Ffmpeg Compose Mask

Ffmpeg Compose Mask

Context: $ARGUMENTS

Advanced compositing and channel surgery. If you are keying green-screens, see ffmpeg-chromakey. If you are grading color, see ffmpeg-lut-grade. This skill is for pixel-exact compositing math, alpha plumbing, and plane-level operations.

Quick start

• Composite A and B through a mask (white→A, black→B): → Step 1, 2 (maskedmerge)
• Add an alpha channel from a separate mask video: → Step 2 (alphamerge)
• Pull mask out of an RGBA/yuva file: → Step 2 (alphaextract)
• Warp one video using another as a displacement map: → Step 2 (displace)
• Grab a single plane (Y, U, V, R, G, B, A): → Step 2 (extractplanes)
• Rebuild image from separated planes: → Step 2 (mergeplanes)
• Chromatic aberration / per-channel pixel shift: → Step 2 (rgbashift)
• Apply a Photoshop blend mode (screen, multiply, overlay...): → Step 2 (blend)
• Fill connected pixel region (Photoshop paint bucket): → Step 2 (floodfill)

When to use

• Roto-style compositing where you already have (or can compute) a per-pixel matte.
• Alpha plumbing: moving masks between files, reconstructing RGBA from split streams.
• Plane surgery: YUV/RGB channel isolation, swapping, rebuilding across pixel formats.
• Procedural effects: displacement warps, chromatic aberration, blend-mode layers.
• Edge-linking (hysteresis paired with edgedetect), mask refinement (maskfun, maskedthreshold, maskedclamp).
• Pixel remapping (remap), connected-region filling (floodfill).

Step 1 — Identify sources and the mask

Count your inputs. maskedmerge and displace need three; alphamerge and blend need two; alphaextract, extractplanes, rgbashift, maskfun, floodfill need one.

Verify each input with ffprobe before building the filter graph:

ffprobe -v error -select_streams v:0 -show_entries stream=width,height,pix_fmt,r_frame_rate -of default=nw=1 INPUT

All inputs to multi-input filters must share width, height, and pixel format. If they don't, scale/convert first (scale=W:H, format=yuva420p, fps=FPS).

Step 2 — Pick the compositing operation

Masked merge (3 inputs: source0, source1, mask)

Mask pixel value drives the blend: white = take from source0, black = from source1, gray = weighted mix.

ffmpeg -i A.mp4 -i B.mp4 -i MASK.mp4 \
  -filter_complex "[2:v]format=gray[m];[0:v][1:v][m]maskedmerge[out]" \
  -map "[out]" -c:v libx264 -crf 18 out.mp4

Options: planes (bitmask of planes to process; default 0xF = all).

Alpha merge (2 inputs: video + grayscale mask)

Bolts a grayscale matte onto the primary input as alpha. Output is yuva-format.

ffmpeg -i video.mp4 -i mask.mp4 \
  -filter_complex "[0:v][1:v]alphamerge[out]" \
  -map "[out]" -c:v prores_ks -profile:v 4444 out.mov

Alpha extract (1 RGBA/yuva input → grayscale mask)

ffmpeg -i rgba.mov -vf "alphaextract" -c:v ffv1 mask.mkv

Pre-/un-premultiply around compositing

Straight alpha inputs must NOT be premultiplied into RGB before compositing; premultiplied inputs must be un-multiplied before filters that expect straight, then remultiplied after.

# Composite with correct alpha math:
ffmpeg -i premul.mov -i bg.mov \
  -filter_complex "[0:v]unpremultiply=inplace=1[a];\
                   [1:v][a]overlay=format=auto[composed];\
                   [composed]premultiply=inplace=1[out]" \
  -map "[out]" -c:v prores_ks -profile:v 4444 out.mov

Both premultiply and unpremultiply accept inplace=1 to use the input's own alpha plane (no 2nd stream needed).

Displacement warp (3 inputs: base, xmap, ymap)

xmap and ymap are grayscale; mid-gray (128) means no shift, white = +scale, black = -scale. One map can supply both if you split.

ffmpeg -i base.mp4 -i xmap.mp4 -i ymap.mp4 \
  -filter_complex "[0:v][1:v][2:v]displace=edge=smear[out]" \
  -map "[out]" -c:v libx264 -crf 18 out.mp4

Options: edge ∈ {blank, smear, wrap, mirror} (default smear).

Plane surgery

Extract a single plane (always output is grayscale):

# Luma only:
ffmpeg -i in.mp4 -vf "format=yuv420p,extractplanes=y" -c:v ffv1 y.mkv
# All three at once:
ffmpeg -i in.mp4 -filter_complex "extractplanes=y+u+v[y][u][v]" \
  -map "[y]" y.mkv -map "[u]" u.mkv -map "[v]" v.mkv

Rebuild planes from multiple inputs via mergeplanes with a hex selector:

# 3 gray streams → single yuv444p:
ffmpeg -i y.mkv -i u.mkv -i v.mkv \
  -filter_complex "[0:v][1:v][2:v]mergeplanes=0x001020:yuv444p[out]" \
  -map "[out]" out.mov

# Add a gray stream as alpha onto a yuv444p stream → yuva444p:
# [yuv][mask]mergeplanes=0x00010210:yuva444p

Swap planes within a single stream with shuffleplanes (per-plane index into input 0):

# Y=plane 0, swap U and V:
ffmpeg -i in.mp4 -vf "shuffleplanes=0:2:1:3" out.mp4

Per-channel pixel shift / chromatic aberration

ffmpeg -i in.mp4 -vf "format=gbrp,rgbashift=rh=2:bh=-2" out.mp4

Options: rh, rv, gh, gv, bh, bv, ah, av (horizontal/vertical shift in pixels per channel; default 0). edge ∈ {smear, wrap}.

Spatial blend (2 inputs)

Photoshop-style blend modes applied pixelwise between two aligned streams:

ffmpeg -i base.mp4 -i overlay.mp4 \
  -filter_complex "[0:v][1:v]blend=all_mode=screen:all_opacity=0.5[out]" \
  -map "[out]" out.mp4

Modes: addition, and, average, burn, darken, difference, divide, dodge, exclusion, freeze, glow, grainextract, grainmerge, hardlight, hardmix, heat, lighten, linearlight, multiply, negation, normal, or, overlay, phoenix, pinlight, reflect, screen, softdifference, softlight, subtract, vividlight, xor. Per-plane variants: c0_mode, c1_mode, c2_mode, c3_mode plus matching c0_opacity, etc. all_expr accepts arbitrary expressions on A, B, X, Y, W, H, SW, SH, T, N.

Floodfill (paint-bucket fill)

# Replace pixels at (100, 100) and all connected same-color ones with red:
ffmpeg -i in.mp4 -vf "floodfill=x=100:y=100:s0=10:s1=10:s2=10:d0=255:d1=0:d2=0" out.mp4

x, y are the seed pixel (integer). s0,s1,s2,s3 match the source pixel (must equal them exactly to flood); d0,d1,d2,d3 are the destination values. Single component = single plane match/replace.

Hysteresis (edge-linking)

Pair with edgedetect. First input = strong edges (high threshold), second = weak edges (low threshold). Weak edges connected to strong edges are kept.

ffmpeg -i in.mp4 -filter_complex \
  "[0:v]split[s][w];[s]edgedetect=high=0.3:low=0.3,format=gray[strong];\
   [w]edgedetect=high=0.1:low=0.1,format=gray[weak];\
   [strong][weak]hysteresis=threshold=0:planes=1[out]" \
  -map "[out]" out.mp4

Mask generation helpers

maskfun=low=N:high=M:fill=F:sum=S — turn input into a mask (low/high thresholds, fill color for masked pixels, sum window). maskedthreshold=threshold=N:mode=abs|diff — two-input: output 1st where |a-b| ≤ threshold, else 2nd. maskedclamp=undershoot=U:overshoot=O:planes=P — three-input: clamp input0 to the band defined by input1 (lower) and input2 (upper). maskedmax, maskedmin — three-input: pick whichever of input1/input2 is farther/closer to input0 per pixel.

Step 3 — Build the filter graph (labels + -map)

Always name every intermediate with [label] when you have more than one branch. Every label that leaves -filter_complex must be consumed by -map "[label]" on an output.

ffmpeg \
  -i source0.mp4 \
  -i source1.mp4 \
  -i mask.mp4 \
  -filter_complex "\
    [0:v]format=yuva420p[a];\
    [1:v]format=yuva420p[b];\
    [2:v]format=gray[m];\
    [a][b][m]maskedmerge[composed];\
    [composed]format=yuv420p[out_v];\
    [0:a][1:a]amix=inputs=2[out_a]\
  " \
  -map "[out_v]" -map "[out_a]" \
  -c:v libx264 -crf 18 -c:a aac -b:a 192k out.mp4

Step 4 — Verify alpha and channels

After the render, confirm pixel format and plane counts:

ffprobe -v error -show_entries stream=pix_fmt,width,height -of default=nw=1 out.mov

Expected: yuva420p, yuva444p, rgba, or argb if alpha was intended. If you see yuv420p, alpha was lost (either the container doesn't support it or a later filter/encoder dropped it; see Gotchas).

Preview the mask-only pass before committing:

ffplay -f lavfi "movie=out.mov,alphaextract"

Available scripts

• scripts/compose.py — argparse-driven wrapper that builds and (optionally) runs the correct filter graph for each operation: mask-merge, alpha-merge, alpha-extract, displace, planes, blend, rgbashift. Stdlib only; supports --dry-run and --verbose.

# Preview the command:
python3 ${CLAUDE_SKILL_DIR}/scripts/compose.py mask-merge \
  --source0 A.mp4 --source1 B.mp4 --mask M.mp4 --output out.mp4 --dry-run

# Run it:
python3 ${CLAUDE_SKILL_DIR}/scripts/compose.py blend \
  --source0 a.mp4 --source1 b.mp4 --mode screen --output out.mp4

Reference docs

• references/filters.md — per-filter option tables, masking mental model, pre/un-premultiply decision tree, plane-selector hex notation, recipe book (rotoscope, sky replacement, chromatic aberration, displacement warp).

Gotchas

• maskedmerge takes three inputs in exact order: source0, source1, mask. Reordering silently produces garbage because each role is positional.
• Masks must be single-plane grayscale. Insert format=gray on the mask branch before handing it to maskedmerge, alphamerge, displace (x/y maps), or hysteresis. Skipping this is the #1 cause of "why is my composite tinted?"
• alphamerge expects [video][mask] in that order; swapped → inverted matte.
• Alpha operations require yuva420p / yuva444p / rgba / gbrap pixel formats. Insert format=yuva420p before any filter that expects alpha; libx264 by default uses yuv420p and will silently drop alpha — use libx264rgb, prores_ks -profile:v 4444, qtrle, png, ffv1, vp9 (yuva420p), or PNG sequences for alpha-carrying output.
• Output container must support alpha. MP4 doesn't for most codecs; use MOV (ProRes 4444, QTRLE), MKV (FFV1, VP9 yuva420p), WebM (VP9 yuva420p), or PNG sequence.
• Premultiply vs straight alpha: if your composite has dark halos or incorrect edge colors, insert unpremultiply=inplace=1 before compositing and premultiply=inplace=1 after.
• extractplanes requires planar pixel formats. Force with format=yuv420p or format=gbrp first if unsure. The option syntax is extractplanes=y+u+v (plus-separated names), not a bitmask.
• mergeplanes uses a packed hex selector: 0xAABBCC for 3 planes, 0xAABBCCDD for 4. Each byte is (plane_idx << 4) | input_idx. Read right-to-left for output plane 0, 1, 2... Example: 0x001020 → out plane 0 = input 0 plane 0, out plane 1 = input 1 plane 0, out plane 2 = input 2 plane 0.
• shuffleplanes=a:b:c:d — each number is which plane of input 0 becomes output planes 0, 1, 2, 3 respectively. It is NOT per-input addressing (use mergeplanes for that). shuffleplanes=0:2:1:3 → Y stays, U and V swap.
• rgbashift shifts by pixels, not subsampled positions; on chroma-subsampled formats convert to gbrp first for pixel-accurate results.
• displace x/y maps must be grayscale where mid-gray (~128) = zero shift. White shifts positively, black negatively; amplitude scales with the plane bit-depth range.
• blend=all_mode=screen and similar modes operate on whatever range the pixel format uses — in YUV the math is technically correct but differs visually from RGB-space reference blends. Convert to gbrp first (format=gbrp) to match Photoshop output.
• maskedclamp is three-input (source, lower-bound, upper-bound) — lower/upper must bracket source pixel-for-pixel.
• floodfill coordinates are integer pixels measured from top-left. The filter floods only pixels whose plane values exactly match s0..s3; it's not a soft-threshold match.
• hysteresis requires two single-plane inputs of identical dimensions. Run edgedetect with different thresholds into each branch.
• Every label produced inside -filter_complex must be consumed by -map "[label]" when you have multiple outputs; unmapped labels are silently dropped.

Examples

Example 1 — Replace sky via luminance mask

ffmpeg -i footage.mp4 -i sky.mp4 \
  -filter_complex "\
    [0:v]split[src][lum];\
    [lum]hue=s=0,curves=master='0/0 0.7/0 0.75/1 1/1',format=gray[m];\
    [1:v]scale=iw:ih[sky];\
    [src][sky][m]maskedmerge[out]\
  " -map "[out]" -c:v libx264 -crf 18 sky_replaced.mp4

Example 2 — Extract alpha, transmit YUV + mask separately, reconstruct

# Sender:
ffmpeg -i rgba.mov -vf "alphaextract" -c:v ffv1 mask.mkv
ffmpeg -i rgba.mov -vf "format=yuv420p" -c:v libx264 -crf 12 color.mp4
# Receiver:
ffmpeg -i color.mp4 -i mask.mkv \
  -filter_complex "[0:v][1:v]alphamerge" \
  -c:v prores_ks -profile:v 4444 reconstructed.mov

Example 3 — Chromatic aberration look

ffmpeg -i in.mp4 -vf "format=gbrp,rgbashift=rh=3:bh=-3,format=yuv420p" \
  -c:v libx264 -crf 18 aberrated.mp4

Example 4 — Displacement warp from noise

ffmpeg -i base.mp4 -f lavfi -i "nullsrc=s=1920x1080:d=10,\
geq='128+40*sin(2*PI*X/100+T):128:128'" \
  -filter_complex "[1:v]split[x][y];[0:v][x][y]displace=edge=smear[out]" \
  -map "[out]" warp.mp4

Troubleshooting

Error: "Input link parameters do not match"

Cause: mismatched width, height, or pixel format between filter inputs. Solution: scale and format-convert each branch before the multi-input filter. Insert scale=W:H,format=yuva420p on every input.

Output has no alpha even though I merged one

Cause: encoder or container silently dropped alpha. Solution: confirm pix_fmt with ffprobe. Switch to prores_ks -profile:v 4444 in .mov, qtrle in .mov, ffv1 in .mkv, or libvpx-vp9 -pix_fmt yuva420p in .webm.

Composite has dark halos around soft edges

Cause: premultiplied alpha treated as straight (or vice versa). Solution: wrap the compositing chain with unpremultiply=inplace=1 before, premultiply=inplace=1 after.

extractplanes errors with "does not support this format"

Cause: non-planar input (e.g., NV12, packed YUYV). Solution: insert format=yuv420p (or format=gbrp for RGB work) before extractplanes.

mergeplanes produces weird colors

Cause: wrong hex selector. The byte order is little-endian-ish per-plane; read each byte as (plane_idx << 4) | input_idx. Solution: walk through the selector on paper. For 3 gray inputs → yuv444p you want 0x001020 (out0=in0.0, out1=in1.0, out2=in2.0).

blend result looks wrong in YUV

Cause: blend modes are defined mathematically over the sample range; YUV samples aren't linear RGB. Solution: format=gbrp before blend, then convert back to yuv420p for the encoder.