ffmpeg-opencv-integration

Quick Reference

Task	Best Library	Why
Simple video read	OpenCV cv2.VideoCapture	Built-in, easy API
GPU video I/O	ffmpegcv	NVDEC/NVENC, OpenCV-compatible API
Multi-threaded streaming	VidGear	RTSP/RTMP, camera capture
ML batch loading	Decord	2x faster than OpenCV, batch GPU decode
Frame-level precision	PyAV	Direct libav access, precise seeking
Complex filter graphs	ffmpeg-python subprocess	Full FFmpeg power

Color Format	Library	Conversion
BGR	OpenCV (cv2)	cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
RGB	FFmpeg, PIL, PyAV	cv2.cvtColor(img, cv2.COLOR_RGB2BGR)
YUV	FFmpeg internal	Convert to RGB/BGR for processing

When to Use This Skill

Use for FFmpeg + OpenCV combined workflows:

• Reading video with FFmpeg, processing frames with OpenCV
• Piping frames between FFmpeg and OpenCV processes
• GPU-accelerated video I/O with OpenCV processing
• Fixing color format mismatches (BGR vs RGB)
• Memory-efficient processing of large videos
• Parallel frame processing on Modal.com

---

FFmpeg + OpenCV Integration Guide

Complete guide to combining FFmpeg's video I/O power with OpenCV's image processing capabilities.

Critical Gotchas (Must Know!)

1. Color Format Mismatch (BGR vs RGB)

The #1 source of bugs in FFmpeg + OpenCV pipelines.

import cv2
import numpy as np

# OpenCV uses BGR by default
img_bgr = cv2.imread("image.jpg")  # BGR format!

# FFmpeg outputs RGB by default (in most configs)
# PyAV outputs RGB
# PIL/Pillow uses RGB

# WRONG: Using FFmpeg RGB output directly with OpenCV
# Colors will be swapped (red becomes blue)

# CORRECT: Always convert explicitly
def bgr_to_rgb(frame: np.ndarray) -> np.ndarray:
    """Convert OpenCV BGR to RGB for FFmpeg/PIL/ML."""
    return cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)

def rgb_to_bgr(frame: np.ndarray) -> np.ndarray:
    """Convert RGB (FFmpeg/PyAV/PIL) to OpenCV BGR."""
    return cv2.cvtColor(frame, cv2.COLOR_RGB2BGR)

# When piping FFmpeg to OpenCV:
def read_ffmpeg_frame_for_opencv(raw_bytes: bytes, width: int, height: int) -> np.ndarray:
    """Read FFmpeg raw frame and convert to OpenCV BGR."""
    # FFmpeg rawvideo is RGB by default
    frame_rgb = np.frombuffer(raw_bytes, dtype=np.uint8).reshape(height, width, 3)
    # Convert to BGR for OpenCV
    frame_bgr = cv2.cvtColor(frame_rgb, cv2.COLOR_RGB2BGR)
    return frame_bgr

2. Frame Dimension Order (y, x) vs (x, y)

OpenCV and NumPy use (row, col) = (y, x), not (x, y).

import cv2
import numpy as np

img = cv2.imread("image.jpg")
print(img.shape)  # (height, width, channels) = (y, x, c)

# WRONG: Accessing pixel at x=100, y=200
# pixel = img[100, 200]  # This gets row=100, col=200 = (y=100, x=200)

# CORRECT: Accessing pixel at x=100, y=200
pixel = img[200, 100]  # row=200 (y), col=100 (x)

# WRONG: Creating image with width=1920, height=1080
# img = np.zeros((1920, 1080, 3))  # Creates 1920 rows × 1080 cols!

# CORRECT: Creating 1920×1080 image
img = np.zeros((1080, 1920, 3), dtype=np.uint8)  # (height, width, channels)

# When reading frame dimensions:
cap = cv2.VideoCapture("video.mp4")
width = int(cap.get(cv2.CAP_PROP_FRAME_WIDTH))   # x dimension
height = int(cap.get(cv2.CAP_PROP_FRAME_HEIGHT))  # y dimension

# NumPy array will be: frame.shape = (height, width, 3)

3. Audio Stream Loss with FFmpeg Filters

Video filters drop audio by default in ffmpeg-python.

import ffmpeg

# WRONG: Audio is silently dropped
(
    ffmpeg
    .input('input.mp4')
    .filter('scale', 1280, 720)
    .output('output.mp4')
    .run()
)

# CORRECT: Explicitly handle both streams
input_file = ffmpeg.input('input.mp4')
video = input_file.video.filter('scale', 1280, 720)
audio = input_file.audio  # Preserve audio stream
(
    ffmpeg
    .output(video, audio, 'output.mp4')
    .overwrite_output()
    .run()
)

# CORRECT: For complex pipelines
input_file = ffmpeg.input('input.mp4')
video = (
    input_file.video
    .filter('scale', 1280, 720)
    .filter('fps', fps=30)
)
audio = input_file.audio.filter('loudnorm')
ffmpeg.output(video, audio, 'output.mp4', vcodec='libx264', acodec='aac').run()

4. Memory Leaks with VideoCapture

Always release VideoCapture and destroy windows.

import cv2

# WRONG: No cleanup
cap = cv2.VideoCapture("video.mp4")
while cap.isOpened():
    ret, frame = cap.read()
    if not ret:
        break
    cv2.imshow("Frame", frame)
# Memory leak!

# CORRECT: Use try/finally
cap = cv2.VideoCapture("video.mp4")
try:
    while cap.isOpened():
        ret, frame = cap.read()
        if not ret:
            break
        cv2.imshow("Frame", frame)
        if cv2.waitKey(1) & 0xFF == ord('q'):
            break
finally:
    cap.release()
    cv2.destroyAllWindows()

# CORRECT: Use context manager pattern
class VideoReader:
    def __init__(self, path: str):
        self.cap = cv2.VideoCapture(path)
        if not self.cap.isOpened():
            raise IOError(f"Cannot open video: {path}")

    def __enter__(self):
        return self

    def __exit__(self, *args):
        self.cap.release()

    def frames(self):
        while True:
            ret, frame = self.cap.read()
            if not ret:
                break
            yield frame

# Usage
with VideoReader("video.mp4") as reader:
    for frame in reader.frames():
        # Process frame...
        pass

---

Pattern 1: Pipe FFmpeg to OpenCV

For complex input handling (network streams, unusual formats), use FFmpeg to decode and pipe raw frames to OpenCV.

import subprocess
import numpy as np
import cv2

def ffmpeg_to_opencv_pipe(input_path: str, width: int, height: int):
    """Read video with FFmpeg, process frames with OpenCV."""

    # FFmpeg command to output raw BGR24 frames
    cmd = [
        'ffmpeg',
        '-i', input_path,
        '-f', 'rawvideo',
        '-pix_fmt', 'bgr24',  # BGR for OpenCV!
        '-s', f'{width}x{height}',
        '-'  # Output to stdout
    ]

    # Start FFmpeg process
    process = subprocess.Popen(
        cmd,
        stdout=subprocess.PIPE,
        stderr=subprocess.DEVNULL,  # Suppress FFmpeg logs
        bufsize=10**8  # Large buffer for video
    )

    frame_size = width * height * 3

    try:
        while True:
            raw_frame = process.stdout.read(frame_size)
            if len(raw_frame) != frame_size:
                break

            # Convert to NumPy array (already BGR for OpenCV)
            frame = np.frombuffer(raw_frame, dtype=np.uint8).reshape(height, width, 3)

            # Process with OpenCV
            gray = cv2.cvtColor(frame, cv2.COLOR_BGR2GRAY)
            edges = cv2.Canny(gray, 100, 200)

            yield edges
    finally:
        process.stdout.close()
        process.wait()

# Get video dimensions first
def get_video_dimensions(path: str) -> tuple[int, int]:
    """Get video width and height using ffprobe."""
    cmd = [
        'ffprobe', '-v', 'error',
        '-select_streams', 'v:0',
        '-show_entries', 'stream=width,height',
        '-of', 'csv=p=0',
        path
    ]
    result = subprocess.run(cmd, capture_output=True, text=True)
    width, height = map(int, result.stdout.strip().split(','))
    return width, height

# Usage
width, height = get_video_dimensions("input.mp4")
for edge_frame in ffmpeg_to_opencv_pipe("input.mp4", width, height):
    cv2.imshow("Edges", edge_frame)
    if cv2.waitKey(1) & 0xFF == ord('q'):
        break

---

Pattern 2: OpenCV to FFmpeg Pipe

Process frames with OpenCV, encode output with FFmpeg.

import subprocess
import numpy as np
import cv2

def opencv_to_ffmpeg_pipe(
    input_path: str,
    output_path: str,
    process_frame: callable,
    fps: float = 30.0
):
    """Process video frames with OpenCV, encode with FFmpeg."""

    # Open input with OpenCV
    cap = cv2.VideoCapture(input_path)
    width = int(cap.get(cv2.CAP_PROP_FRAME_WIDTH))
    height = int(cap.get(cv2.CAP_PROP_FRAME_HEIGHT))

    # FFmpeg command to receive raw BGR frames
    cmd = [
        'ffmpeg', '-y',
        '-f', 'rawvideo',
        '-vcodec', 'rawvideo',
        '-s', f'{width}x{height}',
        '-pix_fmt', 'bgr24',  # OpenCV BGR format
        '-r', str(fps),
        '-i', '-',  # Read from stdin
        '-c:v', 'libx264',
        '-preset', 'fast',
        '-crf', '23',
        '-pix_fmt', 'yuv420p',
        output_path
    ]

    process = subprocess.Popen(
        cmd,
        stdin=subprocess.PIPE,
        stderr=subprocess.DEVNULL
    )

    try:
        while cap.isOpened():
            ret, frame = cap.read()
            if not ret:
                break

            # Process frame with user function
            processed = process_frame(frame)

            # Write to FFmpeg
            process.stdin.write(processed.tobytes())
    finally:
        cap.release()
        process.stdin.close()
        process.wait()

# Example usage
def add_blur(frame: np.ndarray) -> np.ndarray:
    return cv2.GaussianBlur(frame, (15, 15), 0)

opencv_to_ffmpeg_pipe("input.mp4", "blurred.mp4", add_blur)

---

Pattern 3: Bidirectional Pipe (FFmpeg ↔ OpenCV ↔ FFmpeg)

For full control over input/output codecs while processing with OpenCV.

import subprocess
import numpy as np
import cv2
from concurrent.futures import ThreadPoolExecutor

def ffmpeg_opencv_ffmpeg_pipeline(
    input_path: str,
    output_path: str,
    process_frame: callable,
    preserve_audio: bool = True
):
    """Complete pipeline: FFmpeg decode → OpenCV process → FFmpeg encode."""

    # Get video info
    probe_cmd = [
        'ffprobe', '-v', 'error',
        '-select_streams', 'v:0',
        '-show_entries', 'stream=width,height,r_frame_rate',
        '-of', 'csv=p=0',
        input_path
    ]
    probe_result = subprocess.run(probe_cmd, capture_output=True, text=True)
    parts = probe_result.stdout.strip().split(',')
    width, height = int(parts[0]), int(parts[1])
    fps_parts = parts[2].split('/')
    fps = int(fps_parts[0]) / int(fps_parts[1])

    # FFmpeg decode command (output BGR24 for OpenCV)
    decode_cmd = [
        'ffmpeg',
        '-i', input_path,
        '-f', 'rawvideo',
        '-pix_fmt', 'bgr24',
        '-'
    ]

    # FFmpeg encode command
    encode_cmd = [
        'ffmpeg', '-y',
        '-f', 'rawvideo',
        '-vcodec', 'rawvideo',
        '-s', f'{width}x{height}',
        '-pix_fmt', 'bgr24',
        '-r', str(fps),
        '-i', '-',
    ]

    # Add audio from original file if preserving
    if preserve_audio:
        encode_cmd.extend(['-i', input_path, '-map', '0:v', '-map', '1:a', '-c:a', 'copy'])

    encode_cmd.extend([
        '-c:v', 'libx264',
        '-preset', 'fast',
        '-crf', '23',
        '-pix_fmt', 'yuv420p',
        output_path
    ])

    # Start processes
    decoder = subprocess.Popen(
        decode_cmd,
        stdout=subprocess.PIPE,
        stderr=subprocess.DEVNULL
    )

    encoder = subprocess.Popen(
        encode_cmd,
        stdin=subprocess.PIPE,
        stderr=subprocess.DEVNULL
    )

    frame_size = width * height * 3

    try:
        while True:
            raw_frame = decoder.stdout.read(frame_size)
            if len(raw_frame) != frame_size:
                break

            # Convert to NumPy, process, convert back
            frame = np.frombuffer(raw_frame, dtype=np.uint8).reshape(height, width, 3)
            processed = process_frame(frame)
            encoder.stdin.write(processed.tobytes())
    finally:
        decoder.stdout.close()
        decoder.wait()
        encoder.stdin.close()
        encoder.wait()

# Usage
def detect_edges(frame: np.ndarray) -> np.ndarray:
    gray = cv2.cvtColor(frame, cv2.COLOR_BGR2GRAY)
    edges = cv2.Canny(gray, 50, 150)
    # Convert back to BGR for encoding
    return cv2.cvtColor(edges, cv2.COLOR_GRAY2BGR)

ffmpeg_opencv_ffmpeg_pipeline("input.mp4", "edges.mp4", detect_edges)

---

ffmpegcv: GPU-Accelerated Video I/O

ffmpegcv provides an OpenCV-compatible API with GPU acceleration (NVDEC/NVENC).

Installation

pip install ffmpegcv

Basic Usage

import ffmpegcv

# Read video (uses NVDEC if available)
cap = ffmpegcv.VideoCapture("video.mp4")

# CPU-only reading
cap = ffmpegcv.VideoCapture("video.mp4", gpu=-1)

# Force GPU 0
cap = ffmpegcv.VideoCapture("video.mp4", gpu=0)

# Read frames (returns BGR like OpenCV!)
while True:
    ret, frame = cap.read()
    if not ret:
        break
    # frame is BGR NumPy array, just like cv2.VideoCapture
    print(frame.shape)  # (height, width, 3)

cap.release()

GPU Video Writing

import ffmpegcv
import cv2
import numpy as np

# GPU-accelerated writing with NVENC
writer = ffmpegcv.VideoWriter(
    "output.mp4",
    codec="h264_nvenc",  # NVIDIA GPU encoding
    fps=30,
    frameSize=(1920, 1080)
)

# Write frames
for i in range(300):
    frame = np.random.randint(0, 255, (1080, 1920, 3), dtype=np.uint8)
    writer.write(frame)  # Accepts BGR like OpenCV

writer.release()

Read Specific Range

import ffmpegcv

# Read frames 100-200 only (efficient seeking)
cap = ffmpegcv.VideoCapture("video.mp4")
cap.set(cv2.CAP_PROP_POS_FRAMES, 100)

for i in range(100):
    ret, frame = cap.read()
    if not ret:
        break
    # Process frame...

cap.release()

Integration with OpenCV Processing

import ffmpegcv
import cv2
import numpy as np

def process_with_ffmpegcv(input_path: str, output_path: str):
    """GPU decode → OpenCV process → GPU encode."""

    # GPU reader
    cap = ffmpegcv.VideoCapture(input_path, gpu=0)
    fps = cap.get(cv2.CAP_PROP_FPS)
    width = int(cap.get(cv2.CAP_PROP_FRAME_WIDTH))
    height = int(cap.get(cv2.CAP_PROP_FRAME_HEIGHT))

    # GPU writer
    writer = ffmpegcv.VideoWriter(
        output_path,
        codec="h264_nvenc",
        fps=fps,
        frameSize=(width, height)
    )

    try:
        while True:
            ret, frame = cap.read()
            if not ret:
                break

            # OpenCV processing (CPU)
            processed = cv2.GaussianBlur(frame, (5, 5), 0)
            processed = cv2.Canny(processed, 100, 200)
            processed = cv2.cvtColor(processed, cv2.COLOR_GRAY2BGR)

            writer.write(processed)
    finally:
        cap.release()
        writer.release()

process_with_ffmpegcv("input.mp4", "processed.mp4")

---

VidGear: Multi-Threaded Video I/O

VidGear provides multi-threaded, high-performance video streaming with OpenCV integration.

Installation

pip install vidgear[core]

CamGear: High-Performance Capture

from vidgear.gears import CamGear
import cv2

# Multi-threaded video reading (faster than cv2.VideoCapture)
stream = CamGear(source="video.mp4").start()

while True:
    frame = stream.read()
    if frame is None:
        break

    # frame is BGR like OpenCV
    cv2.imshow("Frame", frame)
    if cv2.waitKey(1) & 0xFF == ord('q'):
        break

stream.stop()
cv2.destroyAllWindows()

RTSP/Network Streaming

from vidgear.gears import CamGear

# RTSP stream capture
options = {
    "THREADED_QUEUE_MODE": True,
    "CAP_PROP_FRAME_WIDTH": 1920,
    "CAP_PROP_FRAME_HEIGHT": 1080,
}

stream = CamGear(
    source="rtsp://192.168.1.100:554/stream",
    stream_mode=True,
    **options
).start()

while True:
    frame = stream.read()
    if frame is None:
        break
    # Process frame...

stream.stop()

WriteGear: FFmpeg-Backed Writing

from vidgear.gears import WriteGear
import cv2
import numpy as np

# High-performance writing with FFmpeg backend
output_params = {
    "-vcodec": "libx264",
    "-crf": 23,
    "-preset": "fast",
    "-pix_fmt": "yuv420p",
}

writer = WriteGear(output="output.mp4", **output_params)

# Generate and write frames
for i in range(300):
    frame = np.random.randint(0, 255, (1080, 1920, 3), dtype=np.uint8)
    writer.write(frame)

writer.close()

GPU-Accelerated WriteGear

from vidgear.gears import WriteGear

# NVENC GPU encoding
output_params = {
    "-vcodec": "h264_nvenc",
    "-preset": "p4",
    "-rc": "vbr",
    "-cq": 23,
    "-pix_fmt": "yuv420p",
}

writer = WriteGear(output="output.mp4", **output_params)
# ... write frames ...
writer.close()

---

Decord: ML Batch Loading

Decord provides 2x faster video loading than OpenCV, optimized for deep learning.

Installation

pip install decord

Basic Usage

from decord import VideoReader, cpu, gpu
import numpy as np

# CPU video reader
vr = VideoReader("video.mp4", ctx=cpu(0))

# Get video info
print(f"Total frames: {len(vr)}")
print(f"FPS: {vr.get_avg_fps()}")

# Read single frame (returns RGB!)
frame = vr[0]
print(frame.shape)  # (height, width, 3) - RGB format!

# CRITICAL: Decord returns RGB, not BGR
# Convert for OpenCV:
frame_bgr = frame[:, :, ::-1]  # RGB to BGR
# Or use cv2:
import cv2
frame_bgr = cv2.cvtColor(frame.asnumpy(), cv2.COLOR_RGB2BGR)

Batch Loading for ML

from decord import VideoReader, cpu
import numpy as np

vr = VideoReader("video.mp4", ctx=cpu(0))

# Load batch of frames (very efficient!)
frame_indices = [0, 10, 20, 30, 40]
batch = vr.get_batch(frame_indices)
print(batch.shape)  # (5, height, width, 3) - batch of RGB frames

# Load every 10th frame
all_frames = vr.get_batch(range(0, len(vr), 10))

# Convert batch to PyTorch tensor
import torch
tensor = torch.from_numpy(batch.asnumpy())
tensor = tensor.permute(0, 3, 1, 2)  # NHWC → NCHW for PyTorch
tensor = tensor.float() / 255.0  # Normalize

GPU Decoding

from decord import VideoReader, gpu

# GPU video reader (NVDEC)
vr = VideoReader("video.mp4", ctx=gpu(0))

# Batch read with GPU
frames = vr.get_batch([0, 1, 2, 3, 4])
# frames is on GPU, can be used directly with PyTorch

Decord with OpenCV Processing

from decord import VideoReader, cpu
import cv2
import numpy as np

def process_video_with_decord(path: str, batch_size: int = 32):
    """Efficient batch processing with Decord and OpenCV."""

    vr = VideoReader(path, ctx=cpu(0))
    total_frames = len(vr)

    results = []
    for start in range(0, total_frames, batch_size):
        end = min(start + batch_size, total_frames)
        batch = vr.get_batch(range(start, end))

        for frame_rgb in batch.asnumpy():
            # Convert RGB (Decord) to BGR (OpenCV)
            frame_bgr = cv2.cvtColor(frame_rgb, cv2.COLOR_RGB2BGR)

            # OpenCV processing
            gray = cv2.cvtColor(frame_bgr, cv2.COLOR_BGR2GRAY)
            edges = cv2.Canny(gray, 100, 200)

            results.append(edges)

    return results

---

PyAV: Frame-Level Precision

PyAV provides direct access to libav for precise frame-level control.

Installation

pip install av

Basic Usage

import av
import numpy as np

# Open video
container = av.open("video.mp4")

for frame in container.decode(video=0):
    # frame.to_ndarray() returns RGB by default!
    img_rgb = frame.to_ndarray(format="rgb24")

    # Convert to BGR for OpenCV
    import cv2
    img_bgr = cv2.cvtColor(img_rgb, cv2.COLOR_RGB2BGR)

    print(img_bgr.shape)  # (height, width, 3)

container.close()

Precise Seeking

import av

container = av.open("video.mp4")
stream = container.streams.video[0]

# Get frame at timestamp
target_pts = 10.0  # 10 seconds
container.seek(int(target_pts * av.time_base))

for frame in container.decode(video=0):
    # Process frame at approximately 10 seconds
    img = frame.to_ndarray(format="rgb24")
    break

container.close()

Efficient Frame Extraction

import av
import numpy as np
from typing import Generator

def extract_frames_pyav(
    path: str,
    fps: float = None
) -> Generator[np.ndarray, None, None]:
    """Extract frames with PyAV (yields BGR for OpenCV)."""

    container = av.open(path)
    stream = container.streams.video[0]

    # Set frame rate if specified
    if fps:
        stream.codec_context.framerate = fps

    for frame in container.decode(stream):
        # Get RGB array
        img_rgb = frame.to_ndarray(format="rgb24")
        # Convert to BGR for OpenCV
        img_bgr = img_rgb[:, :, ::-1]
        yield img_bgr

    container.close()

# Usage
for frame_bgr in extract_frames_pyav("video.mp4"):
    # Direct OpenCV processing
    edges = cv2.Canny(frame_bgr, 100, 200)

Write Video with PyAV

import av
import numpy as np

def write_video_pyav(frames: list, output_path: str, fps: float = 30.0):
    """Write frames to video with PyAV."""

    height, width = frames[0].shape[:2]

    container = av.open(output_path, mode='w')
    stream = container.add_stream('libx264', rate=fps)
    stream.width = width
    stream.height = height
    stream.pix_fmt = 'yuv420p'
    stream.options = {'crf': '23', 'preset': 'fast'}

    for frame_bgr in frames:
        # Convert BGR to RGB for PyAV
        frame_rgb = frame_bgr[:, :, ::-1]

        # Create VideoFrame
        av_frame = av.VideoFrame.from_ndarray(frame_rgb, format='rgb24')

        # Encode
        for packet in stream.encode(av_frame):
            container.mux(packet)

    # Flush encoder
    for packet in stream.encode():
        container.mux(packet)

    container.close()

---

Modal.com Integration: FFmpeg + OpenCV + GPU

Deploy FFmpeg + OpenCV pipelines on Modal's serverless infrastructure.

Image Configuration

import modal

# Complete image with FFmpeg, OpenCV, and GPU libraries
video_image = (
    modal.Image.debian_slim(python_version="3.12")
    .apt_install(
        "ffmpeg",           # FFmpeg CLI
        "libsm6",           # OpenCV dependencies
        "libxext6",
        "libgl1",
        "libglib2.0-0",
    )
    .pip_install(
        "opencv-python-headless",  # No GUI for server
        "ffmpeg-python",
        "numpy",
        "Pillow",
    )
)

# GPU image with additional libraries
gpu_video_image = (
    modal.Image.debian_slim(python_version="3.12")
    .apt_install("ffmpeg", "libsm6", "libxext6", "libgl1", "libglib2.0-0")
    .pip_install(
        "opencv-python-headless",
        "ffmpeg-python",
        "numpy",
        "torch",
        "decord",
        "ffmpegcv",
    )
)

app = modal.App("ffmpeg-opencv-pipeline", image=video_image)

Basic Frame Processing on Modal

import modal

app = modal.App("opencv-processing")

image = (
    modal.Image.debian_slim()
    .apt_install("ffmpeg", "libsm6", "libxext6", "libgl1")
    .pip_install("opencv-python-headless", "numpy")
)

@app.function(image=image)
def process_frame(frame_bytes: bytes) -> bytes:
    """Process single frame with OpenCV on Modal."""
    import cv2
    import numpy as np

    # Decode image (PNG or JPEG)
    nparr = np.frombuffer(frame_bytes, np.uint8)
    frame = cv2.imdecode(nparr, cv2.IMREAD_COLOR)  # BGR

    # OpenCV processing
    processed = cv2.GaussianBlur(frame, (15, 15), 0)

    # Encode back to PNG
    _, encoded = cv2.imencode('.png', processed)
    return encoded.tobytes()

@app.function(image=image, timeout=600)
def extract_and_process(video_bytes: bytes) -> list[bytes]:
    """Extract frames with FFmpeg, process with OpenCV."""
    import subprocess
    import tempfile
    from pathlib import Path
    import cv2
    import numpy as np

    with tempfile.TemporaryDirectory() as tmpdir:
        input_path = Path(tmpdir) / "input.mp4"
        input_path.write_bytes(video_bytes)

        # Extract frames with FFmpeg
        subprocess.run([
            "ffmpeg", "-i", str(input_path),
            "-vf", "fps=1",  # 1 frame per second
            f"{tmpdir}/frame_%04d.png"
        ], check=True, capture_output=True)

        # Process each frame with OpenCV
        results = []
        for frame_path in sorted(Path(tmpdir).glob("frame_*.png")):
            frame = cv2.imread(str(frame_path))

            # Apply edge detection
            gray = cv2.cvtColor(frame, cv2.COLOR_BGR2GRAY)
            edges = cv2.Canny(gray, 100, 200)

            # Encode result
            _, encoded = cv2.imencode('.png', edges)
            results.append(encoded.tobytes())

        return results

@app.local_entrypoint()
def main():
    video_bytes = Path("input.mp4").read_bytes()
    processed_frames = extract_and_process.remote(video_bytes)

    for i, frame_bytes in enumerate(processed_frames):
        Path(f"output/frame_{i:04d}.png").write_bytes(frame_bytes)

Parallel Frame Processing with map()

import modal
from pathlib import Path

app = modal.App("parallel-opencv")

image = (
    modal.Image.debian_slim()
    .apt_install("ffmpeg", "libsm6", "libxext6", "libgl1")
    .pip_install("opencv-python-headless", "numpy")
)

@app.function(image=image)
def extract_frames(video_bytes: bytes) -> list[bytes]:
    """Extract all frames from video."""
    import subprocess
    import tempfile
    from pathlib import Path

    with tempfile.TemporaryDirectory() as tmpdir:
        input_path = Path(tmpdir) / "input.mp4"
        input_path.write_bytes(video_bytes)

        subprocess.run([
            "ffmpeg", "-i", str(input_path),
            "-vsync", "0",
            f"{tmpdir}/frame_%06d.png"
        ], check=True, capture_output=True)

        frames = []
        for path in sorted(Path(tmpdir).glob("frame_*.png")):
            frames.append(path.read_bytes())

        return frames

@app.function(image=image)
def process_single_frame(frame_data: tuple[int, bytes]) -> tuple[int, bytes]:
    """Process a single frame."""
    import cv2
    import numpy as np

    frame_idx, frame_bytes = frame_data

    nparr = np.frombuffer(frame_bytes, np.uint8)
    frame = cv2.imdecode(nparr, cv2.IMREAD_COLOR)

    # Heavy OpenCV processing
    gray = cv2.cvtColor(frame, cv2.COLOR_BGR2GRAY)
    edges = cv2.Canny(gray, 50, 150)
    dilated = cv2.dilate(edges, None, iterations=2)
    contours, _ = cv2.findContours(dilated, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)

    # Draw contours on original frame
    result = frame.copy()
    cv2.drawContours(result, contours, -1, (0, 255, 0), 2)

    _, encoded = cv2.imencode('.png', result)
    return frame_idx, encoded.tobytes()

@app.function(image=image, timeout=600)
def combine_frames(processed_frames: list[tuple[int, bytes]], fps: float) -> bytes:
    """Combine processed frames back into video."""
    import subprocess
    import tempfile
    from pathlib import Path

    with tempfile.TemporaryDirectory() as tmpdir:
        # Sort by frame index and write
        for idx, frame_bytes in sorted(processed_frames):
            path = Path(tmpdir) / f"frame_{idx:06d}.png"
            path.write_bytes(frame_bytes)

        output_path = Path(tmpdir) / "output.mp4"

        subprocess.run([
            "ffmpeg", "-y",
            "-framerate", str(fps),
            "-i", f"{tmpdir}/frame_%06d.png",
            "-c:v", "libx264",
            "-preset", "fast",
            "-crf", "23",
            "-pix_fmt", "yuv420p",
            str(output_path)
        ], check=True, capture_output=True)

        return output_path.read_bytes()

@app.local_entrypoint()
def main():
    video_bytes = Path("input.mp4").read_bytes()

    # Extract frames (single container)
    frames = extract_frames.remote(video_bytes)
    print(f"Extracted {len(frames)} frames")

    # Process frames in parallel (many containers!)
    inputs = [(i, frame) for i, frame in enumerate(frames)]
    processed = list(process_single_frame.map(inputs))
    print(f"Processed {len(processed)} frames")

    # Combine back into video
    output = combine_frames.remote(processed, fps=30.0)
    Path("output.mp4").write_bytes(output)
    print("Done!")

GPU-Accelerated Pipeline with ffmpegcv

import modal

app = modal.App("gpu-video-pipeline")

# GPU image with ffmpegcv
gpu_image = (
    modal.Image.from_registry("nvidia/cuda:12.4.0-runtime-ubuntu22.04", add_python="3.12")
    .apt_install("ffmpeg", "libsm6", "libxext6", "libgl1", "libglib2.0-0")
    .pip_install("opencv-python-headless", "numpy", "ffmpegcv")
)

@app.function(image=gpu_image, gpu="T4")
def gpu_video_processing(video_bytes: bytes) -> bytes:
    """GPU-accelerated video processing with ffmpegcv."""
    import cv2
    import numpy as np
    import ffmpegcv
    import tempfile
    from pathlib import Path

    with tempfile.TemporaryDirectory() as tmpdir:
        input_path = Path(tmpdir) / "input.mp4"
        output_path = Path(tmpdir) / "output.mp4"
        input_path.write_bytes(video_bytes)

        # GPU reader (NVDEC)
        cap = ffmpegcv.VideoCapture(str(input_path), gpu=0)
        fps = cap.fps
        width = int(cap.width)
        height = int(cap.height)

        # GPU writer (NVENC)
        writer = ffmpegcv.VideoWriter(
            str(output_path),
            codec="h264_nvenc",
            fps=fps,
            frameSize=(width, height)
        )

        try:
            while True:
                ret, frame = cap.read()
                if not ret:
                    break

                # OpenCV processing (CPU - can't avoid this)
                processed = cv2.bilateralFilter(frame, 9, 75, 75)

                writer.write(processed)
        finally:
            cap.release()
            writer.release()

        return output_path.read_bytes()

@app.local_entrypoint()
def main():
    video = Path("input.mp4").read_bytes()
    result = gpu_video_processing.remote(video)
    Path("output.mp4").write_bytes(result)

---

Common Patterns Summary

Color Conversion Cheat Sheet

import cv2
import numpy as np

# OpenCV BGR → RGB (for FFmpeg, PIL, PyTorch)
rgb = cv2.cvtColor(bgr, cv2.COLOR_BGR2RGB)
rgb = bgr[:, :, ::-1]  # Faster, pure NumPy

# RGB → BGR (for OpenCV from FFmpeg, PyAV, Decord)
bgr = cv2.cvtColor(rgb, cv2.COLOR_RGB2BGR)
bgr = rgb[:, :, ::-1]  # Faster

# Grayscale
gray = cv2.cvtColor(bgr, cv2.COLOR_BGR2GRAY)
bgr_from_gray = cv2.cvtColor(gray, cv2.COLOR_GRAY2BGR)

# HSV (for color filtering)
hsv = cv2.cvtColor(bgr, cv2.COLOR_BGR2HSV)

Library Selection Decision Tree

Need video I/O?
├── Simple local file?
│   └── cv2.VideoCapture (built-in, easy)
├── Need GPU acceleration?
│   └── ffmpegcv (NVDEC/NVENC, OpenCV-compatible)
├── Network streaming (RTSP/RTMP)?
│   └── VidGear CamGear (multi-threaded)
├── ML batch training?
│   └── Decord (2x faster, batch GPU decode)
├── Frame-level precision/seeking?
│   └── PyAV (direct libav access)
└── Complex filters/formats?
    └── FFmpeg subprocess with pipes

Memory-Efficient Generators

def frame_generator(path: str, batch_size: int = 1):
    """Memory-efficient frame generator."""
    cap = cv2.VideoCapture(path)
    try:
        batch = []
        while True:
            ret, frame = cap.read()
            if not ret:
                if batch:
                    yield batch
                break
            batch.append(frame)
            if len(batch) == batch_size:
                yield batch
                batch = []
    finally:
        cap.release()

# Usage - never loads entire video into memory
for batch in frame_generator("large_video.mp4", batch_size=32):
    # Process batch of 32 frames
    pass

---

Related Skills

• ffmpeg-python-integration-reference - Type-safe Python-FFmpeg parameter mappings, color conversions, time units
• ffmpeg-fundamentals-2025 - Core FFmpeg operations
• ffmpeg-captions-subtitles - Subtitle processing with Python

References

• OpenCV Documentation
• FFmpeg Documentation
• ffmpegcv GitHub
• VidGear Documentation
• Decord GitHub
• PyAV Documentation
• Modal.com Documentation