import gradio as gr from PIL import Image import os import imageio # Importa a instância do nosso serviço # O modelo será carregado quando este módulo for importado from video_service import video_generation_service # --- FUNÇÕES DE AJUDA PARA A UI (não relacionadas ao modelo) --- TARGET_FIXED_SIDE = 768 MIN_DIM_SLIDER = 256 MAX_IMAGE_SIZE = 1280 def calculate_new_dimensions(orig_w, orig_h): if orig_w == 0 or orig_h == 0: return int(TARGET_FIXED_SIDE), int(TARGET_FIXED_SIDE) if orig_w >= orig_h: new_h, aspect_ratio = TARGET_FIXED_SIDE, orig_w / orig_h new_w = round((new_h * aspect_ratio) / 32) * 32 new_w = max(MIN_DIM_SLIDER, min(new_w, MAX_IMAGE_SIZE)) new_h = max(MIN_DIM_SLIDER, min(new_h, MAX_IMAGE_SIZE)) else: new_w, aspect_ratio = TARGET_FIXED_SIDE, orig_h / orig_w new_h = round((new_w * aspect_ratio) / 32) * 32 new_h = max(MIN_DIM_SLIDER, min(new_h, MAX_IMAGE_SIZE)) new_w = max(MIN_DIM_SLIDER, min(new_w, MAX_IMAGE_SIZE)) return int(new_h), int(new_w) def handle_media_upload_for_dims(filepath, current_h, current_w): if not filepath or not os.path.exists(str(filepath)): return gr.update(value=current_h), gr.update(value=current_w) try: if str(filepath).lower().endswith(('.png', '.jpg', '.jpeg', '.webp')): with Image.open(filepath) as img: orig_w, orig_h = img.size else: # Assumir que é um vídeo with imageio.get_reader(filepath) as reader: meta = reader.get_meta_data() orig_w, orig_h = meta.get('size', (current_w, current_h)) new_h, new_w = calculate_new_dimensions(orig_w, orig_h) return gr.update(value=new_h), gr.update(value=new_w) except Exception as e: print(f"Erro ao processar mídia para dimensões: {e}") return gr.update(value=current_h), gr.update(value=current_w) # --- FUNÇÃO WRAPPER PARA CHAMAR O SERVIÇO A PARTIR DO GRADIO --- def gradio_generate_wrapper(prompt, negative_prompt, input_image, input_video, height, width, mode, duration, frames_to_use, seed, randomize_seed, guidance_scale, improve_texture, progress=gr.Progress(track_tqdm=True)): """ Esta função atua como uma ponte entre a interface Gradio e o nosso VideoService. """ try: # Define a função de callback para a barra de progresso do Gradio def progress_handler(step, total_steps): progress(step / total_steps, desc="Salvando vídeo...") output_path, used_seed = video_generation_service.generate( prompt=prompt, negative_prompt=negative_prompt, input_image_filepath=input_image, input_video_filepath=input_video, height=int(height), width=int(width), mode=mode, duration=float(duration), frames_to_use=int(frames_to_use), seed=int(seed), randomize_seed=bool(randomize_seed), guidance_scale=float(guidance_scale), improve_texture=bool(improve_texture), progress_callback=progress_handler # Passamos o handler para o serviço ) return output_path, used_seed except ValueError as e: # Captura erros de validação do serviço e os exibe na UI raise gr.Error(str(e)) except Exception as e: # Captura outros erros inesperados print(f"Erro inesperado na geração: {e}") raise gr.Error("Ocorreu um erro inesperado. Verifique os logs.") # --- DEFINIÇÃO DA INTERFACE GRADIO --- css = "#col-container { margin: 0 auto; max-width: 900px; }" with gr.Blocks(css=css) as demo: gr.Markdown("# LTX Video 0.9.8 13B Distilled") gr.Markdown("Geração de vídeo rápida e de alta qualidade.") with gr.Row(): with gr.Column(): # ... (Layout das abas e componentes exatamente como antes) ... with gr.Tab("image-to-video") as image_tab: video_i_hidden = gr.Textbox(label="video_i", visible=False, value=None) image_i2v = gr.Image(label="Input Image", type="filepath", sources=["upload", "webcam", "clipboard"]) i2v_prompt = gr.Textbox(label="Prompt", value="The creature from the image starts to move", lines=3) i2v_button = gr.Button("Generate Image-to-Video", variant="primary") with gr.Tab("text-to-video") as text_tab: image_n_hidden = gr.Textbox(label="image_n", visible=False, value=None) video_n_hidden = gr.Textbox(label="video_n", visible=False, value=None) t2v_prompt = gr.Textbox(label="Prompt", value="A majestic dragon flying over a medieval castle", lines=3) t2v_button = gr.Button("Generate Text-to-Video", variant="primary") with gr.Tab("video-to-video", visible=True) as video_tab: image_v_hidden = gr.Textbox(label="image_v", visible=False, value=None) video_v2v = gr.Video(label="Input Video", sources=["upload", "webcam"]) frames_to_use = gr.Slider(label="Frames to use from input video", minimum=9, maximum=257, value=9, step=8, info="Must be N*8+1.") v2v_prompt = gr.Textbox(label="Prompt", value="Change the style to cinematic anime", lines=3) v2v_button = gr.Button("Generate Video-to-Video", variant="primary") duration_input = gr.Slider(label="Video Duration (seconds)", minimum=0.3, maximum=8.5, value=2, step=0.1) improve_texture = gr.Checkbox(label="Improve Texture (multi-scale)", value=True, visible=True) with gr.Column(): output_video = gr.Video(label="Generated Video", interactive=False) with gr.Accordion("Advanced settings", open=False): mode = gr.Dropdown(["text-to-video", "image-to-video", "video-to-video"], label="task", value="image-to-video", visible=False) negative_prompt_input = gr.Textbox(label="Negative Prompt", value="worst quality, inconsistent motion, blurry, jittery, distorted", lines=2) with gr.Row(): seed_input = gr.Number(label="Seed", value=42, precision=0) randomize_seed_input = gr.Checkbox(label="Randomize Seed", value=True) guidance_scale_input = gr.Slider(label="Guidance Scale (CFG)", minimum=1.0, maximum=10.0, value=3.0, step=0.1) with gr.Row(): height_input = gr.Slider(label="Height", value=512, step=32, minimum=MIN_DIM_SLIDER, maximum=MAX_IMAGE_SIZE) width_input = gr.Slider(label="Width", value=704, step=32, minimum=MIN_DIM_SLIDER, maximum=MAX_IMAGE_SIZE) # --- LÓGICA DE EVENTOS DA UI --- image_i2v.upload(fn=handle_media_upload_for_dims, inputs=[image_i2v, height_input, width_input], outputs=[height_input, width_input]) video_v2v.upload(fn=handle_media_upload_for_dims, inputs=[video_v2v, height_input, width_input], outputs=[height_input, width_input]) image_tab.select(fn=lambda: "image-to-video", outputs=[mode]) text_tab.select(fn=lambda: "text-to-video", outputs=[mode]) video_tab.select(fn=lambda: "video-to-video", outputs=[mode]) common_inputs = [negative_prompt_input, height_input, width_input, mode, duration_input, frames_to_use, seed_input, randomize_seed_input, guidance_scale_input, improve_texture] common_outputs = [output_video, seed_input] t2v_button.click(fn=gradio_generate_wrapper, inputs=[t2v_prompt, *common_inputs[:1], image_n_hidden, video_n_hidden, *common_inputs[1:]], outputs=common_outputs, api_name="text_to_video") i2v_button.click(fn=gradio_generate_wrapper, inputs=[i2v_prompt, *common_inputs[:1], image_i2v, video_i_hidden, *common_inputs[1:]], outputs=common_outputs, api_name="image_to_video") v2v_button.click(fn=gradio_generate_wrapper, inputs=[v2v_prompt, *common_inputs[:1], image_v_hidden, video_v2v, *common_inputs[1:]], outputs=common_outputs, api_name="video_to_video") if __name__ == "__main__": demo.queue().launch(debug=True, share=False)