# app.py import gradio as gr import os import sys import traceback from pathlib import Path # --- Import dos Serviços de Backend --- try: from api.ltx_server_refactored import video_generation_service except ImportError: print("ERRO FATAL: Não foi possível importar 'video_generation_service'.") sys.exit(1) try: from api.seedvr_server import SeedVRServer except ImportError: print("AVISO: Não foi possível importar SeedVRServer.") SeedVRServer = None seedvr_inference_server = SeedVRServer() if SeedVRServer else None # --- ESTADO DA SESSÃO --- def create_initial_state(): return {"low_res_video": None, "low_res_latents": None, "used_seed": None} # --- FUNÇÕES WRAPPER PARA A UI --- def run_generate_base_video( generation_mode, prompt, neg_prompt, start_img, height, width, duration, seed, randomize_seed, fp_num_inference_steps, ship_initial_inference_steps, ship_final_inference_steps, fp_guidance_preset, fp_guidance_scale_list, fp_stg_scale_list, fp_timesteps_list, progress=gr.Progress(track_tqdm=True) ): print(f"UI: Iniciando geração no modo: {generation_mode}") try: initial_image_conditions = [] if start_img: num_frames_estimate = int(duration * 24) items_list = [[start_img, 0, 1.0]] initial_image_conditions = video_generation_service.prepare_condition_items(items_list, height, width, num_frames_estimate) used_seed = None if randomize_seed else seed ltx_configs = { "guidance_preset": fp_guidance_preset, "guidance_scale_list": fp_guidance_scale_list, "stg_scale_list": fp_stg_scale_list, "timesteps_list": fp_timesteps_list, "fp_num_inference_steps": fp_num_inference_steps, "ship_initial_inference_steps": ship_initial_inference_steps, "ship_final_inference_steps": ship_final_inference_steps, } if generation_mode == "Narrativa (Múltiplos Prompts)": video_path, tensor_path, final_seed = video_generation_service.generate_narrative_low( prompt=prompt, negative_prompt=neg_prompt, height=height, width=width, duration=duration, seed=used_seed, initial_image_conditions=initial_image_conditions, ltx_configs_override=ltx_configs, ) else: video_path, tensor_path, final_seed = video_generation_service.generate_single_low( prompt=prompt, negative_prompt=neg_prompt, height=height, width=width, duration=duration, seed=used_seed, initial_image_conditions=initial_image_conditions, ltx_configs_override=ltx_configs, ) if video_path is None: raise gr.Error("A geração do vídeo falhou no backend. Verifique os logs do console para detalhes.") new_state = {"low_res_video": video_path, "low_res_latents": tensor_path, "used_seed": final_seed} return video_path, new_state, gr.update(visible=True) except Exception as e: error_message = f"❌ Ocorreu um erro na Geração Base:\n{e}" print(f"{error_message}\nDetalhes: {traceback.format_exc()}") raise gr.Error(error_message) def run_ltx_refinement(state, prompt, neg_prompt, cfg, progress=gr.Progress(track_tqdm=True)): if not state or not state.get("low_res_latents"): raise gr.Error("Erro: Gere um vídeo base primeiro na Etapa 1.") try: video_path, tensor_path = video_generation_service.generate_upscale_denoise( latents_path=state["low_res_latents"], prompt=prompt, negative_prompt=neg_prompt, guidance_scale=cfg, seed=state["used_seed"] ) state["refined_video_ltx"] = video_path; state["refined_latents_ltx"] = tensor_path return video_path, state except Exception as e: raise gr.Error(f"Erro no Refinamento LTX: {e}") def run_seedvr_upscaling(state, seed, resolution, batch_size, fps, progress=gr.Progress(track_tqdm=True)): if not state or not state.get("low_res_video"): raise gr.Error("Erro: Gere um vídeo base primeiro na Etapa 1.") if not seedvr_inference_server: raise gr.Error("Erro: O servidor SeedVR não está disponível.") try: def progress_wrapper(p, desc=""): progress(p, desc=desc) output_filepath = seedvr_inference_server.run_inference( file_path=state["low_res_video"], seed=seed, resolution=resolution, batch_size=batch_size, fps=fps, progress=progress_wrapper ) return gr.update(value=output_filepath), gr.update(value=f"✅ Concluído!\nSalvo em: {output_filepath}") except Exception as e: return None, gr.update(value=f"❌ Erro no SeedVR:\n{e}") # --- DEFINIÇÃO DA INTERFACE GRADIO --- with gr.Blocks() as demo: gr.Markdown("# LTX Video - Geração e Pós-Produção por Etapas") app_state = gr.State(value=create_initial_state()) with gr.Row(): with gr.Column(scale=1): gr.Markdown("### Etapa 1: Configurações de Geração") generation_mode_input = gr.Radio( label="Modo de Geração", choices=["Simples (Prompt Único)", "Narrativa (Múltiplos Prompts)"], value="Narrativa (Múltiplos Prompts)", info="Simples para uma ação, Narrativa para uma sequência (uma cena por linha)." ) prompt_input = gr.Textbox(label="Prompt(s)", value="Um leão majestoso caminha pela savana\nEle sobe em uma grande pedra e olha o horizonte", lines=4) neg_prompt_input = gr.Textbox(label="Negative Prompt", value="blurry, low quality, bad anatomy", lines=2) start_image = gr.Image(label="Imagem de Início (Opcional)", type="filepath", sources=["upload"]) with gr.Accordion("Parâmetros Principais", open=True): duration_input = gr.Slider(label="Duração Total (s)", value=8, step=1, minimum=1, maximum=40) with gr.Row(): height_input = gr.Slider(label="Height", value=512, step=32, minimum=256, maximum=1024) width_input = gr.Slider(label="Width", value=704, step=32, minimum=256, maximum=1024) with gr.Row(): seed_input = gr.Number(label="Seed", value=42, precision=0) randomize_seed = gr.Checkbox(label="Randomize Seed", value=True) with gr.Accordion("Opções Adicionais LTX (Avançado)", open=False): with gr.Tabs(): with gr.TabItem("Guiagem (First Pass)"): fp_guidance_preset = gr.Dropdown( label="Preset de Guiagem", choices=["Padrão (Recomendado)", "Agressivo", "Suave", "Customizado"], value="Padrão (Recomendado)", info="Muda o comportamento da guiagem ao longo da difusão." ) with gr.Group(visible=False) as custom_guidance_group: gr.Markdown("⚠️ Edite as listas em formato JSON. Ex: `[1, 2, 3]`") fp_guidance_scale_list = gr.Textbox(label="Lista de Guidance Scale", value="[1, 1, 6, 8, 6, 1, 1]") fp_stg_scale_list = gr.Textbox(label="Lista de STG Scale (Movimento)", value="[0, 0, 4, 4, 4, 2, 1]") fp_timesteps_list = gr.Textbox(label="Lista de Guidance Timesteps", value="[1.0, 0.996, 0.9933, 0.9850, 0.9767, 0.9008, 0.6180]") with gr.TabItem("Scheduler"): fp_num_inference_steps = gr.Slider(label="Passos de Inferência", minimum=10, maximum=100, step=1, value=30) ship_initial_inference_steps = gr.Slider(label="Pular Passos Iniciais (img2vid)", minimum=0, maximum=100, step=1, value=0) ship_final_inference_steps = gr.Slider(label="Pular Passos Finais", minimum=0, maximum=100, step=1, value=0) generate_low_btn = gr.Button("1. Gerar Vídeo Base", variant="primary") with gr.Column(scale=1): gr.Markdown("### Vídeo Base Gerado") low_res_video_output = gr.Video(label="O resultado da Etapa 1 aparecerá aqui", interactive=False) def update_custom_guidance_visibility(preset_choice): return gr.update(visible=(preset_choice == "Customizado")) fp_guidance_preset.change(fn=update_custom_guidance_visibility, inputs=fp_guidance_preset, outputs=custom_guidance_group) all_ltx_inputs = [ fp_num_inference_steps, ship_initial_inference_steps, ship_final_inference_steps, fp_guidance_preset, fp_guidance_scale_list, fp_stg_scale_list, fp_timesteps_list, ] generate_low_btn.click( fn=run_generate_base_video, inputs=[ generation_mode_input, prompt_input, neg_prompt_input, start_image, height_input, width_input, duration_input, seed_input, randomize_seed, *all_ltx_inputs, # Passa o `cfg` por último, pois não está no `all_ltx_inputs` ], outputs=[low_res_video_output, app_state, post_prod_group] ) ltx_refine_btn.click( fn=run_ltx_refinement, inputs=[app_state, prompt_input, neg_prompt_input, cfg_input], outputs=[ltx_refined_video_output, app_state] ) run_seedvr_button.click( fn=run_seedvr_upscaling, inputs=[app_state, seedvr_seed, seedvr_resolution, seedvr_batch_size, seedvr_fps_output], outputs=[seedvr_video_output, seedvr_status_box] ) if __name__ == "__main__": demo.queue().launch(server_name="0.0.0.0", server_port=7860, debug=True, show_error=True)