Update: Ultra Realistic Model + Dynamic LoRA loader

app.py

@@ -4,238 +4,140 @@ import torch
 import numpy as np
 from PIL import Image
 import tempfile, os
+from huggingface_hub import hf_hub_download
 
-# ─── MODELOS ───────────────────────────────────────────────────────────────────
-TEXT2IMG_MODEL  = "SG161222/RealVisXL_V4.0"          # Realista, sin censura
-IMG2IMG_MODEL   = "SG161222/RealVisXL_V4.0"
-LTX_MODEL       = "Lightricks/LTX-Video"
-LTX_LORA        = "Lora-Daddy/Ltx2.3-real-nudity-early-alpha-30k-steps"
+# ─── CONFIGURACIÓN DE MODELOS ──────────────────────────────────────────────────
+# Usamos CyberRealistic Pony por su alto realismo en humanos
+BASE_MODEL = "cyberdelia/CyberRealisticPony"
+LTX_MODEL  = "Lightricks/LTX-Video"
+DEFAULT_LORA = "John6666/nsfw-master-flux-lora-merged" # Sugerido para imagenes
+LTX_NSFW_LORA = "Lora-Daddy/Ltx2.3-real-nudity-early-alpha-30k-steps"
 
-pipe_t2i   = None
-pipe_i2i   = None
+pipe_t2i = None
+pipe_i2i = None
 pipe_video = None
 
-NEG_DEFAULT = (
-    "censored, blurry, low quality, bad anatomy, deformed, ugly, "
-    "watermark, logo, text, worst quality, jpeg artifacts"
-)
+NEG_DEFAULT = "censored, blurry, low quality, bad anatomy, deformed, ugly, watermark, text, signature"
 
 # ─── LOADERS ───────────────────────────────────────────────────────────────────
-def load_t2i():
+def load_t2i(lora_id=None, lora_scale=1.0):
     global pipe_t2i
+    from diffusers import StableDiffusionXLPipeline
     if pipe_t2i is None:
-        from diffusers import StableDiffusionXLPipeline
         pipe_t2i = StableDiffusionXLPipeline.from_pretrained(
-            TEXT2IMG_MODEL, torch_dtype=torch.float16, use_safetensors=True, variant="fp16"
+            BASE_MODEL, torch_dtype=torch.float16, use_safetensors=True, variant="fp16"
         )
+    
+    if lora_id:
+        try:
+            pipe_t2i.unload_lora_weights()
+            pipe_t2i.load_lora_weights(lora_id)
+            pipe_t2i.fuse_lora(lora_scale=lora_scale)
+        except Exception as e:
+            print(f"Error cargando LoRA: {e}")
     return pipe_t2i
 
-def load_i2i():
+def load_i2i(lora_id=None, lora_scale=1.0):
     global pipe_i2i
+    from diffusers import StableDiffusionXLImg2ImgPipeline
     if pipe_i2i is None:
-        from diffusers import StableDiffusionXLImg2ImgPipeline
         pipe_i2i = StableDiffusionXLImg2ImgPipeline.from_pretrained(
-            IMG2IMG_MODEL, torch_dtype=torch.float16, use_safetensors=True, variant="fp16"
+            BASE_MODEL, torch_dtype=torch.float16, use_safetensors=True, variant="fp16"
         )
+    if lora_id:
+        try:
+            pipe_i2i.unload_lora_weights()
+            pipe_i2i.load_lora_weights(lora_id)
+            pipe_i2i.fuse_lora(lora_scale=lora_scale)
+        except Exception as e:
+            print(f"Error cargando LoRA: {e}")
     return pipe_i2i
 
-def load_video(use_lora=True):
+def load_video():
     global pipe_video
+    from diffusers import LTXPipeline
     if pipe_video is None:
-        from diffusers import LTXPipeline
-        pipe_video = LTXPipeline.from_pretrained(
-            LTX_MODEL, torch_dtype=torch.bfloat16
-        )
-        if use_lora:
-            try:
-                pipe_video.load_lora_weights(LTX_LORA)
-                print("[OK] LoRA NSFW cargado.")
-            except Exception as e:
-                print(f"[WARN] LoRA no cargado: {e}")
+        pipe_video = LTXPipeline.from_pretrained(LTX_MODEL, torch_dtype=torch.bfloat16)
+        try:
+            pipe_video.load_lora_weights(LTX_NSFW_LORA)
+        except: pass
     return pipe_video
 
 # ─── FUNCIONES ─────────────────────────────────────────────────────────────────
 @spaces.GPU(duration=120)
-def text2img(prompt, neg, steps, cfg, w, h, seed):
-    pipe = load_t2i().to("cuda")
+def generate_t2i(prompt, neg, lora_id, lora_scale, steps, cfg, w, h, seed):
+    pipe = load_t2i(lora_id if lora_id else None, lora_scale).to("cuda")
     gen = torch.Generator("cuda").manual_seed(int(seed))
-    img = pipe(prompt=prompt, negative_prompt=neg,
-                num_inference_steps=int(steps), guidance_scale=cfg,
-                width=int(w), height=int(h), generator=gen).images[0]
+    img = pipe(prompt=prompt, negative_prompt=neg, num_inference_steps=int(steps), 
+               guidance_scale=cfg, width=int(w), height=int(h), generator=gen).images[0]
     pipe.to("cpu"); torch.cuda.empty_cache()
     return img
 
 @spaces.GPU(duration=120)
-def img2img(prompt, neg, init_image, strength, steps, cfg, seed):
-    if init_image is None:
-        return None
-    pipe = load_i2i().to("cuda")
+def generate_i2i(prompt, neg, init_image, strength, lora_id, lora_scale, steps, cfg, seed):
+    if init_image is None: return None
+    pipe = load_i2i(lora_id if lora_id else None, lora_scale).to("cuda")
     gen = torch.Generator("cuda").manual_seed(int(seed))
     img = Image.fromarray(init_image).convert("RGB").resize((1024, 1024))
-    result = pipe(prompt=prompt, negative_prompt=neg, image=img,
-                  strength=strength, num_inference_steps=int(steps),
-                  guidance_scale=cfg, generator=gen).images[0]
+    res = pipe(prompt=prompt, negative_prompt=neg, image=img, strength=strength,
+               num_inference_steps=int(steps), guidance_scale=cfg, generator=gen).images[0]
     pipe.to("cpu"); torch.cuda.empty_cache()
-    return result
+    return res
 
 @spaces.GPU(duration=200)
-def img2video(prompt, neg, init_image, num_frames, fps, steps, lora_scale, seed):
+def generate_video(prompt, neg, init_image, num_frames, fps, steps, lora_scale, seed):
     from diffusers.utils import export_to_video
-    pipe = load_video(use_lora=True).to("cuda")
+    pipe = load_video().to("cuda")
     gen = torch.Generator("cuda").manual_seed(int(seed))
-
-    kwargs = dict(
-        prompt=prompt, negative_prompt=neg,
-        num_frames=int(num_frames), num_inference_steps=int(steps),
-        generator=gen,
-    )
-    if hasattr(pipe, "image") and init_image is not None:
-        img = Image.fromarray(init_image).convert("RGB").resize((768, 512))
-        kwargs["image"] = img
+    
+    kwargs = {"prompt": prompt, "negative_prompt": neg, "num_frames": int(num_frames), 
+              "num_inference_steps": int(steps), "generator": gen}
+    
+    if init_image is not None:
+        kwargs["image"] = Image.fromarray(init_image).convert("RGB").resize((768, 512))
+    
     if lora_scale > 0:
         kwargs["cross_attention_kwargs"] = {"scale": lora_scale}
-
-    output = pipe(**kwargs)
-    frames = output.frames[0]
-
-    tmp = tempfile.NamedTemporaryFile(suffix=".mp4", delete=False)
-    export_to_video(frames, tmp.name, fps=int(fps))
-    pipe.to("cpu"); torch.cuda.empty_cache()
-    return tmp.name
-
-@spaces.GPU(duration=120)
-def text2video(prompt, neg, num_frames, fps, w, h, steps, lora_scale, seed):
-    from diffusers import LTXPipeline
-    from diffusers.utils import export_to_video
-    pipe = load_video(use_lora=True).to("cuda")
-    gen = torch.Generator("cuda").manual_seed(int(seed))
-
-    kwargs = dict(
-        prompt=prompt, negative_prompt=neg,
-        num_frames=int(num_frames), width=int(w), height=int(h),
-        num_inference_steps=int(steps), generator=gen,
-    )
-    if lora_scale > 0:
-        kwargs["cross_attention_kwargs"] = {"scale": lora_scale}
-
+        
     output = pipe(**kwargs)
-    frames = output.frames[0]
     tmp = tempfile.NamedTemporaryFile(suffix=".mp4", delete=False)
-    export_to_video(frames, tmp.name, fps=int(fps))
+    export_to_video(output.frames[0], tmp.name, fps=int(fps))
     pipe.to("cpu"); torch.cuda.empty_cache()
     return tmp.name
 
 # ─── UI ────────────────────────────────────────────────────────────────────────
-THEME = gr.themes.Base(
-    primary_hue="violet", secondary_hue="purple", neutral_hue="slate"
-).set(
-    body_background_fill="#0f0f1a",
-    block_background_fill="#1a1a2e",
-    block_border_color="#7c3aed",
-    input_background_fill="#16213e",
-    button_primary_background_fill="linear-gradient(135deg, #7c3aed, #db2777)",
-    button_primary_text_color="white",
+THEME = gr.themes.Soft(primary_hue="violet", neutral_hue="slate").set(
+    body_background_fill="#0b0f19", block_background_fill="#161b22", 
+    button_primary_background_fill="linear-gradient(90deg, #7c3aed, #db2777)"
 )
 
-CSS = """
-h1 { text-align:center; color:#c084fc; font-size:2rem; margin-bottom:4px; }
-.subtitle { text-align:center; color:#94a3b8; margin-bottom:1rem; }
-.tab-nav button { font-weight:600; }
-"""
-
-with gr.Blocks(theme=THEME, css=CSS, title="Studio Privado NSFW") as demo:
-    gr.HTML("<h1>🔥 Studio Privado</h1>")
-    gr.HTML('<p class="subtitle">Generador multimedia sin censura · Tus creaciones son privadas y solo las ves vos</p>')
-
+with gr.Blocks(theme=THEME, title="Studio Privado V2") as demo:
+    gr.HTML("<h1 style='text-align:center; color:#db2777;'>🔥 Studio Privado V2 - Ultra Realismo</h1>")
+    
     with gr.Tabs():
-        # ── TEXT → IMAGE ──────────────────────────────────────────────────────
-        with gr.Tab("🖼 Text → Image"):
-            with gr.Row():
-                with gr.Column():
-                    t2i_prompt = gr.Textbox(label="Prompt", lines=3,
-                        placeholder="A beautiful woman, photorealistic, 8k, detailed...")
-                    t2i_neg    = gr.Textbox(label="Negative Prompt", value=NEG_DEFAULT, lines=2)
-                    with gr.Row():
-                        t2i_steps = gr.Slider(10, 60, 30, step=1, label="Pasos")
-                        t2i_cfg   = gr.Slider(1, 20, 7.5, step=0.5, label="CFG")
-                    with gr.Row():
-                        t2i_w   = gr.Slider(512, 1280, 1024, step=64, label="Ancho")
-                        t2i_h   = gr.Slider(512, 1280, 1024, step=64, label="Alto")
-                    t2i_seed  = gr.Number(42, label="Seed")
-                    t2i_btn   = gr.Button("🚀 Generar Imagen", variant="primary", size="lg")
-                with gr.Column():
-                    t2i_out = gr.Image(label="Resultado", type="pil", height=500)
-            t2i_btn.click(text2img,
-                [t2i_prompt, t2i_neg, t2i_steps, t2i_cfg, t2i_w, t2i_h, t2i_seed],
-                t2i_out)
-
-        # ── IMAGE → IMAGE ─────────────────────────────────────────────────────
-        with gr.Tab("🔄 Image → Image"):
+        with gr.Tab("🖼 Imagen (Text2Img)"):
             with gr.Row():
                 with gr.Column():
-                    i2i_input  = gr.Image(label="Imagen Base", type="numpy")
-                    i2i_prompt = gr.Textbox(label="Prompt", lines=3,
-                        placeholder="Modify the image to...")
-                    i2i_neg    = gr.Textbox(label="Negative Prompt", value=NEG_DEFAULT, lines=2)
+                    t2i_p = gr.Textbox(label="Prompt", lines=3, placeholder="A photorealistic woman...")
+                    t2i_n = gr.Textbox(label="Negativo", value=NEG_DEFAULT)
                     with gr.Row():
-                        i2i_str   = gr.Slider(0.1, 1.0, 0.6, step=0.05, label="Intensidad")
-                        i2i_steps = gr.Slider(10, 60, 30, step=1, label="Pasos")
+                        t2i_lora = gr.Textbox(label="Custom LoRA ID", placeholder="ej: John6666/nsfw-master-flux-lora-merged")
+                        t2i_ls = gr.Slider(0, 1.5, 0.8, label="Fuerza LoRA")
                     with gr.Row():
-                        i2i_cfg  = gr.Slider(1, 20, 7.5, step=0.5, label="CFG")
-                        i2i_seed = gr.Number(42, label="Seed")
-                    i2i_btn = gr.Button("🚀 Transformar Imagen", variant="primary", size="lg")
-                with gr.Column():
-                    i2i_out = gr.Image(label="Resultado", type="pil", height=500)
-            i2i_btn.click(img2img,
-                [i2i_prompt, i2i_neg, i2i_input, i2i_str, i2i_steps, i2i_cfg, i2i_seed],
-                i2i_out)
+                        t2i_w = gr.Slider(512, 1280, 1024, step=64, label="Ancho")
+                        t2i_h = gr.Slider(512, 1280, 1024, step=64, label="Alto")
+                    t2i_btn = gr.Button("Generar Imagen", variant="primary")
+                t2i_out = gr.Image(label="Resultado")
+            t2i_btn.click(generate_t2i, [t2i_p, t2i_n, t2i_lora, t2i_ls, gr.Number(30), gr.Number(7.5), t2i_w, t2i_h, gr.Number(42)], t2i_out)
 
-        # ── TEXT → VIDEO (LTX) ────────────────────────────────────────────────
-        with gr.Tab("🎬 Text → Video (LTX)"):
-            gr.Markdown("*Usa LTX-Video con LoRA NSFW. Genera entre 25 y 121 frames.*")
+        with gr.Tab("🎬 Video (LTX-Video)"):
             with gr.Row():
                 with gr.Column():
-                    t2v_prompt = gr.Textbox(label="Prompt", lines=3,
-                        placeholder="A woman walking in slow motion, cinematic, 4k...")
-                    t2v_neg    = gr.Textbox(label="Negative Prompt", value=NEG_DEFAULT, lines=2)
-                    with gr.Row():
-                        t2v_frames = gr.Slider(25, 121, 49, step=8, label="Frames")
-                        t2v_fps    = gr.Slider(8, 30, 24, step=1, label="FPS")
-                    with gr.Row():
-                        t2v_w  = gr.Slider(256, 768, 512, step=64, label="Ancho")
-                        t2v_h  = gr.Slider(256, 768, 512, step=64, label="Alto")
-                    with gr.Row():
-                        t2v_steps = gr.Slider(10, 50, 30, step=1, label="Pasos")
-                        t2v_lora  = gr.Slider(0.0, 1.5, 0.8, step=0.05, label="LoRA Scale")
-                    t2v_seed = gr.Number(42, label="Seed")
-                    t2v_btn  = gr.Button("🎬 Generar Video", variant="primary", size="lg")
-                with gr.Column():
-                    t2v_out = gr.Video(label="Video Generado")
-            t2v_btn.click(text2video,
-                [t2v_prompt, t2v_neg, t2v_frames, t2v_fps, t2v_w, t2v_h, t2v_steps, t2v_lora, t2v_seed],
-                t2v_out)
-
-        # ── IMAGE → VIDEO (LTX) ───────────────────────────────────────────────
-        with gr.Tab("📹 Image → Video (LTX)"):
-            gr.Markdown("*Animá una imagen usando LTX-Video + LoRA NSFW.*")
-            with gr.Row():
-                with gr.Column():
-                    i2v_input  = gr.Image(label="Imagen Base (se usará como frame inicial)", type="numpy")
-                    i2v_prompt = gr.Textbox(label="Prompt de movimiento", lines=3,
-                        placeholder="The woman slowly turns her head, smooth motion...")
-                    i2v_neg    = gr.Textbox(label="Negative Prompt", value=NEG_DEFAULT, lines=2)
-                    with gr.Row():
-                        i2v_frames = gr.Slider(25, 121, 49, step=8, label="Frames")
-                        i2v_fps    = gr.Slider(8, 30, 24, step=1, label="FPS")
-                    with gr.Row():
-                        i2v_steps = gr.Slider(10, 50, 30, step=1, label="Pasos")
-                        i2v_lora  = gr.Slider(0.0, 1.5, 0.8, step=0.05, label="LoRA Scale")
-                    i2v_seed = gr.Number(42, label="Seed")
-                    i2v_btn  = gr.Button("📹 Animar Imagen", variant="primary", size="lg")
-                with gr.Column():
-                    i2v_out = gr.Video(label="Video Generado")
-            i2v_btn.click(img2video,
-                [i2v_prompt, i2v_neg, i2v_input, i2v_frames, i2v_fps, i2v_steps, i2v_lora, i2v_seed],
-                i2v_out)
+                    v_p = gr.Textbox(label="Prompt de Video", lines=3)
+                    v_img = gr.Image(label="Imagen Base (opcional)", type="numpy")
+                    v_ls = gr.Slider(0, 1.5, 0.8, label="Fuerza LoRA NSFW")
+                    v_btn = gr.Button("Generar Video", variant="primary")
+                v_out = gr.Video(label="Resultado")
+            v_btn.click(generate_video, [v_p, gr.Textbox(value=NEG_DEFAULT), v_img, gr.Number(49), gr.Number(24), gr.Number(30), v_ls, gr.Number(42)], v_out)
 
 demo.launch()