Add main app with CogVideoX-5B I2V + ZeroGPU

app.py

@@ -0,0 +1,262 @@
+"""
+🎬 Image-to-Video AI Studio — Real AI Animation
+Uses CogVideoX-5B-I2V diffusion model to generate REAL animated videos from images.
+Runs on HuggingFace ZeroGPU (free) — actual AI video generation, not zoom/warp tricks.
+"""
+
+import spaces
+import torch
+import gradio as gr
+import tempfile
+import gc
+from PIL import Image
+from diffusers import CogVideoXImageToVideoPipeline
+from diffusers.utils import export_to_video, load_image
+
+# ─── Model Loading (CPU at startup, GPU inside @spaces.GPU) ─────────────────
+print("📦 Loading CogVideoX-5B-I2V pipeline to CPU...")
+pipe = CogVideoXImageToVideoPipeline.from_pretrained(
+    "THUDM/CogVideoX-5b-I2V",
+    torch_dtype=torch.bfloat16,
+)
+print("✅ Pipeline loaded to CPU successfully!")
+
+
+# ─── Prompt Templates ────────────────────────────────────────────────────────
+CAMERA_PROMPTS = {
+    "None (Auto)": "",
+    "Static Shot": "static camera, no camera movement, steady shot",
+    "Slow Zoom In": "camera slowly zooms in, push in shot",
+    "Slow Zoom Out": "camera slowly zooms out, pull back reveal",
+    "Pan Left": "camera pans slowly to the left, horizontal tracking",
+    "Pan Right": "camera pans slowly to the right, horizontal tracking",
+    "Tilt Up": "camera tilts upward slowly, revealing the sky",
+    "Tilt Down": "camera tilts downward slowly",
+    "Orbit Shot": "camera orbits around the subject slowly",
+    "Dolly Forward": "camera moves forward smoothly, dolly shot",
+    "Crane Up": "camera rises upward, crane shot, aerial movement",
+    "Handheld": "slight handheld camera movement, natural sway",
+}
+
+STYLE_PROMPTS = {
+    "Cinematic": "cinematic lighting, film grain, dramatic, movie quality, 4K",
+    "Realistic": "photorealistic, natural lighting, high detail, sharp focus",
+    "Dreamy": "soft focus, ethereal glow, dreamy atmosphere, gentle lighting",
+    "Dynamic": "fast motion, energetic, vibrant, dynamic movement",
+    "Slow Motion": "slow motion, graceful movement, smooth, elegant",
+    "Nature Documentary": "nature documentary style, wildlife, natural beauty, BBC quality",
+    "Anime / Artistic": "anime style, artistic, vibrant colors, stylized movement",
+}
+
+DURATION_FRAMES = {
+    "~3 seconds (24 frames)": 24,
+    "~6 seconds (49 frames)": 49,
+}
+
+
+# ─── Core Video Generation ───────────────────────────────────────────────────
+@spaces.GPU(duration=300)
+def generate_video(
+    image,
+    prompt,
+    camera_motion,
+    style,
+    duration,
+    num_inference_steps,
+    guidance_scale,
+    seed,
+    use_random_seed,
+    progress=gr.Progress(track_tqdm=True),
+):
+    """Generate real animated video using CogVideoX-5B-I2V diffusion model."""
+
+    if image is None:
+        raise gr.Error("⚠️ Please upload an image first!")
+
+    if not prompt or prompt.strip() == "":
+        raise gr.Error("⚠️ Please describe what animation you want!")
+
+    # Move pipeline to GPU
+    pipe.to("cuda")
+    pipe.vae.enable_tiling()
+    pipe.vae.enable_slicing()
+
+    # Build the full prompt
+    parts = [prompt.strip()]
+    camera_desc = CAMERA_PROMPTS.get(camera_motion, "")
+    if camera_desc:
+        parts.append(camera_desc)
+    style_desc = STYLE_PROMPTS.get(style, "")
+    if style_desc:
+        parts.append(style_desc)
+    parts.append("high quality, detailed, smooth motion")
+    full_prompt = ". ".join(parts)
+
+    print(f"🎬 Prompt: {full_prompt}")
+
+    # Seed
+    if use_random_seed:
+        seed = torch.randint(0, 2**32, (1,)).item()
+    generator = torch.Generator(device="cuda").manual_seed(int(seed))
+
+    # Frames
+    num_frames = DURATION_FRAMES.get(duration, 49)
+
+    # Resize image (CogVideoX native: 720x480 landscape, 480x720 portrait)
+    if isinstance(image, str):
+        image = load_image(image)
+    img_w, img_h = image.size
+    if img_w >= img_h:
+        target_w, target_h = 720, 480
+    else:
+        target_w, target_h = 480, 720
+    image_resized = image.resize((target_w, target_h), Image.LANCZOS)
+
+    # Generate
+    print(f"🚀 Generating: {num_frames} frames, {num_inference_steps} steps")
+    result = pipe(
+        prompt=full_prompt,
+        image=image_resized,
+        num_videos_per_prompt=1,
+        num_inference_steps=int(num_inference_steps),
+        num_frames=num_frames,
+        guidance_scale=float(guidance_scale),
+        generator=generator,
+    ).frames[0]
+
+    # Export to MP4
+    out_path = tempfile.mktemp(suffix=".mp4")
+    export_to_video(result, out_path, fps=8)
+
+    # Cleanup
+    pipe.to("cpu")
+    gc.collect()
+    torch.cuda.empty_cache()
+
+    print(f"✅ Video saved: {out_path}")
+    return out_path, f"🎲 Seed: {seed} | Frames: {num_frames} | Steps: {num_inference_steps}"
+
+
+# ─── Gradio UI ───────────────────────────────────────────────────────────────
+CSS = """
+.main-header { text-align: center; padding: 20px 0 10px 0; }
+.main-header h1 {
+    background: linear-gradient(135deg, #667eea 0%, #764ba2 100%);
+    -webkit-background-clip: text; -webkit-text-fill-color: transparent;
+    font-size: 2.2em; margin-bottom: 5px;
+}
+.note-box {
+    background: linear-gradient(135deg, #fff3cd 0%, #ffeeba 100%);
+    border-left: 4px solid #ffc107; border-radius: 8px;
+    padding: 12px 16px; margin: 10px 0;
+}
+.info-box {
+    background: linear-gradient(135deg, #f5f7fa 0%, #c3cfe2 100%);
+    border-radius: 10px; padding: 15px; margin: 10px 0;
+}
+footer { display: none !important; }
+"""
+
+with gr.Blocks(title="🎬 Image-to-Video AI") as demo:
+
+    gr.HTML("""
+    <div class="main-header">
+        <h1>🎬 Image-to-Video AI</h1>
+        <p style="color: #666; font-size: 1.15em;">
+            Transform any image into a <b>real animated video</b> with AI
+        </p>
+        <p style="color: #999; font-size: 0.85em;">
+            CogVideoX-5B · Free ZeroGPU · Real Diffusion Animation
+        </p>
+    </div>
+    """)
+
+    gr.HTML("""
+    <div class="note-box">
+        ⚡ <b>Real AI Animation:</b> This uses a 5-billion parameter diffusion model to generate 
+        <b>actual new video frames</b> — not zoom/pan tricks. Water flows, hair blows, 
+        people walk, clouds drift. Takes ~2-4 min per video on free GPU.
+    </div>
+    """)
+
+    with gr.Row():
+        with gr.Column(scale=1):
+            image_input = gr.Image(label="📸 Upload Image", type="pil", height=320)
+
+            prompt_input = gr.Textbox(
+                label="✍️ Animation Prompt — describe what should move",
+                placeholder="e.g., 'The river flows gently, trees sway in the breeze, clouds drift across the sky'",
+                lines=3,
+            )
+
+            with gr.Row():
+                camera_motion = gr.Dropdown(
+                    choices=list(CAMERA_PROMPTS.keys()),
+                    value="None (Auto)", label="🎥 Camera Motion",
+                )
+                style = gr.Dropdown(
+                    choices=list(STYLE_PROMPTS.keys()),
+                    value="Cinematic", label="🎭 Style",
+                )
+
+            duration = gr.Radio(
+                choices=list(DURATION_FRAMES.keys()),
+                value="~6 seconds (49 frames)", label="⏱️ Duration",
+            )
+
+            with gr.Accordion("⚙️ Advanced", open=False):
+                num_inference_steps = gr.Slider(
+                    minimum=10, maximum=50, value=30, step=5,
+                    label="🔄 Steps (15=fast, 30=good, 50=best)",
+                )
+                guidance_scale = gr.Slider(
+                    minimum=1.0, maximum=12.0, value=6.0, step=0.5,
+                    label="🎯 Guidance Scale",
+                )
+                with gr.Row():
+                    seed = gr.Number(label="🎲 Seed", value=42, precision=0)
+                    use_random_seed = gr.Checkbox(label="Random", value=True)
+
+            generate_btn = gr.Button("🎬 Generate Video", variant="primary", size="lg")
+
+        with gr.Column(scale=1):
+            video_output = gr.Video(label="🎥 Generated Video", autoplay=True, loop=True, height=420)
+            gen_info = gr.Textbox(label="Info", interactive=False)
+
+            gr.HTML("""
+            <div class="info-box">
+                <b>🧠 How it works:</b>
+                <ul style="margin: 5px 0; padding-left: 18px; color: #555;">
+                    <li>CogVideoX-5B generates <b>real new video frames</b> using diffusion</li>
+                    <li>Objects actually move — water, people, animals, particles</li>
+                    <li>Your prompt controls <b>what</b> animates and <b>how</b></li>
+                    <li>Camera motion and style are added to the AI prompt</li>
+                </ul>
+                <p style="color: #888; font-size: 0.85em; margin-top: 8px;">
+                    <a href="https://huggingface.co/THUDM/CogVideoX-5b-I2V" target="_blank">CogVideoX-5B-I2V</a> 
+                    by Tsinghua University · ZeroGPU (Free A10G)
+                </p>
+            </div>
+            """)
+
+    generate_btn.click(
+        fn=generate_video,
+        inputs=[image_input, prompt_input, camera_motion, style, duration,
+                num_inference_steps, guidance_scale, seed, use_random_seed],
+        outputs=[video_output, gen_info],
+    )
+
+    gr.Markdown("### 💡 Example Prompts")
+    gr.Markdown("""
+| Image Type | Example Prompt |
+|-----------|---------------|
+| 🏞️ Landscape | *"The river flows gently, clouds drift, trees sway in breeze"* |
+| 👤 Portrait | *"She smiles and turns her head, hair flowing in the wind"* |
+| 🌊 Ocean | *"Waves crash on shore, sea foam bubbles, seagulls fly"* |
+| 🏙️ City | *"Cars drive, pedestrians walk, city lights flicker"* |
+| 🐾 Animals | *"The cat stretches, yawns, then walks forward"* |
+| 🌸 Nature | *"Petals sway in breeze, butterfly lands on flower"* |
+    """)
+
+if __name__ == "__main__":
+    demo.launch(css=CSS)