Update app.py

app.py

@@ -5,7 +5,6 @@ import spaces
 import torch
 import re
 from diffusers import DiffusionPipeline
-from compel import Compel, ReturnedEmbeddingsType
 from PIL import Image, PngImagePlugin
 import json
 import io
@@ -34,14 +33,9 @@ def add_comma_after_pattern_ti(text):
     modified_text = pattern.sub(lambda x: x.group() + ',', text)
     return modified_text
 
-def get_embed_new(prompt, pipe, compel, only_convert_string=False):
-    """处理提示词的函数"""
-    if only_convert_string:
-        # 简单处理，添加逗号分隔
-        return add_comma_after_pattern_ti(prompt)
-    else:
-        # 如果需要更复杂的处理，可以在这里添加
-        return prompt
+def process_prompt(prompt):
+    """简单的提示词处理函数"""
+    return add_comma_after_pattern_ti(prompt)
 
 DESCRIPTION = "梦羽的模型生成器 - 快速生成 MiaomiaoHarem vPred Dogma 1.1 模型的图片"
 
@@ -53,7 +47,12 @@ MAX_IMAGE_SIZE = 2048
 
 if torch.cuda.is_available():
     dtype = torch.bfloat16
-    pipe = DiffusionPipeline.from_pretrained("Qwen/Qwen-Image", torch_dtype=dtype).to("cuda")
+    device = "cuda"
+else:
+    dtype = torch.float32
+    device = "cpu"
+
+pipe = DiffusionPipeline.from_pretrained("Qwen/Qwen-Image", torch_dtype=dtype).to(device)
 
 def randomize_seed_fn(seed: int, randomize_seed: bool) -> int:
     if randomize_seed:
@@ -68,42 +67,40 @@ def infer(
     seed: int = 7,
     width: int = 1024,
     height: int = 1536,
-    guidance_scale: float = 3,
-    num_inference_steps: int = 30,
+    guidance_scale: float = 4.0,
+    num_inference_steps: int = 50,
     randomize_seed: bool = True,
-    use_resolution_binning: bool = True,
 ):
     seed = int(randomize_seed_fn(seed, randomize_seed))
-    generator = torch.Generator().manual_seed(seed)
-    # 初始化 Compel 实例
-    compel = Compel(
-        tokenizer=[pipe.tokenizer, pipe.tokenizer_2],
-        text_encoder=[pipe.text_encoder, pipe.text_encoder_2],
-        returned_embeddings_type=ReturnedEmbeddingsType.PENULTIMATE_HIDDEN_STATES_NON_NORMALIZED,
-        requires_pooled=[False, True],
-        truncate_long_prompts=False
-    )
-    # 在 infer 函数中调用 get_embed_new
+    generator = torch.Generator(device=device).manual_seed(seed)
+
     if not use_negative_prompt:
         negative_prompt = ""
-    
+
     original_prompt = prompt  # Store original prompt for metadata
-    prompt = get_embed_new(prompt, pipe, compel, only_convert_string=True)
-    negative_prompt = get_embed_new(negative_prompt, pipe, compel, only_convert_string=True)
-    conditioning, pooled = compel([prompt, negative_prompt]) # 必须同时处理来保证长度相等
+
+    # 处理提示词
+    prompt = process_prompt(prompt)
+
+    # 为 Qwen-Image 添加质量提升词条
+    positive_magic_en = "Ultra HD, 4K, cinematic composition."
+    positive_magic_zh = "超清，4K，电影级构图"
+
+    # 简单判断是否包含中文字符
+    if any('\u4e00' <= char <= '\u9fff' for char in prompt):
+        prompt = prompt + " " + positive_magic_zh
+    else:
+        prompt = prompt + " " + positive_magic_en
     
-    # 在调用 pipe 时，使用新的参数名称（确保参数名称正确）
+    # 使用 Qwen-Image 的参数格式
     image = pipe(
-        prompt_embeds=conditioning[0:1],
-        pooled_prompt_embeds=pooled[0:1],
-        negative_prompt_embeds=conditioning[1:2],
-        negative_pooled_prompt_embeds=pooled[1:2],
+        prompt=prompt,
+        negative_prompt=negative_prompt,
         width=width,
         height=height,
-        guidance_scale=guidance_scale,
+        true_cfg_scale=guidance_scale,
         num_inference_steps=num_inference_steps,
         generator=generator,
-        use_resolution_binning=use_resolution_binning,
     ).images[0]
 
     # Create metadata dictionary
@@ -114,11 +111,10 @@ def infer(
         "seed": seed,
         "width": width,
         "height": height,
-        "guidance_scale": guidance_scale,
+        "true_cfg_scale": guidance_scale,
         "num_inference_steps": num_inference_steps,
         "model": "qwen-image",
-        "use_resolution_binning": use_resolution_binning,
-        "PreUrl": "https://huggingface.co/spaces/Menyu/miaomiaoHaremDogma11"
+        "PreUrl": "https://huggingface.co/Qwen/Qwen-Image"
     }
     # Add metadata to the image
     image_with_metadata = add_metadata_to_image(image, metadata)
@@ -189,20 +185,19 @@ with gr.Blocks(css=css) as demo:
             )
         with gr.Row():
             guidance_scale = gr.Slider(
-                label="Guidance Scale",
-                minimum=0.1,
-                maximum=10,
+                label="True CFG Scale",
+                minimum=1.0,
+                maximum=10.0,
                 step=0.1,
-                value=7.0,
+                value=4.0,
             )
             num_inference_steps = gr.Slider(
                 label="生成步数",
                 minimum=1,
-                maximum=50,
+                maximum=100,
                 step=1,
-                value=28,
+                value=50,
             )
-        use_resolution_binning = gr.Checkbox(label="使用分辨率分箱", value=True)
 
     gr.Examples(
         examples=examples,
@@ -230,7 +225,6 @@ with gr.Blocks(css=css) as demo:
             guidance_scale,
             num_inference_steps,
             randomize_seed,
-            use_resolution_binning,
         ],
         outputs=[result, seed],
     )