app.py

"""
Sentinel Tiny Video Space — Frame interpolation with temporal Sentinel conv
"""
import gradio as gr
import torch
import torch.nn as nn
import numpy as np
from PIL import Image
from diffusers import UNet2DModel
import json

# ─── Sentinel Components ─────────────────────────────────────────────────────
class SentinelAct(nn.Module):
    def __init__(self):
        super().__init__()
        self.inv_e = 1.0 / np.e
    def forward(self, x):
        return x * (1.0 / torch.cosh(self.inv_e * x))

class SentinelVideoModel(nn.Module):
    def __init__(self):
        super().__init__()
        self.spatial = UNet2DModel(
            sample_size=32, in_channels=3, out_channels=3, layers_per_block=1,
            block_out_channels=(32, 64), down_block_types=("DownBlock2D",)*2,
            up_block_types=("UpBlock2D",)*2, time_embedding_type="positional",
        )
        self.temporal = nn.Conv3d(3, 3, kernel_size=(3, 1, 1), padding=(1, 0, 0))
        self.inv_e = 1.0 / np.e
    def forward(self, x, t):
        B, C, F, H, W = x.shape
        out = []
        for f in range(F):
            of = self.spatial(x[:, :, f, :, :], t).sample
            out.append(of)
        stack = torch.stack(out, dim=2)
        temp = self.temporal(stack)
        return temp * (1.0 / torch.cosh(self.inv_e * temp))

# ─── Load Model ──────────────────────────────────────────────────────────────
model_status = "⏳ Loading model..."
video_model = None

try:
    from huggingface_hub import hf_hub_download
    model_path = hf_hub_download(repo_id="5dimension/sentinel-tiny-video", filename="model.pt")
    video_model = SentinelVideoModel()
    video_model.load_state_dict(torch.load(model_path, map_location="cpu", weights_only=True))
    video_model.eval()
    model_status = "✅ Model loaded — 30 trainable params on frozen 2D UNet"
except Exception as e:
    video_model = SentinelVideoModel()
    video_model.eval()
    model_status = f"⚠️ Using fresh weights: {str(e)[:100]}"

# ─── Frame Interpolation ──────────────────────────────────────────────────────
def interpolate_video(frame1, frame2, steps=8):
    if video_model is None:
        return [Image.new('RGB', (256, 256), color='gray') for _ in range(steps)]
    
    # Convert PIL to tensor
    def pil_to_tensor(img):
        arr = np.array(img.resize((32, 32)).convert('RGB')) / 255.0
        return torch.from_numpy(arr).permute(2, 0, 1).float()
    
    f1 = pil_to_tensor(frame1)
    f2 = pil_to_tensor(frame2)
    
    # Create interpolation frames
    interpolated = [frame1.resize((256, 256), Image.BILINEAR)]
    
    with torch.no_grad():
        for i in range(1, steps):
            alpha = i / steps
            # Blend frames
            mid_blend = (1 - alpha) * f1 + alpha * f2
            
            # Create triplet
            triplet = torch.stack([f1, mid_blend, f2], dim=0).unsqueeze(0)  # [1, F, C, H, W]
            triplet = triplet.permute(0, 2, 1, 3, 4)  # [1, C, F, H, W]
            
            t = torch.tensor([500], device="cpu").long()
            out = video_model(triplet, t)
            
            # Extract middle frame
            mid = out[0, :, 1, :, :].permute(1, 2, 0).cpu().numpy()
            mid = np.clip(mid, 0, 1)
            mid = (mid * 255).astype(np.uint8)
            
            img = Image.fromarray(mid).resize((256, 256), Image.BILINEAR)
            interpolated.append(img)
        
        interpolated.append(frame2.resize((256, 256), Image.BILINEAR))
    
    return interpolated

# ─── UI ────────────────────────────────────────────────────────────────────────
with gr.Blocks(title="🎬 Sentinel Tiny Video", css="""
    .gradio-container { max-width: 900px; margin: 0 auto; }
    .title { text-align: center; font-size: 2em; font-weight: bold; color: #6b4c9a; }
    .subtitle { text-align: center; color: #888; margin-bottom: 1em; }
""") as demo:
    gr.Markdown("""
    <div class="title">🎬 Sentinel Tiny Video</div>
    <div class="subtitle">30-parameter frame interpolation with Sentinel temporal convolution</div>
    """)
    
    gr.Markdown(f"**Status**: {model_status}")
    
    with gr.Row():
        with gr.Column(scale=1):
            frame1 = gr.Image(label="Frame 1", type="pil", value=None)
            frame2 = gr.Image(label="Frame 2", type="pil", value=None)
        with gr.Column(scale=1):
            steps = gr.Slider(3, 16, value=8, step=1, label="Interpolation Steps")
            generate_btn = gr.Button("🎬 Interpolate", variant="primary")
    
    gallery = gr.Gallery(label="Interpolated Frames", columns=4, height=200)
    
    with gr.Row():
        gr.Markdown("""
        ### About
        - **Architecture**: Frozen 2D UNet + trainable 3D temporal conv
        - **Trainable Params**: 30 (yes, thirty)
        - **Activation**: Sentinel sech on temporal features
        - **Dataset**: CIFAR-10 frame triplets
        - **Input**: Two 32×32 keyframes → smooth transition
        - **Full model**: [sentinel-tiny-video](https://huggingface.co/5dimension/sentinel-tiny-video)
        """)
    
    generate_btn.click(interpolate_video, [frame1, frame2, steps], gallery)

demo.launch()