Update app.py

app.py

@@ -1,65 +1,20 @@
-import os
-import torch
-import gradio as gr
+from diffsynth.pipelines.qwen_image import QwenImagePipeline, ModelConfig
 from PIL import Image
-from diffusers import AutoModel
-from diffusers.utils import export_to_video
-from diffusers.pipelines.ltx.pipeline_ltx_condition import LTXConditionPipeline, LTXVideoCondition
+import torch
 
-# Carrega transformer em bfloat16 (simples e estável)
-transformer = AutoModel.from_pretrained(
-    "Lightricks/LTX-Video",
-    subfolder="transformer",
+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/"),
 )
-
-# Evita chamadas opcionais que quebram em versões diferentes
-# Ex.: enable_layerwise_casting pode não existir; deixe desativado por padrão
-
-pipeline = LTXConditionPipeline.from_pretrained(
-    "Lightricks/LTX-Video",
-    transformer=transformer,
-    torch_dtype=torch.bfloat16,
-).to("cuda")
-
-NEGATIVE = "worst quality, inconsistent motion, blurry, jittery, distorted"
-
-def generate(image: Image.Image, prompt: str, use_shift: bool, mu: float):
-    cond = LTXVideoCondition(image=image, frame_index=0, strength=1.0)
-    kwargs = dict(
-        conditions=[cond],
-        prompt=prompt,
-        negative_prompt=NEGATIVE,
-        width=480,
-        height=480,
-        num_frames=81,
-        num_inference_steps=50,
-    )
-    # Caminho simples: desativa shifting por padrão
-    if use_shift:
-        kwargs["use_dynamic_shifting"] = True
-        kwargs["mu"] = mu
-    else:
-        kwargs["use_dynamic_shifting"] = False
-
-    result = pipeline(**kwargs)
-    video = result.frames[0]
-    export_to_video(video, "output.mp4", fps=24)
-    return "output.mp4"
-
-with gr.Blocks(title="LTX-Video (Simple)", theme=gr.themes.Soft()) as demo:
-    in_image = gr.Image(type="pil", label="Imagem de entrada")
-    in_prompt = gr.Textbox(label="Prompt", placeholder="Descreva...")
-    use_shift = gr.Checkbox(value=False, label="Dynamic shifting (qnprevisa)")
-    mu = gr.Slider(0.0, 1.0, value=0.5, step=0.05, label="mu")
-    run_button = gr.Button("Gerar Vídeo", variant="primary")
-    video_out = gr.Video(label="Vídeo Gerado")
-
-    run_button.click(fn=generate, inputs=[in_image, in_prompt, use_shift, mu], outputs=[video_out])
-
-if __name__ == "__main__":
-    demo.queue().launch(
-        server_name=os.getenv("GRADIO_SERVER_NAME", "0.0.0.0"),
-        server_port=int(os.getenv("GRADIO_SERVER_PORT", "7861")),
-        show_error=True,
-    )
+prompt = "A detailed portrait of a girl underwater, wearing a blue flowing dress, hair gently floating, clear light and shadow, surrounded by bubbles, calm expression, fine details, dreamy and beautiful."
+image = pipe(
+    prompt, seed=0, num_inference_steps=40,
+    # edit_image=Image.open("xxx.jpg").resize((1328, 1328)) # For Qwen-Image-Edit
+)
+image.save("image.jpg")