Fix: Switch to Native Gradio SDK for ZeroGPU stability

app.py

@@ -3,7 +3,7 @@ import os
 import gc
 import torch
 
-# --- PATCH GRADIO RECURSION ---
+# --- PARCHE DE GRADIO ---
 try:
     import gradio_client.utils as client_utils
     if not hasattr(client_utils, "_old_json_schema_to_python_type"):
@@ -19,43 +19,47 @@ import gradio as gr
 from PIL import Image
 import tempfile
 
-# CONFIG MODELOS Y LORAS
+# CONFIG
 MODELS = {
-    "Pony Diffusion V6 XL (Ultra Realismo)": "cyberdelia/CyberRealisticPony",
-    "RealVisXL V4.0 (Fotografía)": "SG161222/RealVisXL_V4.0"
+    "Pony Diffusion V6 XL": "cyberdelia/CyberRealisticPony",
+    "RealVisXL V4.0": "SG161222/RealVisXL_V4.0"
 }
 
 LORAS = {
     "Ninguno": "",
-    "💎 NSFW: Real Nudity (Anatomía)": "Lora-Daddy/Ltx2.3-real-nudity-early-alpha-30k-steps", 
+    "💎 NSFW: Real Nudity": "Lora-Daddy/Ltx2.3-real-nudity-early-alpha-30k-steps", 
     "📜 DOCS: ID Card / Passport": "j0rdan/passport-sdxl",
-    "🔫 WEAPONS: Tactical Gear & Guns": "Ostris/SDXL_LoRA_Test",
-    "✍️ TEXT: Typography Fix": "ntc/Typography-SDXL"
+    "🔫 WEAPONS: Tactical Gear": "Ostris/SDXL_LoRA_Test",
+    "✍️ TEXT: Typography": "ntc/Typography-SDXL"
 }
 
-LTX_MODELS = {"LTX-Video Pro": "Lightricks/LTX-Video"}
-
 def flush():
     gc.collect()
-    torch.cuda.empty_cache()
+    if torch.cuda.is_available():
+        torch.cuda.empty_cache()
 
-# --- MOTORES ---
-@spaces.GPU(duration=150)
+# --- MOTOR DE IMAGEN ---
+@spaces.GPU(duration=120)
 def process_image(prompt, neg, model_name, lora_name, lora_id_custom, lora_scale, steps, cfg, w, h, init_img=None, strength=0.6):
     flush()
+    # Importación local para ahorrar RAM
     from diffusers import StableDiffusionXLPipeline, StableDiffusionXLImg2ImgPipeline
-    model_id = MODELS.get(model_name)
     
-    # Carga base
-    pipe = StableDiffusionXLPipeline.from_pretrained(
-        model_id, torch_dtype=torch.float16, variant="fp16", low_cpu_mem_usage=True
-    ).to("cuda")
+    model_id = MODELS.get(model_name)
     
-    # Inyectar Calidad Pony si es necesario
+    # Inyectar tags de calidad Pony
     if "Pony" in model_name:
         prompt = f"score_9, score_8_up, score_7_up, {prompt}"
         
-    # Cargar LoRA
+    # Carga con bajo consumo de CPU
+    pipe = StableDiffusionXLPipeline.from_pretrained(
+        model_id, torch_dtype=torch.float16, variant="fp16", use_safetensors=True, 
+        low_cpu_mem_usage=True
+    ).to("cuda")
+    
+    # Activar offload para ahorrar RAM
+    pipe.enable_model_cpu_offload()
+
     lora_id = lora_id_custom if lora_id_custom else LORAS.get(lora_name)
     if lora_id:
         try:
@@ -64,29 +68,33 @@ def process_image(prompt, neg, model_name, lora_name, lora_id_custom, lora_scale
         except: pass
 
     if init_img is not None:
-        # Modo Image-to-Image (Modificación)
         pipe_i2i = StableDiffusionXLImg2ImgPipeline.from_pipe(pipe)
         res = pipe_i2i(prompt=prompt, negative_prompt=neg, image=init_img, strength=strength, num_inference_steps=int(steps), guidance_scale=cfg).images[0]
         del pipe_i2i
     else:
-        # Modo Text-to-Image (Creación)
         res = pipe(prompt=prompt, negative_prompt=neg, num_inference_steps=int(steps), guidance_scale=cfg, width=int(w), height=int(h)).images[0]
     
     del pipe
     flush()
     return res
 
+# --- MOTOR DE VIDEO ---
 @spaces.GPU(duration=250)
 def process_video(prompt, init_img, steps, cfg):
     flush()
     from diffusers import LTXPipeline
     from diffusers.utils import export_to_video
     
-    pipe = LTXPipeline.from_pretrained("Lightricks/LTX-Video", torch_dtype=torch.bfloat16, low_cpu_mem_usage=True).to("cuda")
+    pipe = LTXPipeline.from_pretrained(
+        "Lightricks/LTX-Video", torch_dtype=torch.bfloat16, 
+        low_cpu_mem_usage=True
+    ).to("cuda")
+    
+    pipe.enable_model_cpu_offload()
     pipe.enable_vae_slicing()
     
     kwargs = {
-        "prompt": prompt,
+        "prompt": f"score_9, {prompt}",
         "negative_prompt": "low quality, blurry, static, ugly",
         "num_inference_steps": int(steps),
         "guidance_scale": cfg,
@@ -106,44 +114,35 @@ def process_video(prompt, init_img, steps, cfg):
     flush()
     return tmp.name
 
-# --- UI ---
-with gr.Blocks(theme=gr.themes.Default(primary_hue="blue", secondary_hue="indigo")) as demo:
-    gr.HTML("<h1 style='text-align:center;'>🌌 Omni-Studio Pro v3.0</h1>")
+# --- INTERFAZ ---
+with gr.Blocks(theme=gr.themes.Soft()) as demo:
+    gr.HTML("<h1 style='text-align:center;'>🌌 Omni-Studio v3.1 (Power-Optimized)</h1>")
     
     with gr.Tabs():
-        with gr.Tab("🎨 Imagen (Crear / Modificar)"):
+        with gr.Tab("🎨 Imagen"):
             with gr.Row():
-                with gr.Column(scale=1):
-                    prompt = gr.Textbox(label="Prompt Principal", placeholder="Escribe lo que quieres ver...", lines=3)
-                    neg = gr.Textbox(label="Negativo", value="blurry, ugly, distorted, lowres")
-                    with gr.Row():
-                        model = gr.Dropdown(choices=list(MODELS.keys()), value=list(MODELS.keys())[0], label="Motor")
-                        category = gr.Dropdown(choices=list(LORAS.keys()), value="Ninguno", label="Especialidad (LoRA)")
-                    with gr.Row():
-                        l_custom = gr.Textbox(label="LoRA ID Personalizado")
-                        l_scale = gr.Slider(0, 2.0, 0.8, label="Fuerza LoRA")
+                with gr.Column():
+                    p = gr.Textbox(label="Prompt", lines=3)
+                    n = gr.Textbox(label="Negativo", value="blurry, lowres")
+                    m = gr.Dropdown(choices=list(MODELS.keys()), value=list(MODELS.keys())[0], label="Motor")
+                    l = gr.Dropdown(choices=list(LORAS.keys()), value="Ninguno", label="LoRA")
                     with gr.Row():
-                        w = gr.Slider(512, 1024, 1024, step=64, label="Ancho")
-                        h = gr.Slider(512, 1024, 1024, step=64, label="Alto")
-                    img_input = gr.Image(label="Imagen Base (Opcional para Modificar)", type="pil")
-                    strength = gr.Slider(0.1, 0.9, 0.6, label="Fuerza de Modificación (I2I)")
-                    btn_i = gr.Button("🚀 GENERAR / TRANSFORMAR", variant="primary")
-                with gr.Column(scale=1):
-                    out_i = gr.Image(label="Resultado")
+                        w = gr.Slider(512, 1024, 832, step=64, label="Ancho")
+                        h = gr.Slider(512, 1024, 1216, step=64, label="Alto")
+                    img_in = gr.Image(label="Imagen Base (Opcional)", type="pil")
+                    str_i = gr.Slider(0.1, 0.9, 0.6, label="Fuerza Mod")
+                    btn_i = gr.Button("GENERAR IMAGEN", variant="primary")
+                out_i = gr.Image(label="Resultado")
                     
-        with gr.Tab("🎥 Video (T2V / I2V)"):
+        with gr.Tab("🎥 Video"):
             with gr.Row():
                 with gr.Column():
-                    v_prompt = gr.Textbox(label="Video Prompt", lines=3)
-                    v_input = gr.Image(label="Imagen Inicial (Opcional)", type="pil")
-                    with gr.Row():
-                        v_steps = gr.Slider(10, 40, 25, step=1, label="Pasos")
-                        v_cfg = gr.Slider(1, 7, 3.5, label="Guidance")
-                    btn_v = gr.Button("🎬 GENERAR VIDEO", variant="primary")
-                with gr.Column():
-                    out_v = gr.Video(label="Resultado Video")
+                    vp = gr.Textbox(label="Video Prompt", lines=3)
+                    vin = gr.Image(label="Imagen Inicial (Opcional)", type="pil")
+                    btn_v = gr.Button("GENERAR VIDEO", variant="primary")
+                out_v = gr.Video(label="Resultado Video")
 
-    btn_i.click(process_image, [prompt, neg, model, category, l_custom, l_scale, gr.Number(value=30, visible=False), gr.Number(value=7, visible=False), w, h, img_input, strength], out_i)
-    btn_v.click(process_video, [v_prompt, v_input, v_steps, v_cfg], out_v)
+    btn_i.click(process_image, [p, n, m, l, gr.Textbox(visible=False), gr.Number(value=0.8, visible=False), gr.Number(value=30, visible=False), gr.Number(value=7, visible=False), w, h, img_in, str_i], out_i)
+    btn_v.click(process_video, [vp, vin, gr.Number(value=25, visible=False), gr.Number(value=3.5, visible=False)], out_v)
 
 demo.queue().launch(show_api=False, server_name="0.0.0.0", server_port=7860)