Fix lora conversion function for ai-toolkit Qwen Image LoRAs

src/diffusers/loaders/lora_conversion_utils.py

@@ -2129,6 +2129,10 @@ def _convert_non_diffusers_ltxv_lora_to_diffusers(state_dict, non_diffusers_pref
 
 
 def _convert_non_diffusers_qwen_lora_to_diffusers(state_dict):
+    has_diffusion_model = any(k.startswith("diffusion_model.") for k in state_dict)
+    if has_diffusion_model:
+        state_dict = {k.removeprefix("diffusion_model."): v for k, v in state_dict.items()}
+
     has_lora_unet = any(k.startswith("lora_unet_") for k in state_dict)
     if has_lora_unet:
         state_dict = {k.removeprefix("lora_unet_"): v for k, v in state_dict.items()}
@@ -2201,29 +2205,44 @@ def convert_key(key: str) -> str:
     all_keys = list(state_dict.keys())
     down_key = ".lora_down.weight"
     up_key = ".lora_up.weight"
+    a_key = ".lora_A.weight"
+    b_key = ".lora_B.weight"
 
-    def get_alpha_scales(down_weight, alpha_key):
-        rank = down_weight.shape[0]
-        alpha = state_dict.pop(alpha_key).item()
-        scale = alpha / rank  # LoRA is scaled by 'alpha / rank' in forward pass, so we need to scale it back here
-        scale_down = scale
-        scale_up = 1.0
-        while scale_down * 2 < scale_up:
-            scale_down *= 2
-            scale_up /= 2
-        return scale_down, scale_up
+    has_lora_arrow = any(down_key in k or up_key in k for k in all_keys)
+    has_lora_letter = any(a_key in k or b_key in k for k in all_keys)
 
-    for k in all_keys:
-        if k.endswith(down_key):
-            diffusers_down_key = k.replace(down_key, ".lora_A.weight")
-            diffusers_up_key = k.replace(down_key, up_key).replace(up_key, ".lora_B.weight")
-            alpha_key = k.replace(down_key, ".alpha")
-
-            down_weight = state_dict.pop(k)
-            up_weight = state_dict.pop(k.replace(down_key, up_key))
-            scale_down, scale_up = get_alpha_scales(down_weight, alpha_key)
-            converted_state_dict[diffusers_down_key] = down_weight * scale_down
-            converted_state_dict[diffusers_up_key] = up_weight * scale_up
+    if has_lora_arrow:
+
+        def get_alpha_scales(down_weight, alpha_key):
+            rank = down_weight.shape[0]
+            alpha = state_dict.pop(alpha_key).item()
+            scale = alpha / rank  # LoRA is scaled by 'alpha / rank' in forward pass, so we need to scale it back here
+            scale_down = scale
+            scale_up = 1.0
+            while scale_down * 2 < scale_up:
+                scale_down *= 2
+                scale_up /= 2
+            return scale_down, scale_up
+
+        for k in all_keys:
+            if k.endswith(down_key):
+                diffusers_down_key = k.replace(down_key, ".lora_A.weight")
+                diffusers_up_key = k.replace(down_key, up_key).replace(up_key, ".lora_B.weight")
+                alpha_key = k.replace(down_key, ".alpha")
+
+                down_weight = state_dict.pop(k)
+                up_weight = state_dict.pop(k.replace(down_key, up_key))
+                scale_down, scale_up = get_alpha_scales(down_weight, alpha_key)
+                converted_state_dict[diffusers_down_key] = down_weight * scale_down
+                converted_state_dict[diffusers_up_key] = up_weight * scale_up
+
+    # Already in diffusers format (lora_A/lora_B), just pop
+    elif has_lora_letter:
+        for k in all_keys:
+            if a_key in k or b_key in k:
+                converted_state_dict[k] = state_dict.pop(k)
+            elif ".alpha" in k:
+                state_dict.pop(k)
 
     if len(state_dict) > 0:
         raise ValueError(f"`state_dict` should be empty at this point but has {state_dict.keys()=}")

src/diffusers/loaders/lora_pipeline.py

@@ -6684,7 +6684,8 @@ def lora_state_dict(
 
         has_alphas_in_sd = any(k.endswith(".alpha") for k in state_dict)
         has_lora_unet = any(k.startswith("lora_unet_") for k in state_dict)
-        if has_alphas_in_sd or has_lora_unet:
+        has_diffusion_model = any(k.startswith("diffusion_model.") for k in state_dict)
+        if has_alphas_in_sd or has_lora_unet or has_diffusion_model:
             state_dict = _convert_non_diffusers_qwen_lora_to_diffusers(state_dict)
 
         out = (state_dict, metadata) if return_lora_metadata else state_dict