(repo_id, cls_name=diffusers.CogView4Transformer2DModel, load_config=diffusers_load_config, subfolder="transformer")
+    transformer = generic.load_transformer(repo_id, cls_name=diffusers.CogView4Transformer2DModel, load_config=diffusers_load_config, subfolder="transformer", native_spec=COGVIEW4_SPEC)
     text_encoder = generic.load_text_encoder(repo_id, cls_name=transformers.GlmModel, load_config=diffusers_load_config, subfolder="text_encoder", allow_quant=True)
 
     pipe = diffusers.CogView4Pipeline.from_pretrained(
diff --git a/pipelines/model_glm.py b/pipelines/model_glm.py
index c58358f88..97f218bce 100644
--- a/pipelines/model_glm.py
+++ b/pipelines/model_glm.py
@@ -5,6 +5,10 @@ import diffusers
 from modules import shared, devices, sd_models, model_quant, sd_hijack_te
 from modules.logger import log, console
 from pipelines import generic
+from pipelines.native_transformer import TransformerSpec
+
+
+GLM_IMAGE_SPEC = TransformerSpec(cls=diffusers.GlmImageTransformer2DModel)
 
 
 class GLMTokenProgressProcessor(transformers.LogitsProcessor):
@@ -98,7 +102,8 @@ def load_glm_image(checkpoint_info, diffusers_load_config=None):
     transformer = generic.load_transformer(
         repo_id,
         cls_name=diffusers.GlmImageTransformer2DModel,
-        load_config=diffusers_load_config
+        load_config=diffusers_load_config,
+        native_spec=GLM_IMAGE_SPEC,
     )
 
     # Load text encoder (ByT5 for glyph) - cannot use shared T5 as GLM-Image requires specific ByT5 encoder (1472 hidden size)
diff --git a/pipelines/model_hunyuandit.py b/pipelines/model_hunyuandit.py
index b2933a94f..71e75bf48 100644
--- a/pipelines/model_hunyuandit.py
+++ b/pipelines/model_hunyuandit.py
@@ -3,6 +3,10 @@ import diffusers
 from modules import shared, sd_models, devices, model_quant
 from modules.logger import log
 from pipelines import generic
+from pipelines.native_transformer import TransformerSpec
+
+
+HUNYUANDIT_SPEC = TransformerSpec(cls=diffusers.HunyuanDiT2DModel)
 
 
 def load_hunyuandit(checkpoint_info, diffusers_load_config=None):
@@ -19,7 +23,7 @@ def load_hunyuandit(checkpoint_info, diffusers_load_config=None):
     load_args, _quant_args = model_quant.get_dit_args(diffusers_load_config)
     log.debug(f'Load model: type=HunyuanDiT repo="{repo_id}" config={diffusers_load_config} offload={shared.opts.diffusers_offload_mode} dtype={devices.dtype} args={load_args}')
 
-    transformer = generic.load_transformer(repo_id, cls_name=diffusers.HunyuanDiT2DModel, load_config=diffusers_load_config)
+    transformer = generic.load_transformer(repo_id, cls_name=diffusers.HunyuanDiT2DModel, load_config=diffusers_load_config, native_spec=HUNYUANDIT_SPEC)
     repo_te = 'Tencent-Hunyuan/HunyuanDiT-v1.2-Diffusers' if 'HunyuanDiT-v1' in repo_id else repo_id
     text_encoder_2 = generic.load_text_encoder(repo_te, cls_name=transformers.T5EncoderModel, load_config=diffusers_load_config, subfolder="text_encoder_2", allow_shared=False) # this is not normal t5
 
diff --git a/pipelines/model_joy.py b/pipelines/model_joy.py
index 5ed363315..99b8de829 100644
--- a/pipelines/model_joy.py
+++ b/pipelines/model_joy.py
@@ -3,6 +3,10 @@ import transformers
 from modules import shared, devices, sd_models, model_quant, sd_hijack_te, sd_hijack_vae
 from modules.logger import log
 from pipelines import generic
+from pipelines.native_transformer import TransformerSpec
+
+
+JOY_SPEC = TransformerSpec(cls=diffusers.JoyImageEditTransformer3DModel)
 
 
 def load_joy(checkpoint_info, diffusers_load_config=None):
@@ -18,6 +22,7 @@ def load_joy(checkpoint_info, diffusers_load_config=None):
         repo_id,
         cls_name=diffusers.JoyImageEditTransformer3DModel,
         load_config=diffusers_load_config,
+        native_spec=JOY_SPEC,
     )
     text_encoder = generic.load_text_encoder(
         repo_id,
diff --git a/pipelines/model_kandinsky.py b/pipelines/model_kandinsky.py
index 4f06e185d..41a1bb85c 100644
--- a/pipelines/model_kandinsky.py
+++ b/pipelines/model_kandinsky.py
@@ -6,6 +6,7 @@ from pipelines import generic
 from pipelines.native_transformer import TransformerSpec
 
 
+KANDINSKY3_UNET_SPEC = TransformerSpec(cls=diffusers.Kandinsky3UNet, subfolder='unet')
 KANDINSKY5_SPEC = TransformerSpec(cls=diffusers.Kandinsky5Transformer3DModel)
 
 
@@ -54,7 +55,7 @@ def load_kandinsky3(checkpoint_info, diffusers_load_config=None):
     load_args, _quant_args = model_quant.get_dit_args(diffusers_load_config)
     log.debug(f'Load model: type=Kandinsky30 repo="{repo_id}" config={diffusers_load_config} offload={shared.opts.diffusers_offload_mode} dtype={devices.dtype} args={load_args}')
 
-    unet = generic.load_transformer(repo_id, cls_name=diffusers.Kandinsky3UNet, load_config=diffusers_load_config, subfolder="unet", variant="fp16")
+    unet = generic.load_transformer(repo_id, cls_name=diffusers.Kandinsky3UNet, load_config=diffusers_load_config, subfolder="unet", variant="fp16", native_spec=KANDINSKY3_UNET_SPEC)
     text_encoder = generic.load_text_encoder(repo_id, cls_name=transformers.T5EncoderModel, load_config=diffusers_load_config, subfolder="text_encoder", variant="fp16", allow_shared=False)
 
     pipe = diffusers.Kandinsky3Pipeline.from_pretrained(
diff --git a/pipelines/model_pixart.py b/pipelines/model_pixart.py
index 1ced8b659..0a79b3178 100644
--- a/pipelines/model_pixart.py
+++ b/pipelines/model_pixart.py
@@ -4,6 +4,10 @@ from huggingface_hub import file_exists
 from modules import shared, devices, sd_models, model_quant, sd_hijack_te
 from modules.logger import log
 from pipelines import generic
+from pipelines.native_transformer import TransformerSpec
+
+
+PIXART_SPEC = TransformerSpec(cls=diffusers.PixArtTransformer2DModel)
 
 
 def load_pixart(checkpoint_info, diffusers_load_config=None):
@@ -24,7 +28,7 @@ def load_pixart(checkpoint_info, diffusers_load_config=None):
     load_args, _quant_args = model_quant.get_dit_args(diffusers_load_config, allow_quant=False)
     log.debug(f'Load model: type=PixArtSigma repo="{repo_id}" config={diffusers_load_config} offload={shared.opts.diffusers_offload_mode} dtype={devices.dtype} args={load_args}')
 
-    transformer = generic.load_transformer(repo_id, cls_name=diffusers.PixArtTransformer2DModel, load_config=diffusers_load_config)
+    transformer = generic.load_transformer(repo_id, cls_name=diffusers.PixArtTransformer2DModel, load_config=diffusers_load_config, native_spec=PIXART_SPEC)
     text_encoder = generic.load_text_encoder(repo_id_tenc, cls_name=transformers.T5EncoderModel, load_config=diffusers_load_config)
 
     pipe = diffusers.PixArtSigmaPipeline.from_pretrained(
diff --git a/pipelines/model_prx.py b/pipelines/model_prx.py
index 52088444c..8b3fc3a53 100644
--- a/pipelines/model_prx.py
+++ b/pipelines/model_prx.py
@@ -2,6 +2,10 @@ import diffusers
 from modules import shared, devices, sd_models, model_quant, sd_hijack_te
 from modules.logger import log
 from pipelines import generic
+from pipelines.native_transformer import TransformerSpec
+
+
+PRX_SPEC = TransformerSpec(cls=diffusers.PRXTransformer2DModel)
 
 
 def load_prx(checkpoint_info, diffusers_load_config=None):
@@ -14,7 +18,7 @@ def load_prx(checkpoint_info, diffusers_load_config=None):
     log.debug(f'Load model: type=PRX repo="{repo_id}" config={diffusers_load_config} offload={shared.opts.diffusers_offload_mode} dtype={devices.dtype} args={load_args}')
 
     from transformers.models.t5gemma.modeling_t5gemma import T5GemmaEncoder
-    transformer = generic.load_transformer(repo_id, cls_name=diffusers.PRXTransformer2DModel, load_config=diffusers_load_config)
+    transformer = generic.load_transformer(repo_id, cls_name=diffusers.PRXTransformer2DModel, load_config=diffusers_load_config, native_spec=PRX_SPEC)
     text_encoder = generic.load_text_encoder(repo_id, cls_name=T5GemmaEncoder, load_config=diffusers_load_config)
 
     pipe = diffusers.PRXPipeline.from_pretrained(