Merge pull request #921 from modelscope/qwen-image-distill-dmd2-lora

support qwen-image-distill-dmd2-lora

diffsynth/pipelines/qwen_image.py

@@ -371,6 +371,7 @@ class QwenImagePipeline(BasePipeline):
         rand_device: str = "cpu",
         # Steps
         num_inference_steps: int = 30,
+        exponential_shift_mu: float = None,
         # Blockwise ControlNet
         blockwise_controlnet_inputs: list[ControlNetInput] = None,
         # EliGen
@@ -393,7 +394,7 @@ class QwenImagePipeline(BasePipeline):
         progress_bar_cmd = tqdm,
     ):
         # Scheduler
-        self.scheduler.set_timesteps(num_inference_steps, denoising_strength=denoising_strength, dynamic_shift_len=(height // 16) * (width // 16))
+        self.scheduler.set_timesteps(num_inference_steps, denoising_strength=denoising_strength, dynamic_shift_len=(height // 16) * (width // 16), exponential_shift_mu=exponential_shift_mu)
         
         # Parameters
         inputs_posi = {

diffsynth/schedulers/flow_match.py

@@ -31,7 +31,7 @@ class FlowMatchScheduler():
         self.set_timesteps(num_inference_steps)
 
 
-    def set_timesteps(self, num_inference_steps=100, denoising_strength=1.0, training=False, shift=None, dynamic_shift_len=None):
+    def set_timesteps(self, num_inference_steps=100, denoising_strength=1.0, training=False, shift=None, dynamic_shift_len=None, exponential_shift_mu=None):
         if shift is not None:
             self.shift = shift
         sigma_start = self.sigma_min + (self.sigma_max - self.sigma_min) * denoising_strength
@@ -42,7 +42,12 @@ class FlowMatchScheduler():
         if self.inverse_timesteps:
             self.sigmas = torch.flip(self.sigmas, dims=[0])
         if self.exponential_shift:
-            mu = self.calculate_shift(dynamic_shift_len) if dynamic_shift_len is not None else self.exponential_shift_mu
+            if exponential_shift_mu is not None:
+                mu = exponential_shift_mu
+            elif dynamic_shift_len is not None:
+                mu = self.calculate_shift(dynamic_shift_len)
+            else:
+                mu = self.exponential_shift_mu
             self.sigmas = math.exp(mu) / (math.exp(mu) + (1 / self.sigmas - 1))
         else:
             self.sigmas = self.shift * self.sigmas / (1 + (self.shift - 1) * self.sigmas)

examples/qwen_image/model_inference/Qwen-Image-Distill-DMD2.py

@@ -0,0 +1,24 @@
+from diffsynth.pipelines.qwen_image import QwenImagePipeline, ModelConfig, load_state_dict
+from modelscope import snapshot_download
+import torch, math
+
+
+pipe = QwenImagePipeline.from_pretrained(
+    torch_dtype=torch.bfloat16,
+    device="cuda",
+    model_configs=[
+        ModelConfig(model_id="Qwen/Qwen-Image", origin_file_pattern="transformer/diffusion_pytorch_model*.safetensors"),
+        ModelConfig(model_id="Qwen/Qwen-Image", origin_file_pattern="text_encoder/model*.safetensors"),
+        ModelConfig(model_id="Qwen/Qwen-Image", origin_file_pattern="vae/diffusion_pytorch_model.safetensors"),
+    ],
+    tokenizer_config=ModelConfig(model_id="Qwen/Qwen-Image", origin_file_pattern="tokenizer/"),
+)
+
+snapshot_download("MusePublic/Qwen-Image-Distill", allow_file_pattern="qwen_image_distill_3step.safetensors", cache_dir="models")
+lora_state_dict = load_state_dict("models/MusePublic/Qwen-Image-Distill/qwen_image_distill_3step.safetensors")
+lora_state_dict = {i.replace("base_model.model.", ""): j for i, j in lora_state_dict.items()}
+pipe.load_lora(pipe.dit, state_dict=lora_state_dict)
+
+prompt = "精致肖像，水下少女，蓝裙飘逸，发丝轻扬，光影透澈，气泡环绕，面容恬静，细节精致，梦幻唯美。"
+image = pipe(prompt, seed=0, num_inference_steps=3, cfg_scale=1, exponential_shift_mu=math.log(2.5))
+image.save("image.jpg")