FFmpeg on Cloud Run: AI Video Gotchas

2026-02-26 · Jazz Lien

The Setup

My video editor runs on Cloud Run. Users combine AI-generated clips (from Veo, BytePlus, Kling) with music, trim points, and titles. The backend uses FFmpeg to concatenate and export.

Worked perfectly in development. Broke mysteriously in production.

The Bug: "Stream map matches no streams"

Stream map '0:a:0' matches no streams.

Users reported "Failed to concatenate clips" with no other context. The generic error message hid the real problem.

The Root Cause

AI-generated videos from photo animation providers often have no audio track.

When you give Veo a photo and ask it to animate, you get video only. No audio. Not even silence.

My FFmpeg command assumed audio:

ffmpeg -i clip1.mp4 -i clip2.mp4 -filter_complex "[0:v][0:a][1:v][1:a]concat=n=2:v=1:a=1" output.mp4

The [0:a] references audio stream 0 of input 0. If there's no audio stream, FFmpeg fails.

The Fix: Detect Before Processing

Step 1: Probe for Audio

async function hasAudioTrack(filePath: string): Promise<boolean> {
  const result = await exec(
    `ffprobe -v error -select_streams a -show_entries stream=codec_type -of csv=p=0 "${filePath}"`
  );
  return result.stdout.trim() === 'audio';
}

Step 2: Add Silent Audio If Missing

async function ensureAudioTrack(inputPath: string, outputPath: string): Promise<void> {
  if (await hasAudioTrack(inputPath)) {
    // Already has audio, just copy
    await fs.copyFile(inputPath, outputPath);
    return;
  }

  // Get video duration
  const duration = await getVideoDuration(inputPath);

  // Add silent audio track
  await exec(`
    ffmpeg -i "${inputPath}" \
      -f lavfi -i anullsrc=r=44100:cl=stereo \
      -c:v copy -c:a aac -shortest \
      -t ${duration} \
      "${outputPath}"
  `);
}

The anullsrc filter generates silence. -shortest stops when the video ends.

Step 3: Pre-process All Inputs

Before concatenation:

async function prepareClipsForConcat(clips: string[]): Promise<string[]> {
  const prepared: string[] = [];

  for (const clip of clips) {
    const tempPath = `/tmp/${uuid()}.mp4`;
    await ensureAudioTrack(clip, tempPath);
    prepared.push(tempPath);
  }

  return prepared;
}

Safety Net: Optional Stream Mapping

Even with pre-processing, add a safety net. FFmpeg 4.1+ supports optional mapping:

ffmpeg -i input.mp4 -map 0:v:0 -map 0:a:0? output.mp4
#                                       ^ question mark = optional

The ? makes the audio stream optional. If it doesn't exist, FFmpeg continues without it instead of failing.

For concat:

ffmpeg -i clip1.mp4 -i clip2.mp4 \
  -filter_complex "[0:v][1:v]concat=n=2:v=1:a=0[outv]" \
  -map "[outv]" \
  output.mp4

This concatenates video only. Add audio in a separate pass if needed.

Debugging Cloud Run FFmpeg

Check the Actual Error

Client sees: "Failed to concatenate clips"

Server logs have the real error:

gcloud run logs read --service=my-service --limit=50

Look for FFmpeg stderr. The stream mapping error points to the -map flag, but the actual problem is the missing stream.

Common Culprits

Source	Audio?
AI photo animation (Veo, BytePlus, Kling)	Usually no
Screen recordings	Sometimes no
Re-encoded clips	Might be stripped
User uploads	Check before assuming

FFprobe Cheat Sheet

# Check all streams
ffprobe -show_streams input.mp4

# Check just audio
ffprobe -select_streams a -show_entries stream=codec_type input.mp4

# Get duration
ffprobe -v error -show_entries format=duration -of csv=p=0 input.mp4

Cloud Run Specifics

Memory for Video Processing

Video processing is memory-hungry. Default 512MB isn't enough:

gcloud run deploy my-service --memory=2Gi

Timeout for Long Videos

Default 5-minute timeout may not be enough for long exports:

gcloud run deploy my-service --timeout=15m

Concurrency

FFmpeg is CPU-intensive. Lower concurrency to avoid resource contention:

gcloud run deploy my-service --concurrency=4

The Pattern

1. Receive clip URLs
2. Download to /tmp
3. Probe each for audio
4. Add silent audio where missing
5. Concatenate with optional mapping as safety net
6. Clean up temp files
7. Return result

Never assume AI-generated video has audio. Always check.

FFmpeg is infinitely capable and infinitely confusing. The -map documentation is your friend.