app.py

import os
import subprocess
import tempfile

import cv2
import gradio as gr
import numpy as np
import spaces
import torch
import torch.nn.functional as F
from huggingface_hub import hf_hub_download
from imageio_ffmpeg import get_ffmpeg_exe

from train_log.IFNet_HDv3 import IFNet
from upsampler_theme import UPSAMPLER_THEME, UPSAMPLER_CSS, footer_html, header_html

# Official Practical-RIFE v4.25 weights (MIT), mirrored unchanged on our org
# so no third-party repo can drift underneath the Space.
WEIGHTS = hf_hub_download("Upsampler/rife-4-25", "flownet.pkl")

MAX_SECONDS = 6.5
MAX_SIDE = 1280
# Consecutive frames more different than this (mean abs diff on gray
# thumbnails) are treated as a hard cut: duplicate instead of interpolate,
# so cuts don't produce ghost blends.
SCENE_CUT_DIFF = 60.0

flownet = IFNet()
state = torch.load(WEIGHTS, map_location="cpu")
flownet.load_state_dict(
    {k.replace("module.", ""): v for k, v in state.items() if "module." in k},
    strict=False,
)
flownet = flownet.to("cuda").eval()


def probe(path):
    cap = cv2.VideoCapture(path)
    fps = cap.get(cv2.CAP_PROP_FPS) or 30
    n = int(cap.get(cv2.CAP_PROP_FRAME_COUNT))
    w = int(cap.get(cv2.CAP_PROP_FRAME_WIDTH))
    h = int(cap.get(cv2.CAP_PROP_FRAME_HEIGHT))
    cap.release()
    return fps, n, w, h


def get_duration(video, factor, progress=None):
    if not video:
        return 10
    try:
        fps, n, w, h = probe(video)
    except Exception:
        return 60
    mult = int(factor[0]) if factor else 2
    # ~35ms per generated frame at 720p on the ZeroGPU slice, plus decode,
    # encode, and model-to-GPU overhead; never over-request.
    est = 20 + int(n * (mult - 1) * 0.06 * max(1.0, (w * h) / (1280 * 720)))
    return min(110, est)


def midpoint(img0, img1):
    imgs = torch.cat((img0, img1), 1)
    scale_list = [16, 8, 4, 2, 1]
    _, _, merged = flownet(imgs, 0.5, scale_list)
    return merged[-1]


@spaces.GPU(duration=get_duration)
def interpolate(video, factor, progress=gr.Progress()):
    if not video:
        raise gr.Error("Please upload a video first.")
    fps, n, w, h = probe(video)
    seconds = n / max(fps, 1)
    if seconds > MAX_SECONDS:
        raise gr.Error(
            f"This clip is {seconds:.1f}s; the free tool accepts up to {MAX_SECONDS:.0f}s. "
            "Trim it first and try again."
        )
    if max(w, h) > MAX_SIDE:
        raise gr.Error(
            f"This clip is {w}x{h}; the free tool accepts up to {MAX_SIDE}px on the longest side."
        )

    mult = int(factor[0])
    cap = cv2.VideoCapture(video)
    frames = []
    ok, frame = cap.read()
    while ok:
        frames.append(frame)
        ok, frame = cap.read()
    cap.release()
    if len(frames) < 2:
        raise gr.Error("Could not read enough frames from this video.")

    ph = ((h - 1) // 64 + 1) * 64
    pw = ((w - 1) // 64 + 1) * 64

    def to_tensor(frame):
        t = torch.from_numpy(frame).to("cuda", non_blocking=True)
        t = t.permute(2, 0, 1).float().unsqueeze(0) / 255.0
        return F.pad(t, (0, pw - w, 0, ph - h))

    def to_frame(t):
        out = (t[0][:, :h, :w].permute(1, 2, 0) * 255.0).clamp(0, 255)
        return out.byte().cpu().numpy()

    thumbs = [
        cv2.cvtColor(cv2.resize(f, (64, 36)), cv2.COLOR_BGR2GRAY).astype(np.float32)
        for f in frames
    ]

    out_path = os.path.join(tempfile.mkdtemp(), "interpolated.mp4")
    ffmpeg = get_ffmpeg_exe()
    enc = subprocess.Popen(
        [
            ffmpeg, "-y", "-f", "rawvideo", "-pix_fmt", "bgr24",
            "-s", f"{w}x{h}", "-r", str(fps * mult), "-i", "-",
            "-i", video, "-map", "0:v", "-map", "1:a?", "-c:a", "copy",
            "-c:v", "libx264", "-preset", "fast", "-crf", "18",
            "-pix_fmt", "yuv420p", "-shortest", out_path,
        ],
        stdin=subprocess.PIPE,
        stderr=subprocess.DEVNULL,
    )

    with torch.inference_mode():
        for i in range(len(frames) - 1):
            progress((i + 1) / len(frames), desc="Interpolating frames")
            enc.stdin.write(frames[i].tobytes())
            is_cut = float(np.abs(thumbs[i] - thumbs[i + 1]).mean()) > SCENE_CUT_DIFF
            if is_cut:
                for _ in range(mult - 1):
                    enc.stdin.write(frames[i].tobytes())
                continue
            t0, t1 = to_tensor(frames[i]), to_tensor(frames[i + 1])
            if mult == 2:
                enc.stdin.write(to_frame(midpoint(t0, t1)).tobytes())
            else:  # 4x: recursive midpoints -> quarter timesteps
                mid = midpoint(t0, t1)
                for t in (midpoint(t0, mid), mid, midpoint(mid, t1)):
                    enc.stdin.write(to_frame(t).tobytes())
    enc.stdin.write(frames[-1].tobytes())
    enc.stdin.close()
    if enc.wait() != 0:
        raise gr.Error("Video encoding failed. Please try a different clip.")
    return out_path


with gr.Blocks(theme=UPSAMPLER_THEME, css=UPSAMPLER_CSS) as demo:
    gr.HTML(
        header_html(
            "RIFE Video Frame Interpolation",
            "Double or quadruple your video's frame rate for smooth slow motion",
        )
    )
    with gr.Row():
        with gr.Column(scale=1):
            input_video = gr.Video(label="Video (up to 6s, 1280px)")
            factor = gr.Radio(["2x frames", "4x frames"], value="2x frames", label="Interpolation")
            run = gr.Button("Interpolate", variant="primary")
        with gr.Column(scale=1):
            output_video = gr.Video(label="Result", autoplay=True)
    run.click(interpolate, inputs=[input_video, factor], outputs=output_video, api_name="interpolate")
    gr.HTML(
        footer_html(
            "RIFE (Real-Time Intermediate Flow Estimation, Practical-RIFE v4.25) "
            "generates new in-between frames for any video, turning choppy clips "
            "into smooth slow motion and converting frame rates like 30fps to "
            "60fps or 120fps, with scene-cut detection to avoid ghosting.",
            "https://upsampler.com/free-video-frame-interpolation-no-signup",
            "free video frame interpolation tool",
        )
    )

if __name__ == "__main__":
    demo.launch(ssr_mode=False)