detection

modules_forge/packages/huggingface_guess/detection.py

@@ -118,6 +118,14 @@ def detect_unet_config(state_dict: dict, key_prefix: str):
             dit_config["guidance_embed"] = "{}guidance_in.in_layer.weight".format(key_prefix) in state_dict_keys
         return dit_config
 
+    if "{}txt_norm.weight".format(key_prefix) in state_dict_keys:  # Qwen Image
+        _qweight: bool = "{}transformer_blocks.0.attn.to_qkv.qweight".format(key_prefix) in state_dict_keys
+        dit_config = {"nunchaku": _qweight}
+        dit_config["image_model"] = "qwen_image"
+        dit_config["in_channels"] = state_dict["{}img_in.weight".format(key_prefix)].shape[1]
+        dit_config["num_layers"] = count_blocks(state_dict_keys, "{}transformer_blocks.".format(key_prefix) + "{}.")
+        return dit_config
+
     if "{}input_blocks.0.0.weight".format(key_prefix) not in state_dict_keys:
         return None
 

modules_forge/packages/huggingface_guess/model_list.py

@@ -375,6 +375,38 @@ class WAN21_I2V(WAN21_T2V):
     }
 
 
+class QwenImage(BASE):
+    huggingface_repo = "Qwen/Qwen-Image"
+
+    unet_config = {
+        "image_model": "qwen_image",
+    }
+
+    sampling_settings = {
+        "multiplier": 1.0,
+        "shift": 1.15,
+    }
+
+    memory_usage_factor = 1.8
+
+    unet_extra_config = {}
+    latent_format = latent.Wan21
+
+    supported_inference_dtypes = [torch.bfloat16, torch.float32]
+
+    vae_key_prefix = ["vae."]
+    text_encoder_key_prefix = ["text_encoders."]
+
+    unet_target = "transformer"
+
+    def __init__(self, unet_config):
+        super().__init__(unet_config)
+        self.nunchaku: bool = self.unet_config.pop("nunchaku", False)
+
+    def clip_target(self, state_dict={}):
+        return {"qwen25": "text_encoder"}
+
+
 models = [
     SD15,
     SDXL,
@@ -384,4 +416,5 @@ models = [
     Chroma,
     WAN21_T2V,
     WAN21_I2V,
+    QwenImage,
 ]