feat(pipelines): wire native_spec for Qwen-Image and custom-class arches

Qwen-Image: diffusers registers a no-op identity lambda for
QwenImageTransformer2DModel; from_single_file silently accepted any
key naming and loaded with mismatches. QWEN_SPEC sets converter=None
to short-circuit the no-op.

Custom transformer classes (Lens, Step1X-Edit, VIBE Sana, f-lite):
runtime-registered onto diffusers, never in SINGLE_FILE_LOADABLE_CLASSES.
Specs land in each arch's existing pipelines/<arch>/__init__.py.

NextStep skipped: loader returns None up front.

pipelines/f_lite/__init__.py

@@ -1,5 +1,9 @@
+from pipelines.native_transformer import TransformerSpec
 from .pipeline import FLitePipeline, FLitePipelineOutput, APGConfig
 from .model import DiT
 
 
-__all__ = ["APGConfig", "DiT", "FLitePipeline", "FLitePipelineOutput"]
+FLITE_SPEC = TransformerSpec(cls=DiT, subfolder='dit_model')
+
+
+__all__ = ["APGConfig", "DiT", "FLITE_SPEC", "FLitePipeline", "FLitePipelineOutput"]

pipelines/lens/__init__.py

@@ -2,6 +2,7 @@
 
 import diffusers as _diffusers
 import transformers as _transformers
+from pipelines.native_transformer import TransformerSpec
 from .pipeline import LensPipeline, LensPipelineOutput
 from .pipeline_image import LensImg2ImgPipeline, LensInpaintPipeline
 from .reasoner import PromptReasoner
@@ -9,6 +10,9 @@ from .resolution import RESOLUTION_BUCKETS, resolve_resolution
 from .text_encoder import LensGptOssEncoder
 from .transformer import LensTransformer2DModel
 
+
+LENS_SPEC = TransformerSpec(cls=LensTransformer2DModel)
+
 # ---------------------------------------------------------------------------
 # Make our custom subclasses discoverable by ``diffusers.DiffusionPipeline``.
 #
@@ -38,6 +42,7 @@ if not hasattr(_diffusers, "LensInpaintPipeline"):
 del _diffusers, _transformers
 
 __all__ = [
+    "LENS_SPEC",
     "LensPipeline",
     "LensPipelineOutput",
     "LensImg2ImgPipeline",

pipelines/model_flite.py

@@ -19,7 +19,7 @@ def load_flite(checkpoint_info, diffusers_load_config=None):
     diffusers.FLitePipeline = f_lite.FLitePipeline
     sys.modules['f_lite'] = f_lite
 
-    dit_model = generic.load_transformer(repo_id, cls_name=f_lite.DiT, load_config=diffusers_load_config, subfolder="dit_model")
+    dit_model = generic.load_transformer(repo_id, cls_name=f_lite.DiT, load_config=diffusers_load_config, subfolder="dit_model", native_spec=f_lite.FLITE_SPEC)
     text_encoder = generic.load_text_encoder(repo_id, cls_name=transformers.T5EncoderModel, load_config=diffusers_load_config, subfolder="text_encoder")
 
     pipe = f_lite.FLitePipeline.from_pretrained(

pipelines/model_lens.py

@@ -15,7 +15,7 @@ def load_lens(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=Lens repo="{repo_id}" config={diffusers_load_config} offload={shared.opts.diffusers_offload_mode} dtype={devices.dtype} reasoner={shared.opts.model_lens_enable_pe} args={load_args}')
 
-    transformer = generic.load_transformer(repo_id, cls_name=lens.LensTransformer2DModel, load_config=diffusers_load_config)
+    transformer = generic.load_transformer(repo_id, cls_name=lens.LensTransformer2DModel, load_config=diffusers_load_config, native_spec=lens.LENS_SPEC)
     text_encoder = generic.load_text_encoder(repo_id, cls_name=lens.LensGptOssEncoder, load_config=diffusers_load_config, allow_quant=False)
 
     pipe = lens.LensPipeline.from_pretrained(

pipelines/model_qwen.py

@@ -57,12 +57,14 @@ def load_qwen(checkpoint_info, diffusers_load_config=None):
             transformer_subfolder = "transformer"
 
     if transformer is None:
+        from pipelines.qwen import QWEN_SPEC
         transformer = generic.load_transformer(
             repo_transformer,
             subfolder=transformer_subfolder,
             cls_name=diffusers.QwenImageTransformer2DModel,
             load_config=diffusers_load_config,
             modules_to_not_convert=["transformer_blocks.0.img_mod.1.weight"],
+            native_spec=QWEN_SPEC,
         )
 
     repo_te = 'Qwen/Qwen-Image'

pipelines/model_step1x_edit.py

@@ -22,7 +22,8 @@ def load_step1x_edit(checkpoint_info, diffusers_load_config=None):
 
     text_encoder = generic.load_text_encoder(repo_id, cls_name=transformers.Qwen2_5_VLForConditionalGeneration, load_config=diffusers_load_config)
     processor = transformers.Qwen2_5_VLProcessor.from_pretrained(repo_id, cache_dir=shared.opts.hfcache_dir, subfolder='processor')
-    transformer = generic.load_transformer(repo_id, cls_name=Step1XEditTransformer2DModel, load_config=diffusers_load_config)
+    from pipelines.step1x import STEP1X_SPEC
+    transformer = generic.load_transformer(repo_id, cls_name=Step1XEditTransformer2DModel, load_config=diffusers_load_config, native_spec=STEP1X_SPEC)
 
     pipe = Step1XEditPipeline.from_pretrained(
         repo_id,

pipelines/model_vibe.py

@@ -21,11 +21,13 @@ def load_vibe(checkpoint_info, diffusers_load_config=None):
 
     sys.modules['vibe.transformer.vibe_sana_editing'] = diffusers # monkey patch since hf model_index.json points to custom class path
 
+    from pipelines.vibe import VIBE_SPEC
     transformer = generic.load_transformer(
         repo_id,
         cls_name=VIBESanaEditingModel,
         load_config=diffusers_load_config,
         allow_quant=False,
+        native_spec=VIBE_SPEC,
     )
     text_encoder = generic.load_text_encoder(
         repo_id,

pipelines/qwen/__init__.py

@@ -1,2 +1,24 @@
+"""Qwen-Image pipeline package.
+
+Exports :data:`QWEN_SPEC` for use by :mod:`pipelines.model_qwen` together
+with :mod:`pipelines.native_transformer`. Qwen-Image is the lone Mode C
+arch: diffusers registers a no-op identity lambda as its
+``SINGLE_FILE_LOADABLE_CLASSES`` converter, so ``from_single_file`` silently
+accepts whatever key naming the community file uses and loads with mangled
+weights instead of raising.
+
+The spec explicitly sets ``converter=None`` to short-circuit the no-op
+pickup; if a real Qwen-Image converter is needed for some trainer dump
+in the wild, it can be plugged in here. Until then, validation surfaces a
+clear error listing unexpected/missing keys instead of letting a malformed
+file load silently.
+"""
+
+import diffusers
+
+from pipelines.native_transformer import TransformerSpec
 from pipelines.qwen.qwen_nunchaku import load_qwen_nunchaku
 from pipelines.qwen.qwen_pruning import check_qwen_pruning
+
+
+QWEN_SPEC = TransformerSpec(cls=diffusers.QwenImageTransformer2DModel, converter=None)

pipelines/step1x/__init__.py

@@ -1,8 +1,14 @@
+from pipelines.native_transformer import TransformerSpec
 from pipelines.step1x.pipeline_output import Step1XEditPipelineOutput
 from pipelines.step1x.pipeline_step1x_edit import Step1XEditPipeline
 from pipelines.step1x.transformer_step1x_edit import Step1XEditTransformer2DModel
 
+
+STEP1X_SPEC = TransformerSpec(cls=Step1XEditTransformer2DModel)
+
+
 __all__ = [
+    "STEP1X_SPEC",
     "Step1XEditPipeline",
     "Step1XEditPipelineOutput",
     "Step1XEditTransformer2DModel",

pipelines/vibe/__init__.py

@@ -1,6 +1,11 @@
 """VIBE pipeline components for SD.Next."""
 
+from pipelines.native_transformer import TransformerSpec
 from .vibe_sana_editing import VIBESanaEditingModel
 from .vibe_sana_pipeline import VIBESanaEditingPipeline, VIBESanaImagePipeline
 
-__all__ = ["VIBESanaEditingModel", "VIBESanaEditingPipeline", "VIBESanaImagePipeline"]
+
+VIBE_SPEC = TransformerSpec(cls=VIBESanaEditingModel)
+
+
+__all__ = ["VIBE_SPEC", "VIBESanaEditingModel", "VIBESanaEditingPipeline", "VIBESanaImagePipeline"]