Fix: Switch to Native Gradio SDK for ZeroGPU stability

app.py

@@ -1,9 +1,8 @@
-import sys
-import os
-import gc
-import torch
+import sys, os, gc, torch, spaces, tempfile
+import gradio as gr
+from PIL import Image
 
-# --- PARCHE DE GRADIO ---
+# PATCH GRADIO
 try:
     import gradio_client.utils as client_utils
     if not hasattr(client_utils, "_old_json_schema_to_python_type"):
@@ -14,135 +13,79 @@ try:
     client_utils._json_schema_to_python_type = patched_json_schema_to_python_type
 except: pass
 
-import spaces
-import gradio as gr
-from PIL import Image
-import tempfile
-
-# CONFIG
-MODELS = {
-    "Pony Diffusion V6 XL": "cyberdelia/CyberRealisticPony",
-    "RealVisXL V4.0": "SG161222/RealVisXL_V4.0"
-}
+def flush():
+    gc.collect()
+    if torch.cuda.is_available(): torch.cuda.empty_cache()
 
+MODELS = {"Pony Diffusion V6 XL": "cyberdelia/CyberRealisticPony"}
 LORAS = {
     "Ninguno": "",
-    "💎 NSFW: Real Nudity": "Lora-Daddy/Ltx2.3-real-nudity-early-alpha-30k-steps", 
-    "📜 DOCS: ID Card / Passport": "j0rdan/passport-sdxl",
-    "🔫 WEAPONS: Tactical Gear": "Ostris/SDXL_LoRA_Test",
-    "✍️ TEXT: Typography": "ntc/Typography-SDXL"
+    "💎 NSFW: Real Nudity": "Lora-Daddy/Ltx2.3-real-nudity-early-alpha-30k-steps",
+    "📜 DOCS: ID Card": "j0rdan/passport-sdxl",
+    "🔫 WEAPONS: Tactical": "Ostris/SDXL_LoRA_Test"
 }
 
-def flush():
-    gc.collect()
-    if torch.cuda.is_available():
-        torch.cuda.empty_cache()
-
-# --- 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):
+def generate(prompt, lora_name, w, h, init_img=None, strength=0.6):
     flush()
-    # Importación local para ahorrar RAM
     from diffusers import StableDiffusionXLPipeline, StableDiffusionXLImg2ImgPipeline
+    p = f"score_9, score_8_up, score_7_up, {prompt}"
+    pipe = StableDiffusionXLPipeline.from_pretrained("cyberdelia/CyberRealisticPony", torch_dtype=torch.float16, variant="fp16", low_cpu_mem_usage=True).to("cuda")
     
-    model_id = MODELS.get(model_name)
-    
-    # Inyectar tags de calidad Pony
-    if "Pony" in model_name:
-        prompt = f"score_9, score_8_up, score_7_up, {prompt}"
-        
-    # 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:
-            pipe.load_lora_weights(lora_id)
-            pipe.fuse_lora(lora_scale=lora_scale)
+    lid = LORAS.get(lora_name)
+    if lid:
+        try: pipe.load_lora_weights(lid)
         except: pass
 
-    if init_img is not None:
+    if init_img:
         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]
+        res = pipe_i2i(prompt=p, image=init_img, strength=strength, num_inference_steps=25).images[0]
         del pipe_i2i
     else:
-        res = pipe(prompt=prompt, negative_prompt=neg, num_inference_steps=int(steps), guidance_scale=cfg, width=int(w), height=int(h)).images[0]
+        res = pipe(prompt=p, num_inference_steps=30, 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):
+def video(prompt, init_img):
     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.enable_model_cpu_offload()
-    pipe.enable_vae_slicing()
-    
-    kwargs = {
-        "prompt": f"score_9, {prompt}",
-        "negative_prompt": "low quality, blurry, static, ugly",
-        "num_inference_steps": int(steps),
-        "guidance_scale": cfg,
-        "num_frames": 33,
-        "width": 704,
-        "height": 480
-    }
-    
-    if init_img is not None:
-        kwargs["image"] = init_img
-        
-    output = pipe(**kwargs)
+    pipe = LTXPipeline.from_pretrained("Lightricks/LTX-Video", torch_dtype=torch.bfloat16, low_cpu_mem_usage=True).to("cuda")
+    kw = {"prompt": prompt, "num_inference_steps": 20, "num_frames": 25, "width": 704, "height": 480}
+    if init_img: kw["image"] = init_img
+    out = pipe(**kw)
     tmp = tempfile.NamedTemporaryFile(suffix=".mp4", delete=False)
-    export_to_video(output.frames[0], tmp.name, fps=16)
-    
+    export_to_video(out.frames[0], tmp.name, fps=16)
     del pipe
     flush()
     return tmp.name
 
-# --- 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.Blocks() as demo:
+    gr.HTML("<h1 style='text-align:center;'>🌌 Omni-Studio v3.2</h1>")
     with gr.Tabs():
-        with gr.Tab("🎨 Imagen"):
+        with gr.Tab("🖼 Imagen"):
             with gr.Row():
                 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")
+                    p = gr.Textbox(label="Prompt")
                     l = gr.Dropdown(choices=list(LORAS.keys()), value="Ninguno", label="LoRA")
-                    with gr.Row():
-                        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")
-                    
+                    w = gr.Slider(512, 1024, 832, step=64)
+                    h = gr.Slider(512, 1024, 1216, step=64)
+                    img = gr.Image(label="Base", type="pil")
+                    st = gr.Slider(0.1, 0.9, 0.6, label="Mod Strength")
+                    btn = gr.Button("GENERAR")
+                out = gr.Image(label="Resultado")
         with gr.Tab("🎥 Video"):
             with gr.Row():
                 with gr.Column():
-                    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")
+                    vp = gr.Textbox(label="Prompt")
+                    vi = gr.Image(label="Base", type="pil")
+                    vbtn = gr.Button("GENERAR VIDEO")
+                vout = gr.Video(label="Resultado")
 
-    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)
+    btn.click(generate, [p, l, w, h, img, st], out)
+    vbtn.click(video, [vp, vi], vout)
 
 demo.queue().launch(show_api=False, server_name="0.0.0.0", server_port=7860)

requirements.txt

@@ -1,14 +1,11 @@
-spaces
 gradio==4.44.1
-fastapi==0.112.2
-starlette==0.38.2
-huggingface-hub==0.24.2
-audioop-lts
+spaces
+torch
 diffusers>=0.31.0
-transformers>=4.44.0
-accelerate>=0.33.0
+transformers
+accelerate
 peft
-torch
 sentencepiece
-imageio[ffmpeg]
-av
+safetensors
+numpy
+pillow