LTX2 condition pipeline initial commit

change get_image_features API

Signed-off-by: JaredforReal <w13431838023@gmail.com>
Co-authored-by: YiYi Xu <yixu310@gmail.com>

src/diffusers/__init__.py

@@ -552,6 +552,7 @@ else:
             "LEditsPPPipelineStableDiffusionXL",
             "LongCatImageEditPipeline",
             "LongCatImagePipeline",
+            "LTX2ConditionPipeline",
             "LTX2ImageToVideoPipeline",
             "LTX2LatentUpsamplePipeline",
             "LTX2Pipeline",
@@ -1284,6 +1285,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             LEditsPPPipelineStableDiffusionXL,
             LongCatImageEditPipeline,
             LongCatImagePipeline,
+            LTX2ConditionPipeline,
             LTX2ImageToVideoPipeline,
             LTX2LatentUpsamplePipeline,
             LTX2Pipeline,

src/diffusers/commands/fp16_safetensors.py

@@ -35,8 +35,8 @@ from . import BaseDiffusersCLICommand
 def conversion_command_factory(args: Namespace):
     if args.use_auth_token:
         warnings.warn(
-            "The `--use_auth_token` flag is deprecated and will be removed in a future version. Authentication is now"
-            " handled automatically if user is logged in."
+            "The `--use_auth_token` flag is deprecated and will be removed in a future version."
+            "Authentication is now handled automatically if the user is logged in."
         )
     return FP16SafetensorsCommand(args.ckpt_id, args.fp16, args.use_safetensors)
 
@@ -92,8 +92,8 @@ class FP16SafetensorsCommand(BaseDiffusersCLICommand):
         pipeline_class = getattr(import_module("diffusers"), pipeline_class_name)
         self.logger.info(f"Pipeline class imported: {pipeline_class_name}.")
 
-        # Load the appropriate pipeline. We could have use `DiffusionPipeline`
-        # here, but just to avoid any rough edge cases.
+        # Load the appropriate pipeline. We could have used `DiffusionPipeline`
+        # here, but just to avoid potential edge cases.
         pipeline = pipeline_class.from_pretrained(
             self.ckpt_id, torch_dtype=torch.float16 if self.fp16 else torch.float32
         )

src/diffusers/pipelines/__init__.py

@@ -291,7 +291,11 @@ else:
         "LTXLatentUpsamplePipeline",
         "LTXI2VLongMultiPromptPipeline",
     ]
-    _import_structure["ltx2"] = ["LTX2Pipeline", "LTX2ImageToVideoPipeline", "LTX2LatentUpsamplePipeline"]
+    _import_structure["ltx2"] = [
+        "LTX2Pipeline",
+        "LTX2ConditionPipelineLTX2ImageToVideoPipeline",
+        "LTX2LatentUpsamplePipeline",
+    ]
     _import_structure["lumina"] = ["LuminaPipeline", "LuminaText2ImgPipeline"]
     _import_structure["lumina2"] = ["Lumina2Pipeline", "Lumina2Text2ImgPipeline"]
     _import_structure["lucy"] = ["LucyEditPipeline"]
@@ -742,7 +746,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             LTXLatentUpsamplePipeline,
             LTXPipeline,
         )
-        from .ltx2 import LTX2ImageToVideoPipeline, LTX2LatentUpsamplePipeline, LTX2Pipeline
+        from .ltx2 import LTX2ConditionPipeline, LTX2ImageToVideoPipeline, LTX2LatentUpsamplePipeline, LTX2Pipeline
         from .lucy import LucyEditPipeline
         from .lumina import LuminaPipeline, LuminaText2ImgPipeline
         from .lumina2 import Lumina2Pipeline, Lumina2Text2ImgPipeline

src/diffusers/pipelines/glm_image/pipeline_glm_image.py

@@ -407,8 +407,8 @@ class GlmImagePipeline(DiffusionPipeline):
 
             if len(source_grids) > 0:
                 prior_token_image_embed = self.vision_language_encoder.get_image_features(
-                    inputs["pixel_values"], source_grids, return_dict=False
-                )
+                    inputs["pixel_values"], source_grids
+                ).pooler_output
                 prior_token_image_embed = torch.cat(prior_token_image_embed, dim=0)
                 prior_token_image_ids_d32 = self.vision_language_encoder.get_image_tokens(
                     prior_token_image_embed, source_grids

src/diffusers/pipelines/ltx2/__init__.py

@@ -25,6 +25,7 @@ else:
     _import_structure["connectors"] = ["LTX2TextConnectors"]
     _import_structure["latent_upsampler"] = ["LTX2LatentUpsamplerModel"]
     _import_structure["pipeline_ltx2"] = ["LTX2Pipeline"]
+    _import_structure["pipeline_ltx2_condition"] = ["LTX2ConditionPipeline"]
     _import_structure["pipeline_ltx2_image2video"] = ["LTX2ImageToVideoPipeline"]
     _import_structure["pipeline_ltx2_latent_upsample"] = ["LTX2LatentUpsamplePipeline"]
     _import_structure["vocoder"] = ["LTX2Vocoder"]
@@ -40,6 +41,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
         from .connectors import LTX2TextConnectors
         from .latent_upsampler import LTX2LatentUpsamplerModel
         from .pipeline_ltx2 import LTX2Pipeline
+        from .pipeline_ltx2_condition import LTX2ConditionPipeline
         from .pipeline_ltx2_image2video import LTX2ImageToVideoPipeline
         from .pipeline_ltx2_latent_upsample import LTX2LatentUpsamplePipeline
         from .vocoder import LTX2Vocoder

src/diffusers/pipelines/wan/pipeline_wan.py

@@ -496,8 +496,13 @@ class WanPipeline(DiffusionPipeline, WanLoraLoaderMixin):
             num_frames = num_frames // self.vae_scale_factor_temporal * self.vae_scale_factor_temporal + 1
         num_frames = max(num_frames, 1)
 
-        h_multiple_of = self.vae_scale_factor_spatial * self.transformer.config.patch_size[1]
-        w_multiple_of = self.vae_scale_factor_spatial * self.transformer.config.patch_size[2]
+        patch_size = (
+            self.transformer.config.patch_size
+            if self.transformer is not None
+            else self.transformer_2.config.patch_size
+        )
+        h_multiple_of = self.vae_scale_factor_spatial * patch_size[1]
+        w_multiple_of = self.vae_scale_factor_spatial * patch_size[2]
         calc_height = height // h_multiple_of * h_multiple_of
         calc_width = width // w_multiple_of * w_multiple_of
         if height != calc_height or width != calc_width:

src/diffusers/pipelines/wan/pipeline_wan_i2v.py

@@ -637,8 +637,13 @@ class WanImageToVideoPipeline(DiffusionPipeline, WanLoraLoaderMixin):
             num_frames = num_frames // self.vae_scale_factor_temporal * self.vae_scale_factor_temporal + 1
         num_frames = max(num_frames, 1)
 
-        h_multiple_of = self.vae_scale_factor_spatial * self.transformer.config.patch_size[1]
-        w_multiple_of = self.vae_scale_factor_spatial * self.transformer.config.patch_size[2]
+        patch_size = (
+            self.transformer.config.patch_size
+            if self.transformer is not None
+            else self.transformer_2.config.patch_size
+        )
+        h_multiple_of = self.vae_scale_factor_spatial * patch_size[1]
+        w_multiple_of = self.vae_scale_factor_spatial * patch_size[2]
         calc_height = height // h_multiple_of * h_multiple_of
         calc_width = width // w_multiple_of * w_multiple_of
         if height != calc_height or width != calc_width: