Rework img/video reference in Minimax H3 to more closely match the comfy ui implementation.

toolkit/dataloader_mixins.py

@@ -2309,10 +2309,15 @@ class TextEmbeddingCachingMixin:
                         self.sd.set_device_state_preset('cache_text_encoder')
                         did_move = True
 
-                    if file_item.encode_control_in_text_embeddings and file_item.control_path is not None:
+                    control_video_paths = getattr(file_item, 'control_video_paths', None) or []
+                    if file_item.encode_control_in_text_embeddings and (
+                        file_item.control_path is not None or len(control_video_paths) > 0
+                    ):
                         ctrl_img_list = []
                         control_path_list = file_item.control_path
-                        if not isinstance(file_item.control_path, list):
+                        if control_path_list is None:
+                            control_path_list = []
+                        elif not isinstance(control_path_list, list):
                             control_path_list = [control_path_list]
                         for i in range(len(control_path_list)):
                             try:
@@ -2328,6 +2333,14 @@ class TextEmbeddingCachingMixin:
                             except Exception as e:
                                 print_acc(f"Error: {e}")
                                 print_acc(f"Error loading control image: {control_path_list[i]}")
+                        # control VIDEOS ride into the presentation by path (models
+                        # with supports_video_control_images turn them into
+                        # timestamped vision blocks); images first, then videos.
+                        # The model needs the dataset config to treat the clip
+                        # exactly like its latent rows (frame count / trim)
+                        ctrl_img_list.extend(control_video_paths)
+                        if len(control_video_paths) > 0:
+                            self.sd._ref_video_dataset_config = self.dataset_config
                         
                         if len(ctrl_img_list) == 0:
                             ctrl_img = None

toolkit/models/base_model.py

@@ -283,6 +283,13 @@ class BaseModel:
             divisibility = divisibility * 2
         return divisibility
 
+    def prepare_sample_prompt_context(self, gen_config):
+        """Optional hook called right before a sample prompt is encoded, with
+        the sample's GenerateImageConfig, for models whose control conditioning
+        in the text embeds depends on sample settings (e.g. a video reference's
+        length capped at the sample's frame count)."""
+        return None
+
     def get_frame_count_snapper(self):
         """Optional hook for video models whose VAE accepts frame counts on a
         grid other than the default ``temporal_compression * n + 1``.
@@ -605,6 +612,7 @@ class BaseModel:
                             else:
                                 ctrl_img = ctrl_img_list[0] if len(ctrl_img_list) > 0 else None
                         # encode the prompt ourselves so we can do fun stuff with embeddings
+                        self.prepare_sample_prompt_context(gen_config)
                         if isinstance(self.adapter, CustomAdapter):
                             self.adapter.is_unconditional_run = False
                         conditional_embeds = self.encode_prompt(

toolkit/stable_diffusion_model.py

@@ -215,6 +215,9 @@ class StableDiffusion:
         
         # set true for models that encode control image into text embeddings
         self.encode_control_in_text_embeddings = False
+        # control files may be VIDEOS (paths exposed on the batch as
+        # control_video_paths_list); see minimax_h3 ref2va
+        self.supports_video_control_images = False
         # control images will come in as a list for encoding some things if true
         self.has_multiple_control_images = False
         # do not resize control images
@@ -292,7 +295,20 @@ class StableDiffusion:
         if self.is_flux or self.is_v3:
             divisibility = divisibility * 2
         return divisibility * 2 # todo remove this
-        
+
+    def get_frame_count_snapper(self):
+        """Optional hook for video models whose VAE accepts frame counts on a
+        grid other than the default ``temporal_compression * n + 1``. Return a
+        MODULE-LEVEL function ``(num_frames) -> int`` (picklable — file items
+        travel into dataloader workers) that snaps a frame count DOWN to a
+        valid count, or None for the default auto_frame_count math."""
+        return None
+
+    def prepare_sample_prompt_context(self, gen_config):
+        """Optional hook called right before a sample prompt is encoded, with
+        the sample's GenerateImageConfig, for models whose control conditioning
+        in the text embeds depends on sample settings."""
+        return None
 
     def load_model(self):
         if self.is_loaded: