Ffmpeg Video Filter

Ffmpeg Video Filter

Context: $ARGUMENTS

Quick start

• Resize to 720p (keep aspect): -vf "scale=-2:720" → Step 2
• Letterbox to 1080p: -vf "scale=1920:1080:force_original_aspect_ratio=decrease,pad=1920:1080:(ow-iw)/2:(oh-ih)/2" → Example 2
• Watermark (PNG corner): -filter_complex "[1:v]scale=200:-1[wm];[0:v][wm]overlay=W-w-10:H-h-10" → Example 1
• Deinterlace: -vf "yadif=1" → Example 5
• Stack two videos: -filter_complex "[0:v][1:v]hstack=inputs=2" → Example 3

When to use

• Single-input transformation on one video stream (scale/crop/pad/fps/color) → use -vf.
• Combining two or more inputs (overlay, hstack, picture-in-picture) → use -filter_complex.
• Anything with split, multiple outputs, or named labels → -filter_complex + -map.

Step 1 — Pick -vf vs -filter_complex

-vf (and -af for audio) is shorthand for a single-input, single-output chain. Every filter is comma-separated. Example: -vf "scale=1280:720,eq=saturation=1.1,fps=30".

-filter_complex is the general graph. It accepts multiple inputs and outputs, uses [in][out] labels, and chains are separated by semicolons (filters within a chain are still comma-separated). When you use it you must -map every output stream you want in the container — default stream selection is disabled.

Rules of thumb:

• One input file, one video track, linear transform → -vf.
• Any time you write [0:v] or [1:v] → -filter_complex.
• You need the same source twice (split) → -filter_complex.

Step 2 — Build the filter chain

Canonical filter order: deinterlace → crop → scale → pad → color/eq → overlay → drawtext → fps → format. Deviating has real cost: cropping after scaling wastes pixels; scaling before cropping blurs detail you were going to throw away.

Scale recipes (scale is the workhorse):

scale=1280:720                    # exact
scale=-2:720                      # keep aspect, height 720, width rounded to even
scale=1920:-2                     # keep aspect, width 1920
scale=iw/2:ih/2                   # half size
scale=1920:1080:force_original_aspect_ratio=decrease   # fit inside box (letterbox prep)
scale=1920:1080:force_original_aspect_ratio=increase   # fill box (crop prep)
scale=1920:1080:flags=lanczos     # high-quality downscale

Use -2 (not -1) for the auto dimension: most codecs (H.264/HEVC/AV1 with YUV 4:2:0) require even width and height. -1 preserves aspect but can give odd numbers; -2 preserves aspect AND snaps to a multiple of 2.

Color / eq quick picks:

eq=brightness=0.05:saturation=1.2:contrast=1.1:gamma=1.0
curves=preset=increase_contrast
curves=preset=cross_process
hue=h=10:s=1.2                    # shift hue by 10 degrees, saturate
colorchannelmixer=.3:.4:.3:0:.3:.4:.3:0:.3:.4:.3   # luma-weighted grayscale

Overlay with alpha (PNG logo): pre-format the alpha plane so overlay doesn't silently drop it: format=yuva420p.

Step 3 — Run + verify

ffmpeg -i in.mp4 -vf "scale=-2:720" -c:v libx264 -crf 20 -preset medium -c:a copy out.mp4

For -filter_complex, map outputs explicitly:

ffmpeg -i main.mp4 -i logo.png \
  -filter_complex "[1:v]scale=200:-1[wm];[0:v][wm]overlay=W-w-10:H-h-10[v]" \
  -map "[v]" -map 0:a? -c:a copy out.mp4

Verify the output dimensions/fps with ffprobe:

ffprobe -v error -select_streams v:0 -show_entries stream=width,height,r_frame_rate,pix_fmt out.mp4

Available scripts

• scripts/vfilter.py — preset runner (scale-720p, scale-1080p-letterbox, watermark, deinterlace, drawtext-timecode, hstack, vstack, 2x-speed) plus --custom --filter-string "..." for arbitrary -vf expressions. Use --dry-run to print the command without running.

Workflow

uv run ${CLAUDE_SKILL_DIR}/scripts/vfilter.py --preset scale-720p --input in.mp4 --output out.mp4
uv run ${CLAUDE_SKILL_DIR}/scripts/vfilter.py --preset watermark --input in.mp4 --watermark logo.png --output out.mp4
uv run ${CLAUDE_SKILL_DIR}/scripts/vfilter.py --preset hstack --inputs a.mp4,b.mp4 --output side.mp4
uv run ${CLAUDE_SKILL_DIR}/scripts/vfilter.py --custom --filter-string "crop=iw-200:ih:100:0,scale=-2:720" --input in.mp4 --output out.mp4

Reference docs

• Read references/filters.md for exhaustive option tables, expression variables, -filter_complex label/map rules, drawtext escaping, and 50+ recipe snippets.

Gotchas

• Even dimensions. Most encoders demand even width and height. Use scale=-2:720, not scale=-1:720. libx264 with YUV 4:2:0 will hard-fail on odd sizes.
• drawtext needs a fontfile on many builds. If fontconfig is missing, omitting fontfile= throws "Fontconfig not found". Always specify a real path. macOS: /System/Library/Fonts/Supplemental/Arial.ttf. Linux: /usr/share/fonts/truetype/dejavu/DejaVuSans-Bold.ttf. Windows: C\:/Windows/Fonts/arial.ttf (note the escaped colon).
• drawtext has three escape levels. : separates filter params → escape inside text as \:. , separates filters → escape as \,. ' quotes the text → escape as '' or wrap the whole arg in double quotes at the shell level. For % signs use \%.
• -vf cannot take multiple inputs. The moment you need [1:v], switch to -filter_complex.
• -filter_complex disables default stream selection. You must -map every output label you want muxed (-map "[v]" -map 0:a?), or ffmpeg will not pick them up.
• Filter order changes results. Crop before scale to reduce compute and keep detail where it matters. Color ops go after scaling to avoid quantizing twice. drawtext/overlay go last so they render on the final pixel grid.
• overlay silently drops alpha. If your PNG has transparency, precede overlay with format=yuva420p on the overlay input: [1:v]format=yuva420p,scale=...[wm].
• setsar=1 fixes aspect metadata. Some sources (DV, DVD) have non-square pixels. After any scale you should append setsar=1 to force square-pixel output.
• scale=flags=lanczos matters when downscaling. Default is bicubic; lanczos is sharper for 1080p→720p. Upscaling: bicubic/lanczos similar; gauss for blur-friendly upscale.
• hstack/vstack require equal dimensions. Both inputs must match the stacking axis (same height for hstack, same width for vstack). Scale first.
• fps vs minterpolate. fps=30 drops/duplicates frames — cheap, no visual smoothing. minterpolate=fps=60 motion-interpolates new frames — expensive, but smooth.
• Deinterlace field order. Check with ffprobe -show_streams | grep field_order. yadif=1 sends one field per frame (keeps fps); yadif=0 merges fields (halves fps, often what you want for 29.97i → 29.97p).

Examples

Example 1 — Watermark a logo in the bottom-right

Input: main.mp4, logo.png (with alpha). Steps: scale logo to 200px wide, overlay with 10px margin.

ffmpeg -i main.mp4 -i logo.png \
  -filter_complex "[1:v]format=yuva420p,scale=200:-1[wm];[0:v][wm]overlay=W-w-10:H-h-10[v]" \
  -map "[v]" -map 0:a? -c:v libx264 -crf 20 -c:a copy out.mp4

Example 2 — Scale + letterbox to exact 1920x1080

Input: any aspect ratio. Steps: fit inside 1920x1080, pad with black to center.

ffmpeg -i in.mp4 -vf \
  "scale=1920:1080:force_original_aspect_ratio=decrease:flags=lanczos,\
pad=1920:1080:(ow-iw)/2:(oh-ih)/2:color=black,setsar=1" \
  -c:v libx264 -crf 18 -c:a copy out.mp4

Example 3 — 2-up side-by-side (hstack)

Both inputs must share height. Normalize first.

ffmpeg -i a.mp4 -i b.mp4 -filter_complex \
  "[0:v]scale=-2:720[a];[1:v]scale=-2:720[b];[a][b]hstack=inputs=2[v]" \
  -map "[v]" -c:v libx264 -crf 20 out.mp4

Example 4 — Timecode with drawtext

Burns running timecode at top-left with semi-transparent box.

ffmpeg -i in.mp4 -vf \
  "drawtext=fontfile=/System/Library/Fonts/Supplemental/Arial.ttf:\
timecode='00\:00\:00\:00':rate=25:fontsize=32:fontcolor=white:\
box=1:boxcolor=black@0.5:boxborderw=8:x=20:y=20" \
  -c:v libx264 -crf 20 -c:a copy out.mp4

Note the \: inside the timecode= value — the colons are escaped so ffmpeg doesn't interpret them as parameter separators.

Example 5 — Deinterlace 1080i to 1080p

ffmpeg -i interlaced.mxf -vf "yadif=1,setsar=1" \
  -c:v libx264 -crf 18 -preset slow -c:a copy progressive.mp4

Use yadif=0 if you actually want half the output framerate (merge fields, 50i → 25p).

Troubleshooting

Error: "width not divisible by 2"

Cause: scale=-1:720 produced an odd width, or you used a codec that demands even dimensions. Solution: replace -1 with -2. Example: scale=-2:720.

Error: "No such filter" / "Fontconfig not found"

Cause: your ffmpeg build lacks the filter or font system. Solution: check ffmpeg -filters | grep drawtext. If present, pass fontfile=/absolute/path/to/font.ttf explicitly — do not rely on font=Arial.

Error: "Output with label 'v' does not exist"

Cause: -filter_complex defined [v] but you forgot -map "[v]", or you misspelled the label. Solution: every bracketed label on the right side of the last filter must appear in a -map directive.

Error: filtering produces frozen/black overlay

Cause: overlay input has alpha but no explicit pixel format, so it was negotiated to a format without alpha. Solution: add format=yuva420p before overlay on the overlay chain: [1:v]format=yuva420p,scale=200:-1[wm].

Error: "Filter hstack:0 has an unconnected output" or size mismatch

Cause: inputs to hstack/vstack have different heights/widths. Solution: scale both to a common dimension first. [0:v]scale=-2:720[a];[1:v]scale=-2:720[b];[a][b]hstack.

drawtext text displays literally with backslashes visible

Cause: over-escaping. You only escape once at the filter-argument level. Solution: for static text, use text='Hello\, World' (single-quoted, \, for commas). For timecodes, use the timecode= option instead of text=.