POCKET 4P — D-Log 2 LUT Pack v1.0

For: DJI Osmo Pocket 4P shooting D-Log 2 · Format: .cube, 33- and 65-point · Price: free

12 LUTs for DJI Osmo Pocket 4P footage shot in D-Log 2, in 33- and 65-point .cube.

Built on DJI's own official D-Log2 → Rec.709 colorimetry, with their highlight clipping removed. Measured against their file:

	+4→+8 stops	+6→+10 stops	of range sent to pure white
DJI official LUT	30.9 CV/stop	9.1 CV/stop	13.0%
P4P Neutral	29.5 CV/stop	12.2 CV/stop	1.7%

Mid-tones and skin come through identical to DJI — measured deviation below the knee is 0.00 code values at 65-point — while the top end keeps its separation instead of being flattened into white.

What this is honestly worth. The 13.0% → 1.7% figure is measured on a synthetic ramp covering the full D-Log2 range. On real footage the gain depends entirely on whether your scene actually exceeds +7 stops above middle grey. Measured on a real interior clip peaking at +8.1 stops, neither LUT hard-clipped, and the pack held 14–22% more tonal separation above +6 stops while giving up 3–10% between +5 and +6. The transform redistributes highlight range — it does not invent detail the sensor never captured. Expect a large difference on skies, sun and windows shot from indoors, and a modest one on flat interiors.

---

The LUTs

Start here

01 Neutral	The reference render. DJI's colour and contrast, highlights intact. Use this when you want it to just look right.
02 FlatBase	Low-contrast base to grade on top of. Not a finished image — a starting point.

Everyday

03 Natural	True-to-life, skin nudged warm, greens kept honest. The daily driver.
12 Punch	High contrast and saturation. Reads at thumbnail size and on a phone outdoors.

Film-referenced

04 PrintFilm	Photochemical print feel — dense blacks, warm highlights, cool shadow toe.
06 Bleach	Bleach bypass. Contrast up, most of the colour pulled out.
11 RetroFade	Faded stock. Lifted milky blacks, cyan shadows, soft top end.
10 Mono	Black and white, orange-filter response, warm print tone.

Situational

05 TealAmber	Commercial teal/orange with skin protected from the shift.
07 NightNeon	Interiors and after dark. Shadows kept open, magenta/cyan separation.
08 Golden	Warm golden-hour bias, amber highlights.
09 Nordic	Cold and muted. Overcast, slate greens.

33 vs 65 point: 65 is more precise in saturated colour and worth it for delivery. 33 loads faster and is what mobile apps prefer. Identical looks otherwise.

---

Installing

DaVinci Resolve — copy the .cube files into the LUT folder (Project Settings → Color Management → Open LUT Folder), then Refresh. Apply on a node, or as a Timeline LUT. Right-click a clip → 3D LUT also works.

Premiere Pro — Lumetri Color → Creative → Look → Browse… and pick the file. For a technical-first workflow put it under Basic Correction → Input LUT instead.

Final Cut Pro — add the Custom LUT effect to the clip, set LUT to Choose Custom LUT…. Set the clip's camera LUT to None first so you aren't stacking two transforms.

CapCut / LumaFusion / VN — import as a custom LUT/filter. Use the 33-point files; several mobile apps won't load 65-point.

Photoshop / After Effects — Color Lookup adjustment layer, or the Apply Color LUT effect.

ffmpeg

ffmpeg -i input.mp4 -vf "lut3d=file='P4P_01_Neutral_size65.cube':interp=tetrahedral" -c:v libx264 -crf 16 output.mp4

Two things that will bite you

1. Turn off any other D-Log LUT first. If your NLE auto-applies a DJI camera LUT, these stack and the result is unusable.
2. These expect full-range D-Log 2. If your NLE tags the clip as limited range (16–235), the blacks will crush and the highlights will blow. In Resolve, set the clip to Full range in the Clip Attributes if it looks wrong.

---

Shooting D-Log 2 so these work

D-Log 2 puts middle grey much lower than you expect — about 31 IRE, against ~41 for D-Log M and ~46 for Rec.709. It looks underexposed on the screen when it is correct. That low placement is what buys the highlight range.

Derived from the curve (src/dlog2.py):

Subject	Reflectance	Stops	10-bit CV	IRE
Black floor	0%	—	64	6
Deep shadow	1.1%	−4	96	9
Shadow	4.5%	−2	187	18
18% grey card	18%	0	312	31
Skin (average)	28.6%	+0.67	354	35
White shirt / 90% white	90%	+2.3	457	45
Bright sky, window	576%	+5	625	61
DJI's LUT clips past here	2304%	+7	750	73
Sensor clip	~18400%	+10	~938	92

Practical version:

• Put skin around 33–37 IRE. Not 50. It will look flat and dark on the screen; that is correct.
• Zebras at 90–95 catch the real clip point, not the LUT's.
• Expose to the right if the scene is flat — up to a stop over is fine and buys you cleaner shadows, since there are ~10 stops of room above grey and only ~5.6 below. Underexposing D-Log 2 is the one thing that will make it look noisy and cheap.
• ISO 100 is the base. D-Log 2 runs 100–3200 on this camera.
• Shoot HEVC 10-bit. 8-bit log will band under any of these LUTs.

---

Audio — DJI Mic cleanup

Drag files onto ENHANCE-AUDIO.bat. That's the whole workflow — one or several at a time. Each comes back out next to the original as NAME_audio-enhanced.mp4, video untouched, audio rebuilt.

Or from a terminal:

python src/enhance_audio.py DJI_0001.MP4

Nothing is uploaded anywhere.

Targets: --target social (−14 LUFS, default), podcast (−16), broadcast (−23). Strength: --strength light|normal|strong. --wav writes a standalone file instead of remuxing.

Voice EQ

--eq off | natural | podcast | radio — podcast is the default.

Not copied off a generic "podcast EQ" chart. The measured DJI Mic take sat at warmth +17.5 dB and mud +15.6 dB while presence was −10.2 and air −22 — presence 25.8 dB below the mud, which is the classic lavalier sound, because the capsule is on your chest pointing away from your mouth. A broadcast voice wants that gap nearer 12–15 dB.

Measured on speech-only frames, presence minus mud:

profile	presence − mud
source	−25.8 dB	boomy and muffled
off	−23.8 dB	hum + level only
natural	−20.8 dB	gentle, safest on an already-good mic
podcast	−14.8 dB	generic broadcast shape
voice	−13.5 dB	the house preset — default
radio	−12.1 dB	aggressive, for noisy playback

All of them take presence from −10.2 to around 0 dB and air from −22.0 to −12 while leaving the 100–180 Hz fundamental alone — cutting that to kill boom is exactly what makes people sound thin and telephone-ish.

voice is podcast plus dips at the only two resonances that actually measured in this voice and room — 516 Hz boxiness and 6.4 kHz harshness, both +3.3 dB above the spectral trend. Nothing else needed carving: measured roughness was 1.11 dB std. Against podcast it lands the 516 Hz peak 1.7 dB lower, the 6.4 kHz harshness 1.6 dB lower, and keeps 0.7 dB more fundamental, so it's smoother and a little fuller.

Levelling is two-stage — a fast low-ratio stage catching transients and a slow one riding overall level. One compressor doing all the work is what makes a voice sound squashed.

It measures the recording first and tells you what it found — mains frequency, whether the channels are identical, loudness before and after.

Measured on a real DJI Mic take:

	before	after
Loudness	−27.7 LUFS	−14.7 LUFS
True peak	−8.0 dBTP	−1.4 dBTP (no clipping)
Range	14.1 LU	9.4 LU
50 Hz mains hum	+13.6 dB	−5.0 dB
150 Hz harmonic	+11.2 dB	+2.0 dB

Chaining with a neural denoiser

Cleanroom (MIT, local, Rust/Tauri) masters with DeepFilterNet3 — a proper on-device neural denoiser, much stronger than the afftdn here for room noise, hiss and fans. It also does transcription with diarization. What it does not document is any EQ, notch or de-ess stage, so it will not touch mains hum or a lavalier's presence dip.

They compose. Correct order is denoise → tonal work → loudness last:

python src/enhance_audio.py cleanroom_output.wav --denoise off

--denoise off skips the broadband stage so the signal is not denoised twice, and still applies hum notches, voice EQ and final loudness.

Denoisers

--denoise neural | fft | off — neural is the default, running DeepFilterNet3 locally. Nothing is uploaded.

Measured on a 12 dB SNR version of a real take, against the clean original. Noise is compared after level-matching on speech, because the chain normalises loudness and raw noise floors are otherwise not comparable:

variant	noise vs clean	speech distortion
noisy input	+17.0 dB	1.85 dB
off	+23.0 dB	5.25 dB
fft (afftdn)	+18.2 dB	3.58 dB
neural (DFN3)	+3.6 dB	2.30 dB

Both columns matter. Noise reduction alone is a meaningless score — a filter that deletes everything wins it — so speech distortion measures how far the voice's own spectrum drifted from the clean reference. DFN3 wins on both: within 3.6 dB of the original noise floor, 14.6 dB better than afftdn, while distorting the voice least.

--strength sets the attenuation limit: light 12 dB, normal 24 dB, strong 100 dB (full). Capping it beats full reduction — a background scrubbed to perfect silence pumps audibly every time speech starts.

Installing DeepFilterNet3. The deepfilternet pip package is a dead end: it pins numpy<2 and imports torchaudio.backend.common, which modern torchaudio no longer has. Build the Rust CLI instead:

cargo install --git https://github.com/Rikorose/DeepFilterNet --tag v0.5.6 --bin deep-filter --features "bin,tract,wav-utils,transforms" deep_filter

If that fails on the pinned time crate (it does not compile on rustc 1.98), clone the tag, run cargo update -p time, and build from inside libDF/ — building from the workspace root drags in dataset, which needs system HDF5 the binary never uses.

Why not an AI enhancer. That take measured 36.8 dB signal-to-noise — it was already clean. The faults were level and mains hum, which a deterministic chain fixes exactly and an AI re-synthesiser tends to either ignore or smear. Use Adobe Podcast Enhance or ElevenLabs Voice Isolator when a recording is genuinely damaged — heavy room reverb, wind, crowd, or rescuing camera audio because the mic failed. Not for this.

Three traps this works around, all found by measuring output rather than trusting filters:

1. Single-pass loudnorm is a dynamic estimator and undershot by 3 dB. Loudness is measured first, then applied.
2. loudnorm's linear=true does not enforce the true-peak ceiling — it delivered +1.6 dBTP, which clips. A real limiter follows it.
3. -ac 1 into the AAC encoder produced +1.63 dBTP from a signal that measured −0.99 dBTP by every other route — 2.6 dB of phantom peak from the same waveform. Mono collapse happens with pan= inside the filtergraph instead, and only after checking the channels really are identical (the DJI Mic can put two transmitters on separate channels, and collapsing that would delete someone).

What's here

luts/      the 24 .cube files (12 looks x 33 and 65 point)
vendor/    DJI's official cubes, used as the base. NOT in this repo -
           download them from DJI support to rebuild (see License).
src/       the generator: curve model, look engine, builder, validator
docs/      COLOR-SCIENCE.md - full derivation, measurements and sources
preview/   contact sheets
ref/       reference frames

Rebuild, check and preview:

python src/build_luts.py 33 65

python src/validate.py 65

python src/preview.py path/to/DJI_0001.MP4 --frames 3

Tune a look against real footage rather than by eye — reports how much each LUT tints neutral surfaces, plus skin hue, saturation and luma:

python src/measure_looks.py frame.png --size 65

---

Honesty about the colour science

DJI has not published a D-Log 2 white paper. The transfer curve in src/dlog2.py is a model fitted to published anchors (middle grey at CV 312, ceiling at 47500% reflectance, ~17 stops), not DJI's math. It independently predicts that CV 940 sits at +10.04 stops — which is exactly where DJI's own LUT stops responding, and matches third-party profiling of "~10 stops above grey before hard clip". That agreement is why it's trustworthy enough to build on.

We deliberately ship no D-Gamut2 matrix: two separate attempts to recover one from DJI's LUT failed honestly (a global fit at 122 CV RMS, a Jacobian method that drifted with luminance because DJI ramp saturation with brightness). Rather than invent one, the pack routes colour through DJI's official cube, which contains their real colorimetry.

Full derivation, every measurement and all sources: docs/COLOR-SCIENCE.md.

---

License

The LUTs in luts/, the documentation in docs/ and the previews are CC BY 4.0 — use them in client work, in films you sell, in anything. Credit in your video is appreciated but not required; a link back is what the licence asks for. The build scripts in src/ are MIT.

DJI's own LUTs are not included and are not covered by either licence. This pack is derived from measurements of DJI's D-Log2 to Rec.709 conversion, and those derived transforms are original work. DJI's files themselves are theirs. To rebuild from source, download the official pack from DJI's support site and put the cubes in vendor/.