Ffmpeg Frames Images

Ffmpeg Frames Images

Context: $ARGUMENTS

Frame extraction and image I/O with ffmpeg: single stills, periodic frame dumps, contact sheets, sprite sheets for scrubbers, high-quality GIFs, and building video back from an image sequence.

Quick start

• Single screenshot at a timestamp → Step 1 (screenshot) + Step 2
• Dump N frames per second to PNG/JPG → Step 1 (sequence) + Step 2
• Contact sheet / thumbnail grid → Step 1 (sheet) + Step 2
• Sprite sheet for HLS scrubber → Step 1 (sprite) + Step 2
• High-quality GIF from video → Step 1 (gif) + Step 2 (two-pass)
• Build video from PNG/JPG sequence → Step 1 (images-to-video) + Step 2

When to use

• Grab a poster/thumbnail frame at a specific time.
• Produce preview grids (contact sheet) for a long video.
• Generate scrubber sprites for a web player (HLS/DASH image tracks).
• Export frames for ML, computer vision, or VFX pipelines.
• Make a looping GIF for docs/social without visible banding.
• Re-encode a rendered image sequence (e.g. Blender/Nuke output) into MP4.

Step 1 — Pick your output

Choose exactly one shape; each has its own command pattern.

1. screenshot — one still image at a timestamp.
2. sequence — many images at regular intervals (fps=1 = 1/sec, fps=1/10 = every 10s).
3. sheet — one image that is a grid (tile=MxN) of stamps across the video.
4. sprite — same as sheet but sized for a web player's scrubber.
5. gif — animated GIF (two-command palette workflow).
6. images-to-video — reverse direction: image sequence → MP4.
7. thumbnail-smart — let ffmpeg pick a "representative" frame.
8. keyframes-only — export only frames marked as keyframes.

Step 2 — Build the command

2a. Single screenshot at 1:30

# Input-side seek is FAST (keyframe-snapped, may be a few ms off).
ffmpeg -ss 00:01:30 -i in.mp4 -vframes 1 -q:v 2 out.jpg

# Frame-accurate (SLOW, decodes from 0):
ffmpeg -i in.mp4 -ss 00:01:30 -vframes 1 -q:v 2 out.jpg

# Lossless PNG instead of JPG:
ffmpeg -ss 90 -i in.mp4 -vframes 1 -c:v png out.png

2b. One frame per second (or every N seconds)

# 1 frame / second:
ffmpeg -i in.mp4 -vf fps=1 frame_%04d.jpg

# 1 frame / 10 seconds:
ffmpeg -i in.mp4 -vf fps=1/10 frame_%04d.jpg

# Higher JPG quality (2 = best, 31 = worst):
ffmpeg -i in.mp4 -vf fps=1 -q:v 2 frame_%04d.jpg

2c. Contact sheet 4x4 at 5-second intervals

ffmpeg -i in.mp4 \
  -vf "fps=1/5,scale=320:-1,tile=4x4" \
  -frames:v 1 sheet.jpg

2d. Sprite sheet for HLS scrubber (10x10 grid, 10s interval, 160px wide)

ffmpeg -i in.mp4 \
  -vf "fps=1/10,scale=160:-1,tile=10x10" \
  -frames:v 1 sprites.jpg

Each sprite cell is 160 x (160/aspect). For a 16:9 video that's 160x90. With 100 cells and a 10s interval you cover 1000 seconds of runtime per sheet. See [references/recipes.md](references/recipes.md) for HLS EXT-X-IMAGE-STREAM-INF math.

2e. Smart thumbnail (auto-pick representative frame)

ffmpeg -i in.mp4 -vf "thumbnail=100,scale=640:-1" -frames:v 1 thumb.jpg

thumbnail=100 scans 100 consecutive frames and picks the most "typical" one (lowest histogram distance to the group average). Good for avoiding black/blurred frames.

2f. Keyframes only

ffmpeg -i in.mp4 -vf "select='eq(pict_type,I)'" -vsync vfr key_%04d.jpg

2g. High-quality GIF (TWO commands — palettegen, then paletteuse)

# 1) Generate an optimized palette
ffmpeg -i in.mp4 \
  -vf "fps=15,scale=480:-1:flags=lanczos,palettegen=max_colors=256" \
  -y palette.png

# 2) Apply the palette with dithering
ffmpeg -i in.mp4 -i palette.png \
  -lavfi "fps=15,scale=480:-1:flags=lanczos[x];[x][1:v]paletteuse=dither=bayer:bayer_scale=5" \
  -y out.gif

There is no reliable one-shot GIF command that gives this quality. The two-pass palettegen → paletteuse flow is the standard.

2h. Video from image sequence

# Note: -framerate goes BEFORE -i (it's an input option for image2).
ffmpeg -framerate 30 -i img_%04d.png \
  -c:v libx264 -pix_fmt yuv420p out.mp4

# Glob instead of numeric pattern:
ffmpeg -framerate 30 -pattern_type glob -i 'img_*.png' \
  -c:v libx264 -pix_fmt yuv420p out.mp4

# Start numbering at a non-zero index:
ffmpeg -framerate 24 -start_number 100 -i img_%04d.png out.mp4

Step 3 — Verify frame count

Before shipping, confirm you got what you expected.

# How many images landed on disk:
ls frame_*.jpg | wc -l

# Expected count for fps=1/N on a duration T:
#   floor(T / N) + 1
# (ffmpeg rounds to the nearest frame; off-by-one is normal.)

# Re-check a produced sheet's dimensions:
ffprobe -v error -select_streams v:0 -show_entries stream=width,height \
  -of csv=p=0 sheet.jpg

Available scripts

• scripts/frames.py — wrapper around all of the above. Subcommands:

screenshot, every, sheet, sprite, gif, images-to-video. Supports --dry-run and --verbose. Stdlib-only Python; no prompts.

Workflow

# Screenshot at 1m30s:
uv run ${CLAUDE_SKILL_DIR}/scripts/frames.py screenshot \
  --input in.mp4 --time 00:01:30 --output thumb.jpg

# 4x4 contact sheet at 5s intervals:
uv run ${CLAUDE_SKILL_DIR}/scripts/frames.py sheet \
  --input in.mp4 --cols 4 --rows 4 --interval 5 --width 320 --output sheet.jpg

# Two-pass GIF:
uv run ${CLAUDE_SKILL_DIR}/scripts/frames.py gif \
  --input in.mp4 --fps 15 --width 480 --output out.gif

# Rebuild video from PNGs:
uv run ${CLAUDE_SKILL_DIR}/scripts/frames.py images-to-video \
  --pattern 'img_%04d.png' --fps 30 --output out.mp4

Reference docs

Read [references/recipes.md](references/recipes.md) when:

• You need more than 8 recipes (there are 30+ in there).
• You need the JPEG quality scale, PNG compression flags, or image2 demuxer options (pattern_type, start_number, loop, framerate).
• You're tuning GIF size vs quality (dither tables, fps trade-offs).
• You're computing sprite sheet geometry for EXT-X-IMAGE-STREAM-INF playlists.

Gotchas

• -ss before -i is FAST but keyframe-snapped — the chosen time snaps to the nearest preceding keyframe. Off by up to the GOP length (often ~2s). Great for thumbnails.
• -ss after -i is FRAME-ACCURATE but slow — ffmpeg decodes from frame 0 to the target. Use only when exact frame matters.
• -q:v 2 is best-JPEG-quality. Scale is 2–31, LOWER is better. 2 is visually lossless-ish, 5 is web-quality, 31 is unusable.
• -frames:v 1 vs -vframes 1 — both work. Prefer -frames:v (stream-specific, modern form).
• image2 numeric pattern zero-padding must match the %0Nd width. img_%04d.png means files must be img_0001.png, not img_1.png, or you'll get "could not find codec parameters" errors.
• Glob only works if you add -pattern_type glob. Otherwise ffmpeg treats img_*.png as a literal filename.
• GIF quality REQUIRES the two-pass palette workflow. There is no acceptable single-command alternative; a naked ffmpeg -i in.mp4 out.gif produces banded, washed-out garbage.
• paletteuse=dither=bayer:bayer_scale=5 is the quality/size sweet spot. dither=none bands heavily; dither=floyd_steinberg is prettiest but noticeably larger.
• -framerate vs -r. -framerate is an INPUT option for the image2 demuxer (rate at which images are READ). -r is an OUTPUT option (rate STAMPED on the output). For image sequence → video, set -framerate BEFORE -i.
• JPG kills alpha. If the source has transparency (captures, overlays), use PNG or WebP.
• tile must match what you feed it. A 4x4 tile expects exactly 16 frames — use -frames:v 1 to stop after the first tiled output and trim the fps to produce at least 16 frames' worth.
• thumbnail=N needs N frames in the window. If the source is shorter than N frames, the filter still works but samples what it has. Default N is 100.
• On macOS, ffmpeg may not be on PATH in non-interactive shells. If your script fails with command not found, invoke with an absolute path (e.g. /opt/homebrew/bin/ffmpeg).

Examples

Example 1: YouTube-style poster frame at 1:30

ffmpeg -ss 00:01:30 -i lecture.mp4 -vframes 1 -q:v 2 -vf "scale=1280:-1" poster.jpg

Result: 1280-wide JPEG, keyframe-snapped (fast).

Example 2: Scrubber sprite sheet for a 1h video

For a 3600s video, sampling every 10s = 360 thumbs. A 19x19 grid (361 cells) covers it:

ffmpeg -i movie.mp4 \
  -vf "fps=1/10,scale=160:-1,tile=19x19" \
  -frames:v 1 sprites.jpg

Result: one JPG, ~3040x1710 pixels at 16:9, ready to slice on the client.

Example 3: Animated GIF of a 10s clip

# 1) palette
ffmpeg -ss 10 -t 10 -i clip.mp4 \
  -vf "fps=15,scale=480:-1:flags=lanczos,palettegen" -y palette.png
# 2) use palette
ffmpeg -ss 10 -t 10 -i clip.mp4 -i palette.png \
  -lavfi "fps=15,scale=480:-1:flags=lanczos[x];[x][1:v]paletteuse=dither=bayer:bayer_scale=5" \
  -y demo.gif

Result: ~2–3 MB GIF, no visible banding.

Example 4: Blender render → MP4

Blender wrote render_0001.png … render_0240.png at 24 fps:

ffmpeg -framerate 24 -i render_%04d.png \
  -c:v libx264 -pix_fmt yuv420p -crf 18 render.mp4

Result: 10-second MP4, high quality, broadly compatible.

Troubleshooting

Error: Could not find codec parameters for stream 0 (Video: none): unknown codec

Cause: image pattern's zero-padding doesn't match the actual filenames. Solution: rename files, or change %04d to match the real padding, or switch to -pattern_type glob -i 'img_*.png'.

Error: Output file #0 does not contain any stream

Cause: -ss placed after -i with an offset past the end of the video, OR thumbnail/select filtered everything out. Solution: verify the timestamp with ffprobe -show_entries format=duration; for select filters, loosen the predicate or pair with -vsync vfr.

Error: GIF looks washed out / banded / has dithering noise halo

Cause: skipped the palette step, or used dither=none. Solution: run the two-pass palettegen + paletteuse=dither=bayer:bayer_scale=5 workflow in section 2g.

Error: Image sequence imports at wrong speed / too fast / too slow

Cause: you used -r (output rate) when you needed -framerate (input rate). ffmpeg defaults to 25 fps on the image2 demuxer if neither is set. Solution: put -framerate N BEFORE -i img_%04d.png.

Error: tile filter complains about missing frames / produces a partial grid

Cause: fps didn't produce enough frames to fill MxN. Solution: either lower the interval (more frames), reduce the grid (2x2), or accept a partial tile.

Error: JPG is huge / low-quality

Cause: missing -q:v. Solution: add -q:v 2 (best) through -q:v 7 (web). Remember: lower = better.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: damionrashford
• Source: damionrashford/media-os
• License: MIT
• Homepage: https://damionrashford.github.io/media-os/

Install and usage instructions live in the source repository linked above.