如何提升FFmpeg制作幻灯片视频的性能？缩短至5分钟内

First, let's recap your scenario: you're merging 10 1080x1920 images and 2 same-resolution videos into a final video, currently taking ~11 minutes on a 4GB Linux VM. Below are actionable tweaks to get this under 5 minutes:

1. Optimize Encoder Speed & CPU Utilization #

• Use faster x264 presets: The default medium preset balances quality and speed, but switching to -preset fast or -preset veryfast will cut encoding time drastically with minimal quality loss (barely noticeable for most use cases). Add this right after -c:v libx264.
• Maximize multi-threading: Explicitly enable multi-threading with -threads auto (let FFmpeg use all available vCPUs) or specify the number of cores your VM has (e.g., -threads 4 if you have 4 vCPUs). This ensures you're not leaving CPU power unused.

2. Simplify Resource-Heavy Filters #

• Replace zoompan with loop (if zoom isn't mandatory): The zoompan filter does per-frame scaling, which is CPU-intensive. If you just need images to display for 5 seconds without slow zoom, use -loop 1 -t $IMAGE_TIME_SPAN when inputting images instead. This skips per-frame zoom calculations entirely.
  • Example for a single image: -loop 1 -t $IMAGE_TIME_SPAN -i $INPUT_DIR/_img1.jpg
• Remove redundant scale filters: Your videos are already 1080x1920, so the scale=$SCALE_SIZE filter on [2:v] and [4:v] is unnecessary. Delete those to save processing cycles.
• Shorten fade duration (if acceptable): A 4-second fade-in requires more frame processing than needed. Reducing it to 1-2 seconds (e.g., d=1) cuts filter computation time while still keeping a smooth transition.

3. Avoid Memory Swap Thrashing #

• Split processing into stages: Loading all 12 inputs at once eats up valuable RAM. Instead:
  1. First, convert each image to a short video clip with your desired effects (fade, zoom) and save them as temporary files.
  2. Then, concat all temporary image clips with the original videos in a second FFmpeg command.
     This reduces the number of inputs loaded into memory at once, preventing swap usage (which kills performance on 4GB RAM).
• Stick to lightweight pixel formats: You're already using yuv420p (good for compatibility and low memory), so ensure no intermediate steps switch to higher-bit formats.

4. Hardware Acceleration (If Available) #

If your VM has access to hardware acceleration (e.g., virtio-gpu, NVIDIA passthrough), use a hardware encoder instead of libx264:

• Intel GPUs: -c:v h264_qsv
• NVIDIA GPUs: -c:v h264_nvenc
• V4L2-compatible devices: -c:v h264_v4l2m2m
  Hardware encoders can encode 2-5x faster than software libx264—this is a massive speed boost if you can use it.

5. Tweak Input/Output Settings #

• Explicitly define audio parameters: When using anullsrc, specify sample rate and channels upfront to avoid auto-detection delays: -f lavfi -t 1 -i anullsrc=r=44100:cl=stereo
• Set image input framerate: Add -framerate 25 before your image inputs to match the zoompan frame count (25*5), so FFmpeg doesn't have to guess the framerate.

Modified Command Snippet (Example) #

Here's how your command might look after applying key optimizations:

INPUT_DATA=(_img1.jpg vid1.mp4 _img2.jpg vid2.mp4 _img3.jpg _img4.jpg _img5.jpg _img6.jpg _img7.jpg _img8.jpg _img9.jpg _img10.jpg)
IMAGE_TIME_SPAN=5
INPUT_DIR="input"
OUTPUT_DIR="output"
BACKGROUND_MUSIC="$INPUT_DIR/background.mp3"
OUTPUT_STEP1="$OUTPUT_DIR/tmp_video.mp4"
OUTPUT_STEP3="$OUTPUT_DIR/tmp_background.mp3"
OUTPUT_FILE="$OUTPUT_DIR/final.mp4"
SCALE_SIZE="1080x1920"

ffmpeg -y -threads auto \
-f lavfi -t 1 -i anullsrc=r=44100:cl=stereo \
-loop 1 -t $IMAGE_TIME_SPAN -i $INPUT_DIR/${INPUT_DATA[0]} \
-i $INPUT_DIR/${INPUT_DATA[1]} \
-loop 1 -t $IMAGE_TIME_SPAN -i $INPUT_DIR/${INPUT_DATA[2]} \
-i $INPUT_DIR/${INPUT_DATA[3]} \
-loop 1 -t $IMAGE_TIME_SPAN -i $INPUT_DIR/${INPUT_DATA[4]} \
-loop 1 -t $IMAGE_TIME_SPAN -i $INPUT_DIR/${INPUT_DATA[5]} \
-loop 1 -t $IMAGE_TIME_SPAN -i $INPUT_DIR/${INPUT_DATA[6]} \
-loop 1 -t $IMAGE_TIME_SPAN -i $INPUT_DIR/${INPUT_DATA[7]} \
-loop 1 -t $IMAGE_TIME_SPAN -i $INPUT_DIR/${INPUT_DATA[8]} \
-loop 1 -t $IMAGE_TIME_SPAN -i $INPUT_DIR/${INPUT_DATA[9]} \
-loop 1 -t $IMAGE_TIME_SPAN -i $INPUT_DIR/${INPUT_DATA[10]} \
-loop 1 -t $IMAGE_TIME_SPAN -i $INPUT_DIR/${INPUT_DATA[11]} \
-filter_complex \
"[1:v]fade=t=in:st=0:d=1:alpha=1[img1]; \
[2:v]fade=t=in:st=0:d=1:alpha=1[v1]; \
[3:v]fade=t=in:st=0:d=1:alpha=1[img2]; \
[4:v]fade=t=in:st=0:d=1:alpha=1[v2]; \
[5:v]fade=t=in:st=0:d=1:alpha=1[img3]; \
[6:v]fade=t=in:st=0:d=1:alpha=1[img4]; \
[7:v]fade=t=in:st=0:d=1:alpha=1[img5]; \
[8:v]fade=t=in:st=0:d=1:alpha=1[img6]; \
[9:v]fade=t=in:st=0:d=1:alpha=1[img7]; \
[10:v]fade=t=in:st=0:d=1:alpha=1[img8]; \
[11:v]fade=t=in:st=0:d=1:alpha=1[img9]; \
[12:v]fade=t=in:st=0:d=1:alpha=1[img10]; \
[img1][0:a][v1][0:a][img2][0:a][v2][0:a][img3][0:a][img4][0:a][img5][0:a][img6][0:a][img7][0:a][img8][0:a][img9][0:a][img10][0:a]concat=n=12:v=1:a=1" \
-pix_fmt yuv420p -c:v libx264 -preset fast \
$OUTPUT_STEP1

Note: If you must keep the zoom effect, revert to zoompan but retain the other optimizations (threads, preset, shorter fade, no redundant scale).

内容的提问来源于stack exchange，提问作者huynq9