looks

Named video effects that compile to a backend command, each carrying a licence tier — so you can demand commercial-safe-only and get a refusal rather than a surprise.

from looks.environment import needs_gpl

needs_gpl(["scale", "eq", "lut3d"])  # -> ('eq',)

eq is the obvious brightness/contrast/gamma filter. It exists only in a GPL build of ffmpeg. Nothing in the ffmpeg CLI will tell you, because the binary on your machine runs it fine — and the LGPL-clean way to do the same thing (curves, colorlevels, exposure) is one substitution away. That gap is what this package is for.

pip install looks — zero dependencies. ffmpeg is shelled out to, never linked.

What it does today

from looks.geometry import Size, placement, ffmpeg_chain, social_size

# a vertical phone clip, filled into a 16:9 frame
p = placement(Size(480, 850), social_size("youtube"), mode="fill")
ffmpeg_chain(p)
# 'scale=1920:3400,crop=1920:1080:0:1160'

from looks.lut import Ramp, gradient_map, write_cube

# a look, as a colour ramp indexed by lightness
look = gradient_map(
    Ramp.from_hex(
        [
            (8.2, "#2E0C18"),  # the shadow floor — NOT black
            (46.8, "#D5254A"),
            (100.0, "#FEF0DC"),  # the highlight — NOT white
        ]
    )
)
write_cube(look, "look.cube")  # ffmpeg: -vf lut3d=look.cube

from looks.measure import measure, dispersion

# how far apart do three sources look, after the effect?
stats = [measure(clip, source_id=name, vf="lut3d=look.cube") for name, clip in sources]
dispersion(stats)  # 2.98 -> the spread you are trying to close

Why it exists

PyPI has a dozen ffmpeg wrappers. Surveyed, none of them carries a named-effect registry, and none carries any licence awareness at all — not a field, not a check, not a warning. Meanwhile the licence facts are genuinely surprising:

• eq, boxblur, cropdetect, hqdn3d and 34 others exist only in a GPL ffmpeg. Only three of the 38 are colour operations, and each has an LGPL substitute in the same binary — so nothing you want is unreachable. But the gated ones are exactly the obvious first reach, and eq is the first thing anyone types.
• libx264 and libx265 are FFmpeg's only software H.264/HEVC encoders, both GPL. So an LGPL-tier deliverable is AV1, VP9, ProRes/FFV1 or hardware. Never x264. The wall is in the encoders, not the filters.
• geq is not GPL and has not been since FFmpeg 4.3, contrary to widespread belief.
• av, imageio-ffmpeg and opencv-python's macOS wheels all declare permissive licences while shipping GPL binaries. For av, three layers disagree: the metadata says BSD-3, FFmpeg's own avutil_license() says LGPLv3, and otool -L shows GPL libx264 and libx265 actually linked. A licence check that trusts a declared field is not a check.

Design

Pure-data specs, separable from execution. A Look is inspectable, persistable, diffable and costable before anything runs — the shape falaw.Plan has, except the cost unit is CPU-seconds, not dollars.

Unknown is a refusal, never a warning. That applies to a licence tier, to an ffmpeg build whose -L output matches nothing known, and to a clip whose colour range is untagged.

Two things are deliberately out of scope, and one of them is enforced rather than documented:

Every ffmpeg process looks starts ends in -f null -.

That admits every measurement, probe and diagnostic; it excludes every render, encode, mux and concat, and therefore looks.render() — which cannot be written without violating it. A convenience render function will get used and will rebuild one whole-timeline -filter_complex, which is a measured 2.3 GB regression on the box this ecosystem deploys to. looks emits the chain; you run it. The other exclusion is cut/EDL decisions: an effect says where a look applies, never where a cut is.

Parameters resolve against the clip they apply to. This is the package's most expensive lesson, and it is not a nicety. Building the first real look, one global flattening scale made the softest of three sources softer still — 46 → 38, where the other two went 35 → 72 and 117 → 114 — so it became the softest thing on screen, and it was the one the viewer complained about. The rule that follows is counter-intuitive: normalise the OUTPUT across sources, not the input. Don't sharpen the soft one; measure post-effect and pick parameters that land the clips in family. Full resolution was available and sharper, and was deliberately not used, because it would have made the softest source the sharpest thing in the edit — a new mismatch rather than a fix.

Status

Building. Shipped so far:

474 tests. They compose — looks/tests/test_integration.py walks the whole stack through the public surface: probe the environment, check the chain against the default ceiling, place a vertical clip into 16:9, verify the look cannot flicker, measure at source and post-effect, and confirm the two are correctly incomparable.

Next: the Effect / Look types, the licence-tier ladder, and the compiler.

The design of record is docs/decisions_and_rationale.md, backed by the research notes in docs/research/ — thirteen investigations, each adversarially reviewed by a second reader who re-ran every command rather than taking it on trust. That review found a false permission in the first version of the gate table (five filters are GPL-gated indirectly, through EXTERNAL_LIBRARY_GPL_LIST, with no gpl token on their own line), which is the kind of error this package exists to prevent and is therefore worth reading about.

A note on what this is not

looks is not legal advice, and a licence tier is a mechanical reading of published metadata and source, not an opinion about your situation. What it can do is refuse to let a chain reach a filter your declared ceiling forbids, and tell you which substitution gets you the same result. What it cannot do is know what you are shipping, to whom, under what agreement.

License

MIT