Initial support for qwen image model. (Comfy-Org#9179)

Author: comfyanonymous

comfy/ldm/qwen_image/model.py

@@ -42,6 +42,7 @@
 import comfy.ldm.chroma.model
 import comfy.ldm.ace.model
 import comfy.ldm.omnigen.omnigen2
+import comfy.ldm.qwen_image.model
 
 import comfy.model_management
 import comfy.patcher_extension
@@ -1308,3 +1309,14 @@ def extra_conds_shapes(self, **kwargs):
         if ref_latents is not None:
             out['ref_latents'] = list([1, 16, sum(map(lambda a: math.prod(a.size()), ref_latents)) // 16])
         return out
+
+class QwenImage(BaseModel):
+    def __init__(self, model_config, model_type=ModelType.FLUX, device=None):
+        super().__init__(model_config, model_type, device=device, unet_model=comfy.ldm.qwen_image.model.QwenImageTransformer2DModel)
+
+    def extra_conds(self, **kwargs):
+        out = super().extra_conds(**kwargs)
+        cross_attn = kwargs.get("cross_attn", None)
+        if cross_attn is not None:
+            out['c_crossattn'] = comfy.conds.CONDRegular(cross_attn)
+        return out

comfy/model_base.py

@@ -481,6 +481,11 @@ def detect_unet_config(state_dict, key_prefix, metadata=None):
         dit_config["timestep_scale"] = 1000.0
         return dit_config
 
+    if '{}txt_norm.weight'.format(key_prefix) in state_dict_keys:  # Qwen Image
+        dit_config = {}
+        dit_config["image_model"] = "qwen_image"
+        return dit_config
+
     if '{}input_blocks.0.0.weight'.format(key_prefix) not in state_dict_keys:
         return None
 
@@ -867,7 +872,7 @@ def convert_diffusers_mmdit(state_dict, output_prefix=""):
         depth_single_blocks = count_blocks(state_dict, 'single_transformer_blocks.{}.')
         hidden_size = state_dict["x_embedder.bias"].shape[0]
         sd_map = comfy.utils.flux_to_diffusers({"depth": depth, "depth_single_blocks": depth_single_blocks, "hidden_size": hidden_size}, output_prefix=output_prefix)
-    elif 'transformer_blocks.0.attn.add_q_proj.weight' in state_dict: #SD3
+    elif 'transformer_blocks.0.attn.add_q_proj.weight' in state_dict and 'pos_embed.proj.weight' in state_dict: #SD3
         num_blocks = count_blocks(state_dict, 'transformer_blocks.{}.')
         depth = state_dict["pos_embed.proj.weight"].shape[0] // 64
         sd_map = comfy.utils.mmdit_to_diffusers({"depth": depth, "num_blocks": num_blocks}, output_prefix=output_prefix)

comfy/model_detection.py

@@ -47,6 +47,7 @@
 import comfy.text_encoders.hidream
 import comfy.text_encoders.ace
 import comfy.text_encoders.omnigen2
+import comfy.text_encoders.qwen_image
 
 import comfy.model_patcher
 import comfy.lora
@@ -771,6 +772,7 @@ class CLIPType(Enum):
     CHROMA = 15
     ACE = 16
     OMNIGEN2 = 17
+    QWEN_IMAGE = 18
 
 
 def load_clip(ckpt_paths, embedding_directory=None, clip_type=CLIPType.STABLE_DIFFUSION, model_options={}):
@@ -791,6 +793,7 @@ class TEModel(Enum):
     T5_XXL_OLD = 8
     GEMMA_2_2B = 9
     QWEN25_3B = 10
+    QWEN25_7B = 11
 
 def detect_te_model(sd):
     if "text_model.encoder.layers.30.mlp.fc1.weight" in sd:
@@ -812,7 +815,11 @@ def detect_te_model(sd):
     if 'model.layers.0.post_feedforward_layernorm.weight' in sd:
         return TEModel.GEMMA_2_2B
     if 'model.layers.0.self_attn.k_proj.bias' in sd:
-        return TEModel.QWEN25_3B
+        weight = sd['model.layers.0.self_attn.k_proj.bias']
+        if weight.shape[0] == 256:
+            return TEModel.QWEN25_3B
+        if weight.shape[0] == 512:
+            return TEModel.QWEN25_7B
     if "model.layers.0.post_attention_layernorm.weight" in sd:
         return TEModel.LLAMA3_8
     return None
@@ -917,6 +924,9 @@ class EmptyClass:
         elif te_model == TEModel.QWEN25_3B:
             clip_target.clip = comfy.text_encoders.omnigen2.te(**llama_detect(clip_data))
             clip_target.tokenizer = comfy.text_encoders.omnigen2.Omnigen2Tokenizer
+        elif te_model == TEModel.QWEN25_7B:
+            clip_target.clip = comfy.text_encoders.qwen_image.te(**llama_detect(clip_data))
+            clip_target.tokenizer = comfy.text_encoders.qwen_image.QwenImageTokenizer
         else:
             # clip_l
             if clip_type == CLIPType.SD3:

comfy/sd.py

@@ -19,6 +19,7 @@
 import comfy.text_encoders.wan
 import comfy.text_encoders.ace
 import comfy.text_encoders.omnigen2
+import comfy.text_encoders.qwen_image
 
 from . import supported_models_base
 from . import latent_formats
@@ -1229,7 +1230,36 @@ def clip_target(self, state_dict={}):
         hunyuan_detect = comfy.text_encoders.hunyuan_video.llama_detect(state_dict, "{}qwen25_3b.transformer.".format(pref))
         return supported_models_base.ClipTarget(comfy.text_encoders.omnigen2.Omnigen2Tokenizer, comfy.text_encoders.omnigen2.te(**hunyuan_detect))
 
+class QwenImage(supported_models_base.BASE):
+    unet_config = {
+        "image_model": "qwen_image",
+    }
+
+    sampling_settings = {
+        "multiplier": 1.0,
+        "shift": 2.6,
+    }
+
+    memory_usage_factor = 1.8 #TODO
+
+    unet_extra_config = {}
+    latent_format = latent_formats.Wan21
+
+    supported_inference_dtypes = [torch.bfloat16, torch.float32]
+
+    vae_key_prefix = ["vae."]
+    text_encoder_key_prefix = ["text_encoders."]
+
+    def get_model(self, state_dict, prefix="", device=None):
+        out = model_base.QwenImage(self, device=device)
+        return out
+
+    def clip_target(self, state_dict={}):
+        pref = self.text_encoder_key_prefix[0]
+        hunyuan_detect = comfy.text_encoders.hunyuan_video.llama_detect(state_dict, "{}qwen25_7b.transformer.".format(pref))
+        return supported_models_base.ClipTarget(comfy.text_encoders.qwen_image.QwenImageTokenizer, comfy.text_encoders.qwen_image.te(**hunyuan_detect))
+
 
-models = [LotusD, Stable_Zero123, SD15_instructpix2pix, SD15, SD20, SD21UnclipL, SD21UnclipH, SDXL_instructpix2pix, SDXLRefiner, SDXL, SSD1B, KOALA_700M, KOALA_1B, Segmind_Vega, SD_X4Upscaler, Stable_Cascade_C, Stable_Cascade_B, SV3D_u, SV3D_p, SD3, StableAudio, AuraFlow, PixArtAlpha, PixArtSigma, HunyuanDiT, HunyuanDiT1, FluxInpaint, Flux, FluxSchnell, GenmoMochi, LTXV, HunyuanVideoSkyreelsI2V, HunyuanVideoI2V, HunyuanVideo, CosmosT2V, CosmosI2V, CosmosT2IPredict2, CosmosI2VPredict2, Lumina2, WAN22_T2V, WAN21_T2V, WAN21_I2V, WAN21_FunControl2V, WAN21_Vace, WAN21_Camera, Hunyuan3Dv2mini, Hunyuan3Dv2, HiDream, Chroma, ACEStep, Omnigen2]
+models = [LotusD, Stable_Zero123, SD15_instructpix2pix, SD15, SD20, SD21UnclipL, SD21UnclipH, SDXL_instructpix2pix, SDXLRefiner, SDXL, SSD1B, KOALA_700M, KOALA_1B, Segmind_Vega, SD_X4Upscaler, Stable_Cascade_C, Stable_Cascade_B, SV3D_u, SV3D_p, SD3, StableAudio, AuraFlow, PixArtAlpha, PixArtSigma, HunyuanDiT, HunyuanDiT1, FluxInpaint, Flux, FluxSchnell, GenmoMochi, LTXV, HunyuanVideoSkyreelsI2V, HunyuanVideoI2V, HunyuanVideo, CosmosT2V, CosmosI2V, CosmosT2IPredict2, CosmosI2VPredict2, Lumina2, WAN22_T2V, WAN21_T2V, WAN21_I2V, WAN21_FunControl2V, WAN21_Vace, WAN21_Camera, Hunyuan3Dv2mini, Hunyuan3Dv2, HiDream, Chroma, ACEStep, Omnigen2, QwenImage]
 
 models += [SVD_img2vid]

comfy/supported_models.py

@@ -43,6 +43,23 @@ class Qwen25_3BConfig:
     mlp_activation = "silu"
     qkv_bias = True
 
+@dataclass
+class Qwen25_7BVLI_Config:
+    vocab_size: int = 152064
+    hidden_size: int = 3584
+    intermediate_size: int = 18944
+    num_hidden_layers: int = 28
+    num_attention_heads: int = 28
+    num_key_value_heads: int = 4
+    max_position_embeddings: int = 128000
+    rms_norm_eps: float = 1e-6
+    rope_theta: float = 1000000.0
+    transformer_type: str = "llama"
+    head_dim = 128
+    rms_norm_add = False
+    mlp_activation = "silu"
+    qkv_bias = True
+
 @dataclass
 class Gemma2_2B_Config:
     vocab_size: int = 256000
@@ -348,6 +365,15 @@ def __init__(self, config_dict, dtype, device, operations):
         self.model = Llama2_(config, device=device, dtype=dtype, ops=operations)
         self.dtype = dtype
 
+class Qwen25_7BVLI(BaseLlama, torch.nn.Module):
+    def __init__(self, config_dict, dtype, device, operations):
+        super().__init__()
+        config = Qwen25_7BVLI_Config(**config_dict)
+        self.num_layers = config.num_hidden_layers
+
+        self.model = Llama2_(config, device=device, dtype=dtype, ops=operations)
+        self.dtype = dtype
+
 class Gemma2_2B(BaseLlama, torch.nn.Module):
     def __init__(self, config_dict, dtype, device, operations):
         super().__init__()

comfy/text_encoders/llama.py

@@ -0,0 +1,71 @@
+from transformers import Qwen2Tokenizer
+from comfy import sd1_clip
+import comfy.text_encoders.llama
+import os
+import torch
+import numbers
+
+class Qwen25_7BVLITokenizer(sd1_clip.SDTokenizer):
+    def __init__(self, embedding_directory=None, tokenizer_data={}):
+        tokenizer_path = os.path.join(os.path.dirname(os.path.realpath(__file__)), "qwen25_tokenizer")
+        super().__init__(tokenizer_path, pad_with_end=False, embedding_size=3584, embedding_key='qwen25_7b', tokenizer_class=Qwen2Tokenizer, has_start_token=False, has_end_token=False, pad_to_max_length=False, max_length=99999999, min_length=1, pad_token=151643, tokenizer_data=tokenizer_data)
+
+
+class QwenImageTokenizer(sd1_clip.SD1Tokenizer):
+    def __init__(self, embedding_directory=None, tokenizer_data={}):
+        super().__init__(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data, name="qwen25_7b", tokenizer=Qwen25_7BVLITokenizer)
+        self.llama_template = "<|im_start|>system\nDescribe the image by detailing the color, shape, size, texture, quantity, text, spatial relationships of the objects and background:<|im_end|>\n<|im_start|>user\n{}<|im_end|>\n<|im_start|>assistant\n"
+
+    def tokenize_with_weights(self, text, return_word_ids=False, llama_template=None,**kwargs):
+        if llama_template is None:
+            llama_text = self.llama_template.format(text)
+        else:
+            llama_text = llama_template.format(text)
+        return super().tokenize_with_weights(llama_text, return_word_ids=return_word_ids, **kwargs)
+
+
+class Qwen25_7BVLIModel(sd1_clip.SDClipModel):
+    def __init__(self, device="cpu", layer="last", layer_idx=None, dtype=None, attention_mask=True, model_options={}):
+        super().__init__(device=device, layer=layer, layer_idx=layer_idx, textmodel_json_config={}, dtype=dtype, special_tokens={"pad": 151643}, layer_norm_hidden_state=False, model_class=comfy.text_encoders.llama.Qwen25_7BVLI, enable_attention_masks=attention_mask, return_attention_masks=attention_mask, model_options=model_options)
+
+
+class QwenImageTEModel(sd1_clip.SD1ClipModel):
+    def __init__(self, device="cpu", dtype=None, model_options={}):
+        super().__init__(device=device, dtype=dtype, name="qwen25_7b", clip_model=Qwen25_7BVLIModel, model_options=model_options)
+
+    def encode_token_weights(self, token_weight_pairs):
+        out, pooled, extra = super().encode_token_weights(token_weight_pairs)
+        tok_pairs = token_weight_pairs["qwen25_7b"][0]
+        count_im_start = 0
+        for i, v in enumerate(tok_pairs):
+            elem = v[0]
+            if not torch.is_tensor(elem):
+                if isinstance(elem, numbers.Integral):
+                    if elem == 151644 and count_im_start < 2:
+                        template_end = i
+                        count_im_start += 1
+
+        if out.shape[1] > (template_end + 3):
+            if tok_pairs[template_end + 1][0] == 872:
+                if tok_pairs[template_end + 2][0] == 198:
+                    template_end += 3
+
+        out = out[:, template_end:]
+
+        extra["attention_mask"] = extra["attention_mask"][:, template_end:]
+        if extra["attention_mask"].sum() == torch.numel(extra["attention_mask"]):
+            extra.pop("attention_mask")  # attention mask is useless if no masked elements
+
+        return out, pooled, extra
+
+
+def te(dtype_llama=None, llama_scaled_fp8=None):
+    class QwenImageTEModel_(QwenImageTEModel):
+        def __init__(self, device="cpu", dtype=None, model_options={}):
+            if llama_scaled_fp8 is not None and "scaled_fp8" not in model_options:
+                model_options = model_options.copy()
+                model_options["scaled_fp8"] = llama_scaled_fp8
+            if dtype_llama is not None:
+                dtype = dtype_llama
+            super().__init__(device=device, dtype=dtype, model_options=model_options)
+    return QwenImageTEModel_

comfy/text_encoders/qwen_image.py

@@ -925,7 +925,7 @@ class CLIPLoader:
     @classmethod
     def INPUT_TYPES(s):
         return {"required": { "clip_name": (folder_paths.get_filename_list("text_encoders"), ),
-                              "type": (["stable_diffusion", "stable_cascade", "sd3", "stable_audio", "mochi", "ltxv", "pixart", "cosmos", "lumina2", "wan", "hidream", "chroma", "ace", "omnigen2"], ),
+                              "type": (["stable_diffusion", "stable_cascade", "sd3", "stable_audio", "mochi", "ltxv", "pixart", "cosmos", "lumina2", "wan", "hidream", "chroma", "ace", "omnigen2", "qwen_image"], ),
                               },
                 "optional": {
                               "device": (["default", "cpu"], {"advanced": True}),