diffusers 0.34.0__py3-none-any.whl → 0.35.1__py3-none-any.whl

diffusers/loaders/lora_conversion_utils.py

@@ -817,7 +817,11 @@ def _convert_kohya_flux_lora_to_diffusers(state_dict):
     # has both `peft` and non-peft state dict.
     has_peft_state_dict = any(k.startswith("transformer.") for k in state_dict)
     if has_peft_state_dict:
-        state_dict = {k: v for k, v in state_dict.items() if k.startswith("transformer.")}
+        state_dict = {
+            k.replace("lora_down.weight", "lora_A.weight").replace("lora_up.weight", "lora_B.weight"): v
+            for k, v in state_dict.items()
+            if k.startswith("transformer.")
+        }
         return state_dict
 
     # Another weird one.
@@ -1346,6 +1350,228 @@ def _convert_bfl_flux_control_lora_to_diffusers(original_state_dict):
     return converted_state_dict
 
 
+def _convert_fal_kontext_lora_to_diffusers(original_state_dict):
+    converted_state_dict = {}
+    original_state_dict_keys = list(original_state_dict.keys())
+    num_layers = 19
+    num_single_layers = 38
+    inner_dim = 3072
+    mlp_ratio = 4.0
+
+    # double transformer blocks
+    for i in range(num_layers):
+        block_prefix = f"transformer_blocks.{i}."
+        original_block_prefix = "base_model.model."
+
+        for lora_key in ["lora_A", "lora_B"]:
+            # norms
+            converted_state_dict[f"{block_prefix}norm1.linear.{lora_key}.weight"] = original_state_dict.pop(
+                f"{original_block_prefix}double_blocks.{i}.img_mod.lin.{lora_key}.weight"
+            )
+            if f"double_blocks.{i}.img_mod.lin.{lora_key}.bias" in original_state_dict_keys:
+                converted_state_dict[f"{block_prefix}norm1.linear.{lora_key}.bias"] = original_state_dict.pop(
+                    f"{original_block_prefix}double_blocks.{i}.img_mod.lin.{lora_key}.bias"
+                )
+
+            converted_state_dict[f"{block_prefix}norm1_context.linear.{lora_key}.weight"] = original_state_dict.pop(
+                f"{original_block_prefix}double_blocks.{i}.txt_mod.lin.{lora_key}.weight"
+            )
+
+            # Q, K, V
+            if lora_key == "lora_A":
+                sample_lora_weight = original_state_dict.pop(
+                    f"{original_block_prefix}double_blocks.{i}.img_attn.qkv.{lora_key}.weight"
+                )
+                converted_state_dict[f"{block_prefix}attn.to_v.{lora_key}.weight"] = torch.cat([sample_lora_weight])
+                converted_state_dict[f"{block_prefix}attn.to_q.{lora_key}.weight"] = torch.cat([sample_lora_weight])
+                converted_state_dict[f"{block_prefix}attn.to_k.{lora_key}.weight"] = torch.cat([sample_lora_weight])
+
+                context_lora_weight = original_state_dict.pop(
+                    f"{original_block_prefix}double_blocks.{i}.txt_attn.qkv.{lora_key}.weight"
+                )
+                converted_state_dict[f"{block_prefix}attn.add_q_proj.{lora_key}.weight"] = torch.cat(
+                    [context_lora_weight]
+                )
+                converted_state_dict[f"{block_prefix}attn.add_k_proj.{lora_key}.weight"] = torch.cat(
+                    [context_lora_weight]
+                )
+                converted_state_dict[f"{block_prefix}attn.add_v_proj.{lora_key}.weight"] = torch.cat(
+                    [context_lora_weight]
+                )
+            else:
+                sample_q, sample_k, sample_v = torch.chunk(
+                    original_state_dict.pop(
+                        f"{original_block_prefix}double_blocks.{i}.img_attn.qkv.{lora_key}.weight"
+                    ),
+                    3,
+                    dim=0,
+                )
+                converted_state_dict[f"{block_prefix}attn.to_q.{lora_key}.weight"] = torch.cat([sample_q])
+                converted_state_dict[f"{block_prefix}attn.to_k.{lora_key}.weight"] = torch.cat([sample_k])
+                converted_state_dict[f"{block_prefix}attn.to_v.{lora_key}.weight"] = torch.cat([sample_v])
+
+                context_q, context_k, context_v = torch.chunk(
+                    original_state_dict.pop(
+                        f"{original_block_prefix}double_blocks.{i}.txt_attn.qkv.{lora_key}.weight"
+                    ),
+                    3,
+                    dim=0,
+                )
+                converted_state_dict[f"{block_prefix}attn.add_q_proj.{lora_key}.weight"] = torch.cat([context_q])
+                converted_state_dict[f"{block_prefix}attn.add_k_proj.{lora_key}.weight"] = torch.cat([context_k])
+                converted_state_dict[f"{block_prefix}attn.add_v_proj.{lora_key}.weight"] = torch.cat([context_v])
+
+            if f"double_blocks.{i}.img_attn.qkv.{lora_key}.bias" in original_state_dict_keys:
+                sample_q_bias, sample_k_bias, sample_v_bias = torch.chunk(
+                    original_state_dict.pop(f"{original_block_prefix}double_blocks.{i}.img_attn.qkv.{lora_key}.bias"),
+                    3,
+                    dim=0,
+                )
+                converted_state_dict[f"{block_prefix}attn.to_q.{lora_key}.bias"] = torch.cat([sample_q_bias])
+                converted_state_dict[f"{block_prefix}attn.to_k.{lora_key}.bias"] = torch.cat([sample_k_bias])
+                converted_state_dict[f"{block_prefix}attn.to_v.{lora_key}.bias"] = torch.cat([sample_v_bias])
+
+            if f"double_blocks.{i}.txt_attn.qkv.{lora_key}.bias" in original_state_dict_keys:
+                context_q_bias, context_k_bias, context_v_bias = torch.chunk(
+                    original_state_dict.pop(f"{original_block_prefix}double_blocks.{i}.txt_attn.qkv.{lora_key}.bias"),
+                    3,
+                    dim=0,
+                )
+                converted_state_dict[f"{block_prefix}attn.add_q_proj.{lora_key}.bias"] = torch.cat([context_q_bias])
+                converted_state_dict[f"{block_prefix}attn.add_k_proj.{lora_key}.bias"] = torch.cat([context_k_bias])
+                converted_state_dict[f"{block_prefix}attn.add_v_proj.{lora_key}.bias"] = torch.cat([context_v_bias])
+
+            # ff img_mlp
+            converted_state_dict[f"{block_prefix}ff.net.0.proj.{lora_key}.weight"] = original_state_dict.pop(
+                f"{original_block_prefix}double_blocks.{i}.img_mlp.0.{lora_key}.weight"
+            )
+            if f"{original_block_prefix}double_blocks.{i}.img_mlp.0.{lora_key}.bias" in original_state_dict_keys:
+                converted_state_dict[f"{block_prefix}ff.net.0.proj.{lora_key}.bias"] = original_state_dict.pop(
+                    f"{original_block_prefix}double_blocks.{i}.img_mlp.0.{lora_key}.bias"
+                )
+
+            converted_state_dict[f"{block_prefix}ff.net.2.{lora_key}.weight"] = original_state_dict.pop(
+                f"{original_block_prefix}double_blocks.{i}.img_mlp.2.{lora_key}.weight"
+            )
+            if f"{original_block_prefix}double_blocks.{i}.img_mlp.2.{lora_key}.bias" in original_state_dict_keys:
+                converted_state_dict[f"{block_prefix}ff.net.2.{lora_key}.bias"] = original_state_dict.pop(
+                    f"{original_block_prefix}double_blocks.{i}.img_mlp.2.{lora_key}.bias"
+                )
+
+            converted_state_dict[f"{block_prefix}ff_context.net.0.proj.{lora_key}.weight"] = original_state_dict.pop(
+                f"{original_block_prefix}double_blocks.{i}.txt_mlp.0.{lora_key}.weight"
+            )
+            if f"{original_block_prefix}double_blocks.{i}.txt_mlp.0.{lora_key}.bias" in original_state_dict_keys:
+                converted_state_dict[f"{block_prefix}ff_context.net.0.proj.{lora_key}.bias"] = original_state_dict.pop(
+                    f"{original_block_prefix}double_blocks.{i}.txt_mlp.0.{lora_key}.bias"
+                )
+
+            converted_state_dict[f"{block_prefix}ff_context.net.2.{lora_key}.weight"] = original_state_dict.pop(
+                f"{original_block_prefix}double_blocks.{i}.txt_mlp.2.{lora_key}.weight"
+            )
+            if f"{original_block_prefix}double_blocks.{i}.txt_mlp.2.{lora_key}.bias" in original_state_dict_keys:
+                converted_state_dict[f"{block_prefix}ff_context.net.2.{lora_key}.bias"] = original_state_dict.pop(
+                    f"{original_block_prefix}double_blocks.{i}.txt_mlp.2.{lora_key}.bias"
+                )
+
+            # output projections.
+            converted_state_dict[f"{block_prefix}attn.to_out.0.{lora_key}.weight"] = original_state_dict.pop(
+                f"{original_block_prefix}double_blocks.{i}.img_attn.proj.{lora_key}.weight"
+            )
+            if f"{original_block_prefix}double_blocks.{i}.img_attn.proj.{lora_key}.bias" in original_state_dict_keys:
+                converted_state_dict[f"{block_prefix}attn.to_out.0.{lora_key}.bias"] = original_state_dict.pop(
+                    f"{original_block_prefix}double_blocks.{i}.img_attn.proj.{lora_key}.bias"
+                )
+            converted_state_dict[f"{block_prefix}attn.to_add_out.{lora_key}.weight"] = original_state_dict.pop(
+                f"{original_block_prefix}double_blocks.{i}.txt_attn.proj.{lora_key}.weight"
+            )
+            if f"{original_block_prefix}double_blocks.{i}.txt_attn.proj.{lora_key}.bias" in original_state_dict_keys:
+                converted_state_dict[f"{block_prefix}attn.to_add_out.{lora_key}.bias"] = original_state_dict.pop(
+                    f"{original_block_prefix}double_blocks.{i}.txt_attn.proj.{lora_key}.bias"
+                )
+
+    # single transformer blocks
+    for i in range(num_single_layers):
+        block_prefix = f"single_transformer_blocks.{i}."
+
+        for lora_key in ["lora_A", "lora_B"]:
+            # norm.linear  <- single_blocks.0.modulation.lin
+            converted_state_dict[f"{block_prefix}norm.linear.{lora_key}.weight"] = original_state_dict.pop(
+                f"{original_block_prefix}single_blocks.{i}.modulation.lin.{lora_key}.weight"
+            )
+            if f"{original_block_prefix}single_blocks.{i}.modulation.lin.{lora_key}.bias" in original_state_dict_keys:
+                converted_state_dict[f"{block_prefix}norm.linear.{lora_key}.bias"] = original_state_dict.pop(
+                    f"{original_block_prefix}single_blocks.{i}.modulation.lin.{lora_key}.bias"
+                )
+
+            # Q, K, V, mlp
+            mlp_hidden_dim = int(inner_dim * mlp_ratio)
+            split_size = (inner_dim, inner_dim, inner_dim, mlp_hidden_dim)
+
+            if lora_key == "lora_A":
+                lora_weight = original_state_dict.pop(
+                    f"{original_block_prefix}single_blocks.{i}.linear1.{lora_key}.weight"
+                )
+                converted_state_dict[f"{block_prefix}attn.to_q.{lora_key}.weight"] = torch.cat([lora_weight])
+                converted_state_dict[f"{block_prefix}attn.to_k.{lora_key}.weight"] = torch.cat([lora_weight])
+                converted_state_dict[f"{block_prefix}attn.to_v.{lora_key}.weight"] = torch.cat([lora_weight])
+                converted_state_dict[f"{block_prefix}proj_mlp.{lora_key}.weight"] = torch.cat([lora_weight])
+
+                if f"{original_block_prefix}single_blocks.{i}.linear1.{lora_key}.bias" in original_state_dict_keys:
+                    lora_bias = original_state_dict.pop(f"single_blocks.{i}.linear1.{lora_key}.bias")
+                    converted_state_dict[f"{block_prefix}attn.to_q.{lora_key}.bias"] = torch.cat([lora_bias])
+                    converted_state_dict[f"{block_prefix}attn.to_k.{lora_key}.bias"] = torch.cat([lora_bias])
+                    converted_state_dict[f"{block_prefix}attn.to_v.{lora_key}.bias"] = torch.cat([lora_bias])
+                    converted_state_dict[f"{block_prefix}proj_mlp.{lora_key}.bias"] = torch.cat([lora_bias])
+            else:
+                q, k, v, mlp = torch.split(
+                    original_state_dict.pop(f"{original_block_prefix}single_blocks.{i}.linear1.{lora_key}.weight"),
+                    split_size,
+                    dim=0,
+                )
+                converted_state_dict[f"{block_prefix}attn.to_q.{lora_key}.weight"] = torch.cat([q])
+                converted_state_dict[f"{block_prefix}attn.to_k.{lora_key}.weight"] = torch.cat([k])
+                converted_state_dict[f"{block_prefix}attn.to_v.{lora_key}.weight"] = torch.cat([v])
+                converted_state_dict[f"{block_prefix}proj_mlp.{lora_key}.weight"] = torch.cat([mlp])
+
+                if f"{original_block_prefix}single_blocks.{i}.linear1.{lora_key}.bias" in original_state_dict_keys:
+                    q_bias, k_bias, v_bias, mlp_bias = torch.split(
+                        original_state_dict.pop(f"{original_block_prefix}single_blocks.{i}.linear1.{lora_key}.bias"),
+                        split_size,
+                        dim=0,
+                    )
+                    converted_state_dict[f"{block_prefix}attn.to_q.{lora_key}.bias"] = torch.cat([q_bias])
+                    converted_state_dict[f"{block_prefix}attn.to_k.{lora_key}.bias"] = torch.cat([k_bias])
+                    converted_state_dict[f"{block_prefix}attn.to_v.{lora_key}.bias"] = torch.cat([v_bias])
+                    converted_state_dict[f"{block_prefix}proj_mlp.{lora_key}.bias"] = torch.cat([mlp_bias])
+
+            # output projections.
+            converted_state_dict[f"{block_prefix}proj_out.{lora_key}.weight"] = original_state_dict.pop(
+                f"{original_block_prefix}single_blocks.{i}.linear2.{lora_key}.weight"
+            )
+            if f"{original_block_prefix}single_blocks.{i}.linear2.{lora_key}.bias" in original_state_dict_keys:
+                converted_state_dict[f"{block_prefix}proj_out.{lora_key}.bias"] = original_state_dict.pop(
+                    f"{original_block_prefix}single_blocks.{i}.linear2.{lora_key}.bias"
+                )
+
+    for lora_key in ["lora_A", "lora_B"]:
+        converted_state_dict[f"proj_out.{lora_key}.weight"] = original_state_dict.pop(
+            f"{original_block_prefix}final_layer.linear.{lora_key}.weight"
+        )
+        if f"{original_block_prefix}final_layer.linear.{lora_key}.bias" in original_state_dict_keys:
+            converted_state_dict[f"proj_out.{lora_key}.bias"] = original_state_dict.pop(
+                f"{original_block_prefix}final_layer.linear.{lora_key}.bias"
+            )
+
+    if len(original_state_dict) > 0:
+        raise ValueError(f"`original_state_dict` should be empty at this point but has {original_state_dict.keys()=}.")
+
+    for key in list(converted_state_dict.keys()):
+        converted_state_dict[f"transformer.{key}"] = converted_state_dict.pop(key)
+
+    return converted_state_dict
+
+
 def _convert_hunyuan_video_lora_to_diffusers(original_state_dict):
     converted_state_dict = {k: original_state_dict.pop(k) for k in list(original_state_dict.keys())}
 
@@ -1603,24 +1829,33 @@ def _convert_non_diffusers_wan_lora_to_diffusers(state_dict):
     is_i2v_lora = any("k_img" in k for k in original_state_dict) and any("v_img" in k for k in original_state_dict)
     lora_down_key = "lora_A" if any("lora_A" in k for k in original_state_dict) else "lora_down"
     lora_up_key = "lora_B" if any("lora_B" in k for k in original_state_dict) else "lora_up"
+    has_time_projection_weight = any(
+        k.startswith("time_projection") and k.endswith(".weight") for k in original_state_dict
+    )
 
-    diff_keys = [k for k in original_state_dict if k.endswith((".diff_b", ".diff"))]
-    if diff_keys:
-        for diff_k in diff_keys:
-            param = original_state_dict[diff_k]
-            # The magnitudes of the .diff-ending weights are very low (most are below 1e-4, some are upto 1e-3,
-            # and 2 of them are about 1.6e-2 [the case with AccVideo lora]). The low magnitudes mostly correspond
-            # to norm layers. Ignoring them is the best option at the moment until a better solution is found. It
-            # is okay to ignore because they do not affect the model output in a significant manner.
-            threshold = 1.6e-2
-            absdiff = param.abs().max() - param.abs().min()
-            all_zero = torch.all(param == 0).item()
-            all_absdiff_lower_than_threshold = absdiff < threshold
-            if all_zero or all_absdiff_lower_than_threshold:
-                logger.debug(
-                    f"Removed {diff_k} key from the state dict as it's all zeros, or values lower than hardcoded threshold."
-                )
-                original_state_dict.pop(diff_k)
+    def get_alpha_scales(down_weight, alpha_key):
+        rank = down_weight.shape[0]
+        alpha = original_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 key in list(original_state_dict.keys()):
+        if key.endswith((".diff", ".diff_b")) and "norm" in key:
+            # NOTE: we don't support this because norm layer diff keys are just zeroed values. We can support it
+            # in future if needed and they are not zeroed.
+            original_state_dict.pop(key)
+            logger.debug(f"Removing {key} key from the state dict as it is a norm diff key. This is unsupported.")
+
+        if "time_projection" in key and not has_time_projection_weight:
+            # AccVideo lora has diff bias keys but not the weight keys. This causes a weird problem where
+            # our lora config adds the time proj lora layers, but we don't have the weights for them.
+            # CausVid lora has the weight keys and the bias keys.
+            original_state_dict.pop(key)
 
     # For the `diff_b` keys, we treat them as lora_bias.
     # https://huggingface.co/docs/peft/main/en/package_reference/lora#peft.LoraConfig.lora_bias
@@ -1628,15 +1863,26 @@ def _convert_non_diffusers_wan_lora_to_diffusers(state_dict):
     for i in range(min_block, max_block + 1):
         # Self-attention
         for o, c in zip(["q", "k", "v", "o"], ["to_q", "to_k", "to_v", "to_out.0"]):
-            original_key = f"blocks.{i}.self_attn.{o}.{lora_down_key}.weight"
-            converted_key = f"blocks.{i}.attn1.{c}.lora_A.weight"
-            if original_key in original_state_dict:
-                converted_state_dict[converted_key] = original_state_dict.pop(original_key)
+            alpha_key = f"blocks.{i}.self_attn.{o}.alpha"
+            has_alpha = alpha_key in original_state_dict
+            original_key_A = f"blocks.{i}.self_attn.{o}.{lora_down_key}.weight"
+            converted_key_A = f"blocks.{i}.attn1.{c}.lora_A.weight"
 
-            original_key = f"blocks.{i}.self_attn.{o}.{lora_up_key}.weight"
-            converted_key = f"blocks.{i}.attn1.{c}.lora_B.weight"
-            if original_key in original_state_dict:
-                converted_state_dict[converted_key] = original_state_dict.pop(original_key)
+            original_key_B = f"blocks.{i}.self_attn.{o}.{lora_up_key}.weight"
+            converted_key_B = f"blocks.{i}.attn1.{c}.lora_B.weight"
+
+            if has_alpha:
+                down_weight = original_state_dict.pop(original_key_A)
+                up_weight = original_state_dict.pop(original_key_B)
+                scale_down, scale_up = get_alpha_scales(down_weight, alpha_key)
+                converted_state_dict[converted_key_A] = down_weight * scale_down
+                converted_state_dict[converted_key_B] = up_weight * scale_up
+
+            else:
+                if original_key_A in original_state_dict:
+                    converted_state_dict[converted_key_A] = original_state_dict.pop(original_key_A)
+                if original_key_B in original_state_dict:
+                    converted_state_dict[converted_key_B] = original_state_dict.pop(original_key_B)
 
             original_key = f"blocks.{i}.self_attn.{o}.diff_b"
             converted_key = f"blocks.{i}.attn1.{c}.lora_B.bias"
@@ -1645,15 +1891,24 @@ def _convert_non_diffusers_wan_lora_to_diffusers(state_dict):
 
         # Cross-attention
         for o, c in zip(["q", "k", "v", "o"], ["to_q", "to_k", "to_v", "to_out.0"]):
-            original_key = f"blocks.{i}.cross_attn.{o}.{lora_down_key}.weight"
-            converted_key = f"blocks.{i}.attn2.{c}.lora_A.weight"
-            if original_key in original_state_dict:
-                converted_state_dict[converted_key] = original_state_dict.pop(original_key)
-
-            original_key = f"blocks.{i}.cross_attn.{o}.{lora_up_key}.weight"
-            converted_key = f"blocks.{i}.attn2.{c}.lora_B.weight"
-            if original_key in original_state_dict:
-                converted_state_dict[converted_key] = original_state_dict.pop(original_key)
+            alpha_key = f"blocks.{i}.cross_attn.{o}.alpha"
+            has_alpha = alpha_key in original_state_dict
+            original_key_A = f"blocks.{i}.cross_attn.{o}.{lora_down_key}.weight"
+            converted_key_A = f"blocks.{i}.attn2.{c}.lora_A.weight"
+
+            original_key_B = f"blocks.{i}.cross_attn.{o}.{lora_up_key}.weight"
+            converted_key_B = f"blocks.{i}.attn2.{c}.lora_B.weight"
+
+            if original_key_A in original_state_dict:
+                down_weight = original_state_dict.pop(original_key_A)
+                converted_state_dict[converted_key_A] = down_weight
+            if original_key_B in original_state_dict:
+                up_weight = original_state_dict.pop(original_key_B)
+                converted_state_dict[converted_key_B] = up_weight
+            if has_alpha:
+                scale_down, scale_up = get_alpha_scales(down_weight, alpha_key)
+                converted_state_dict[converted_key_A] *= scale_down
+                converted_state_dict[converted_key_B] *= scale_up
 
             original_key = f"blocks.{i}.cross_attn.{o}.diff_b"
             converted_key = f"blocks.{i}.attn2.{c}.lora_B.bias"
@@ -1662,15 +1917,24 @@ def _convert_non_diffusers_wan_lora_to_diffusers(state_dict):
 
         if is_i2v_lora:
             for o, c in zip(["k_img", "v_img"], ["add_k_proj", "add_v_proj"]):
-                original_key = f"blocks.{i}.cross_attn.{o}.{lora_down_key}.weight"
-                converted_key = f"blocks.{i}.attn2.{c}.lora_A.weight"
-                if original_key in original_state_dict:
-                    converted_state_dict[converted_key] = original_state_dict.pop(original_key)
-
-                original_key = f"blocks.{i}.cross_attn.{o}.{lora_up_key}.weight"
-                converted_key = f"blocks.{i}.attn2.{c}.lora_B.weight"
-                if original_key in original_state_dict:
-                    converted_state_dict[converted_key] = original_state_dict.pop(original_key)
+                alpha_key = f"blocks.{i}.cross_attn.{o}.alpha"
+                has_alpha = alpha_key in original_state_dict
+                original_key_A = f"blocks.{i}.cross_attn.{o}.{lora_down_key}.weight"
+                converted_key_A = f"blocks.{i}.attn2.{c}.lora_A.weight"
+
+                original_key_B = f"blocks.{i}.cross_attn.{o}.{lora_up_key}.weight"
+                converted_key_B = f"blocks.{i}.attn2.{c}.lora_B.weight"
+
+                if original_key_A in original_state_dict:
+                    down_weight = original_state_dict.pop(original_key_A)
+                    converted_state_dict[converted_key_A] = down_weight
+                if original_key_B in original_state_dict:
+                    up_weight = original_state_dict.pop(original_key_B)
+                    converted_state_dict[converted_key_B] = up_weight
+                if has_alpha:
+                    scale_down, scale_up = get_alpha_scales(down_weight, alpha_key)
+                    converted_state_dict[converted_key_A] *= scale_down
+                    converted_state_dict[converted_key_B] *= scale_up
 
                 original_key = f"blocks.{i}.cross_attn.{o}.diff_b"
                 converted_key = f"blocks.{i}.attn2.{c}.lora_B.bias"
@@ -1679,15 +1943,24 @@ def _convert_non_diffusers_wan_lora_to_diffusers(state_dict):
 
         # FFN
         for o, c in zip(["ffn.0", "ffn.2"], ["net.0.proj", "net.2"]):
-            original_key = f"blocks.{i}.{o}.{lora_down_key}.weight"
-            converted_key = f"blocks.{i}.ffn.{c}.lora_A.weight"
-            if original_key in original_state_dict:
-                converted_state_dict[converted_key] = original_state_dict.pop(original_key)
-
-            original_key = f"blocks.{i}.{o}.{lora_up_key}.weight"
-            converted_key = f"blocks.{i}.ffn.{c}.lora_B.weight"
-            if original_key in original_state_dict:
-                converted_state_dict[converted_key] = original_state_dict.pop(original_key)
+            alpha_key = f"blocks.{i}.{o}.alpha"
+            has_alpha = alpha_key in original_state_dict
+            original_key_A = f"blocks.{i}.{o}.{lora_down_key}.weight"
+            converted_key_A = f"blocks.{i}.ffn.{c}.lora_A.weight"
+
+            original_key_B = f"blocks.{i}.{o}.{lora_up_key}.weight"
+            converted_key_B = f"blocks.{i}.ffn.{c}.lora_B.weight"
+
+            if original_key_A in original_state_dict:
+                down_weight = original_state_dict.pop(original_key_A)
+                converted_state_dict[converted_key_A] = down_weight
+            if original_key_B in original_state_dict:
+                up_weight = original_state_dict.pop(original_key_B)
+                converted_state_dict[converted_key_B] = up_weight
+            if has_alpha:
+                scale_down, scale_up = get_alpha_scales(down_weight, alpha_key)
+                converted_state_dict[converted_key_A] *= scale_down
+                converted_state_dict[converted_key_B] *= scale_up
 
             original_key = f"blocks.{i}.{o}.diff_b"
             converted_key = f"blocks.{i}.ffn.{c}.lora_B.bias"
@@ -1754,6 +2027,10 @@ def _convert_non_diffusers_wan_lora_to_diffusers(state_dict):
             converted_key = f"condition_embedder.image_embedder.{img_ours}.lora_B.weight"
             if original_key in original_state_dict:
                 converted_state_dict[converted_key] = original_state_dict.pop(original_key)
+                bias_key_theirs = original_key.removesuffix(f".{lora_up_key}.weight") + ".diff_b"
+                if bias_key_theirs in original_state_dict:
+                    bias_key = converted_key.removesuffix(".weight") + ".bias"
+                    converted_state_dict[bias_key] = original_state_dict.pop(bias_key_theirs)
 
     if len(original_state_dict) > 0:
         diff = all(".diff" in k for k in original_state_dict)
@@ -1849,3 +2126,107 @@ def _convert_non_diffusers_ltxv_lora_to_diffusers(state_dict, non_diffusers_pref
     converted_state_dict = {k.removeprefix(f"{non_diffusers_prefix}."): v for k, v in state_dict.items()}
     converted_state_dict = {f"transformer.{k}": v for k, v in converted_state_dict.items()}
     return converted_state_dict
+
+
+def _convert_non_diffusers_qwen_lora_to_diffusers(state_dict):
+    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()}
+
+        def convert_key(key: str) -> str:
+            prefix = "transformer_blocks"
+            if "." in key:
+                base, suffix = key.rsplit(".", 1)
+            else:
+                base, suffix = key, ""
+
+            start = f"{prefix}_"
+            rest = base[len(start) :]
+
+            if "." in rest:
+                head, tail = rest.split(".", 1)
+                tail = "." + tail
+            else:
+                head, tail = rest, ""
+
+            # Protected n-grams that must keep their internal underscores
+            protected = {
+                # pairs
+                ("to", "q"),
+                ("to", "k"),
+                ("to", "v"),
+                ("to", "out"),
+                ("add", "q"),
+                ("add", "k"),
+                ("add", "v"),
+                ("txt", "mlp"),
+                ("img", "mlp"),
+                ("txt", "mod"),
+                ("img", "mod"),
+                # triplets
+                ("add", "q", "proj"),
+                ("add", "k", "proj"),
+                ("add", "v", "proj"),
+                ("to", "add", "out"),
+            }
+
+            prot_by_len = {}
+            for ng in protected:
+                prot_by_len.setdefault(len(ng), set()).add(ng)
+
+            parts = head.split("_")
+            merged = []
+            i = 0
+            lengths_desc = sorted(prot_by_len.keys(), reverse=True)
+
+            while i < len(parts):
+                matched = False
+                for L in lengths_desc:
+                    if i + L <= len(parts) and tuple(parts[i : i + L]) in prot_by_len[L]:
+                        merged.append("_".join(parts[i : i + L]))
+                        i += L
+                        matched = True
+                        break
+                if not matched:
+                    merged.append(parts[i])
+                    i += 1
+
+            head_converted = ".".join(merged)
+            converted_base = f"{prefix}.{head_converted}{tail}"
+            return converted_base + (("." + suffix) if suffix else "")
+
+        state_dict = {convert_key(k): v for k, v in state_dict.items()}
+
+    converted_state_dict = {}
+    all_keys = list(state_dict.keys())
+    down_key = ".lora_down.weight"
+    up_key = ".lora_up.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
+
+    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 len(state_dict) > 0:
+        raise ValueError(f"`state_dict` should be empty at this point but has {state_dict.keys()=}")
+
+    converted_state_dict = {f"transformer.{k}": v for k, v in converted_state_dict.items()}
+    return converted_state_dict