Update app.py

app.py

@@ -2,110 +2,101 @@ import gradio as gr
 import numpy as np
 import random
 import torch
-from openvino.runtime import Core
 from PIL import Image
-from huggingface_hub import hf_hub_download
-import os
 
-print("🔧 Initializing OpenVINO Runtime...")
+from transformers import CLIPTextModel, CLIPTokenizer
+from diffusers import AutoencoderKL, UNet2DConditionModel
+from optimum.intel import OVStableDiffusionPipeline
 
-# --------- Detect OpenVINO Device ---------
-core = Core()
-available_devices = core.available_devices
-device = "CPU"
+# Nếu muốn đảm bảo reproducibility
+torch.manual_seed(0)
+np.random.seed(0)
+random.seed(0)
 
-if "GPU" in available_devices:
-    device = "GPU"
-elif "MYRIAD" in available_devices:
-    device = "MYRIAD"
-elif "HDDL" in available_devices:
-    device = "HDDL"
+print("🔧 Loading OpenVINO pipeline: HARRY07979/sd-v1-5-openvino")
 
-print(f"✅ Using device: {device}")
+# Load the full pipeline (tokenizer, text encoder, unet, vae) từ repo
+pipeline = OVStableDiffusionPipeline.from_pretrained(
+    "HARRY07979/sd-v1-5-openvino",
+    # nếu model đã convert sẵn IR nên không cần export
+    safety_checker=None,  # nếu muốn bỏ kiểm tra NSFW hoặc model không có
+    feature_extractor=None,
+    torch_dtype=torch.float32,  # nhớ kiểm tra model hỗ trợ dtype nào
+)
 
-# --------- Download model from HF Hub ---------
-model_repo = "HARRY07979/sd-v1-5-openvino"
-model_filename = "stable-diffusion-v1-5.xml"  # you may need to adjust based on actual file
+# Chọn thiết bị phù hợp (CPU, GPU, MYRIAD...) — Optimum Intel sẽ tự chọn nếu có hỗ trợ
+pipeline.to("cpu")
 
-model_path_xml = hf_hub_download(repo_id=model_repo, filename=model_filename)
-model_path_bin = model_path_xml.replace(".xml", ".bin")
-
-print("📦 Model files downloaded")
-
-# --------- Compile OpenVINO model ---------
-compiled_model = core.compile_model(model_path_xml, device)
-
-# Get input and output tensor names
-input_layer = compiled_model.input(0)
-output_layer = compiled_model.output(0)
-
-# --------- Dummy NSFW filter ---------
-def detect_nsfw(text: str):
+# Dummy NSFW detection nếu muốn (bên trong có thể dùng safety_checker của pipeline nếu có)
+def detect_nsfw(prompt: str):
     banned = ["nude", "sex", "porn", "xxx", "nsfw"]
     for word in banned:
-        if word in text.lower():
-            return "NSFW"
-    return "SFW"
+        if word.lower() in prompt.lower():
+            return True
+    return False
 
-# --------- Dummy Preprocessing (you should integrate tokenizer/text encoder for real use) ---------
-def preprocess_input(prompt, width, height, seed):
-    # Just generate dummy latent vector
-    np.random.seed(seed)
-    return np.random.rand(1, 3, height, width).astype(np.float32)
-
-# --------- Inference function ---------
 def infer(
-    prompt,
-    negative_prompt,
-    seed,
-    randomize_seed,
-    width,
-    height,
-    guidance_scale,
-    num_inference_steps,
+    prompt: str,
+    negative_prompt: str,
+    seed: int,
+    randomize_seed: bool,
+    width: int,
+    height: int,
+    guidance_scale: float,
+    num_inference_steps: int,
     progress=gr.Progress(track_tqdm=True),
 ):
-    if detect_nsfw(prompt) == "NSFW":
+    if detect_nsfw(prompt):
         raise gr.Error("⚠️ Prompt contains NSFW content. Please use a safe prompt.")
 
     if randomize_seed:
         seed = random.randint(0, np.iinfo(np.int32).max)
 
+    # Ensure width/height divisible by 8
     width = (width // 8) * 8
     height = (height // 8) * 8
 
-    # Prepare dummy input for now
-    input_data = preprocess_input(prompt, width, height, seed)
-
-    # Run OpenVINO inference
-    result = compiled_model([input_data])[output_layer]
+    generator = torch.Generator("cpu").manual_seed(seed)
 
-    # Postprocess dummy result
-    output_image = np.clip(result[0].transpose(1, 2, 0), 0, 1)
-    output_image = (output_image * 255).astype(np.uint8)
-    pil_image = Image.fromarray(output_image)
+    result = pipeline(
+        prompt=prompt,
+        negative_prompt=negative_prompt,
+        guidance_scale=guidance_scale,
+        num_inference_steps=num_inference_steps,
+        width=width,
+        height=height,
+        generator=generator,
+    )
 
-    return pil_image, seed
+    image = result.images[0]
+    return image, seed
 
-# --------- UI ---------
+# Gradio UI
 examples = [
-    "A futuristic cyberpunk city at night, ultra-detailed, 4k",
-    "A cute corgi wearing sunglasses",
-    "A bowl of ramen in anime style",
+    "A fantasy landscape, vivid colors, sunset light",
+    "Portrait of a cyberpunk robot girl, neon lighting",
+    "An epic sci-fi scene: spaceship battle in space",
 ]
 
-with gr.Blocks(css="#col-container { max-width: 640px; margin: auto; }") as demo:
+css = """
+#col-container {
+    margin: 0 auto;
+    max-width: 640px;
+}
+"""
+
+with gr.Blocks(css=css) as demo:
     with gr.Column(elem_id="col-container"):
-        gr.Markdown("## 🧠 SD 1.5 OpenVINO via HARRY07979 on HuggingFace")
+        gr.Markdown("## Stable Diffusion v1.5 OpenVINO — Full Pipeline (Tokeniser + UNet + VAE)")
 
         with gr.Row():
             prompt = gr.Text(label="Prompt", placeholder="Enter your prompt", show_label=False)
             run_button = gr.Button("Generate", variant="primary")
 
-        result = gr.Image(label="Output", show_label=False)
+        result = gr.Image(label="Generated Image", show_label=False)
 
         with gr.Accordion("Advanced Settings", open=False):
-            negative_prompt = gr.Text(label="Negative Prompt", placeholder="e.g. low quality, blurry")
+            negative_prompt = gr.Text(label="Negative prompt", placeholder="Enter negative prompt")
 
             seed = gr.Slider(label="Seed", minimum=0, maximum=np.iinfo(np.int32).max, step=1, value=0)
             randomize_seed = gr.Checkbox(label="Randomize Seed", value=True)