Support ERNIE-Image (#1389)

* ernie-image pipeline

* ernie-image inference and training

* style fix

* ernie docs

* lowvram

* final style fix

* pr-review

* pr-fix round2

* set uniform training weight

* fix

* update lowvram docs

diffsynth/utils/state_dict_converters/ernie_image_text_encoder.py

@@ -0,0 +1,21 @@
+def ErnieImageTextEncoderStateDictConverter(state_dict):
+    """
+    Maps checkpoint keys from multimodal Mistral3Model format
+    to text-only Ministral3Model format.
+
+    Checkpoint keys (Mistral3Model):
+        language_model.model.layers.0.input_layernorm.weight
+        language_model.model.norm.weight
+
+    Model keys (ErnieImageTextEncoder → self.model = Ministral3Model):
+        model.layers.0.input_layernorm.weight
+        model.norm.weight
+
+    Mapping: language_model. → model.
+    """
+    new_state_dict = {}
+    for key in state_dict:
+        if key.startswith("language_model.model."):
+            new_key = key.replace("language_model.model.", "model.", 1)
+            new_state_dict[new_key] = state_dict[key]
+    return new_state_dict