Huemancer

Film color automation that runs on your own computer, as a Claude skill. Free, local, no cloud credits.

_{A 1960s Yeşilçam (Turkish) film: original on the left, colorized by Huemancer on the right. Video, no flicker.}

English · Türkçe · Gallery

---

English

Huemancer is a Claude Code skill with a set of Python tools behind it. You talk to Claude ("colorize this old film", "make it look cinematic", "match the colors of this clip"), and Claude runs the tools, shows you before/after pages, and renders the result.

What it does

Restoration	For bad copies (old TV, VHS, YouTube rips): deblocking and denoising with ffmpeg, then upscaling to HD with Real-ESRGAN. A cleaner input gives far better colorization.
Black-and-white → color	AI colorization with DDColor. The previous frame's color is carried forward with optical flow, so colors don't flicker, and the carry-over resets at every cut. Brightness always comes from the original full-resolution frame. Letterbox/pillarbox bars stay black.
Auto correction	White balance from near-neutral pixels, black/white point, midtones. Intentional casts (sunsets, neon) are left mostly alone. iPhone HDR (HLG/HDR10) → SDR BT.709.
Shot matching	Finds every cut, then pulls the shots toward the video's common look so the color doesn't jump between shots.
Reference matching	"Make it look like this": carries the tone curve and color character of a reference video or photo onto your footage.
Cinematic looks	11 built-in looks: teal-orange, warm film, print film, cold thriller, bleach bypass, pastel, 70s, day-for-night, noir B&W, sepia, natural. Add your own as small JSON files.
Preview pages	A grid of frames from across the video, one column per look. Click a frame and wipe between original and graded.
LUT export	Every look exports as a .cube file for DaVinci Resolve, Premiere, CapCut and similar editors.
Per-shot control	Override exposure, temperature or the look for single shots. Each render saves its full recipe.

Gallery

Black-and-white → color. The General (1926):

Colorize, then give it a look. The film_baski (print film) look on top of the colorized frame gives a "restored print" result:

Built-in looks on a 1943 Kodachrome by Alfred T. Palmer:

_{The sources are public domain, except the Yeşilçam clip, whose rights stay with their holders. See docs/KAYNAKLAR.md. Regenerate the images with docs/ornekler.py.}

Requirements

• Claude Code (CLI, desktop app or IDE extension)
• Python 3.10–3.13 and ffmpeg (a "full" build, which includes zscale for HDR)
• An NVIDIA GPU is strongly recommended for colorization. Grading works fine on CPU. Apple Silicon uses MPS.
• About 4 GB of disk space: PyTorch, plus ~870 MB for each colorization model.

Install

As a plugin (recommended): in Claude Code, run

/plugin marketplace add thecemozen/huemancer
/plugin install huemancer@huemancer

Or as a plain skill: copy skills/huemancer into ~/.claude/skills/.

Then just ask Claude to colorize or grade a video. On first use Claude runs the setup script, which creates a Python environment in ~/Huemancer, picks the right PyTorch build for your GPU and downloads DDColor. With the plain-skill install you can also run setup yourself:

python ~/.claude/skills/huemancer/scripts/kurulum.py            # install / repair
python ~/.claude/skills/huemancer/scripts/kurulum.py --kontrol  # status only

To keep everything (venv, models, renders) on another drive, run setup once with --kok "D:/Huemancer". The choice is saved in ~/.huemancer_home; the HUEMANCER_HOME environment variable overrides it.

Using it

Just talk to Claude:

Colorize C:\films\old_movie.mp4 and show me the models side by side.

Give wedding.mov a warm film look and match all the shots to each other.

Make my clip look like reference.mp4, a bit less saturated.

Export the teal-orange look as a LUT.

Results go to ~/Huemancer/Renkli/<video name>/. Your source files are never modified.

Speed (RTX 3060 Laptop, 6 GB)

Task	1080p	4K
Colorize (DDColor large)	~11 fps (90-min film ≈ 4.5 h)	—
Grade (correction + look + grain)	~63 fps	~10 fps
Analyze shots	~2-min clip in 30 s

Command line (without Claude)

The tools work on their own too. The command and option names are Turkish:

PY=~/Huemancer/.venv/bin/python   # Windows: ~/Huemancer/.venv/Scripts/python.exe
S=~/.claude/skills/huemancer/scripts

$PY $S/renk.py analiz  video.mp4                       # shots, HDR, B&W detection
$PY $S/onar.py onizle  film.mp4                        # compare restoration strengths
$PY $S/onar.py uygula  film.mp4 --guc 0.7              # clean + upscale -> film_onarilmis.mp4
$PY $S/boya.py onizle  film.mp4                        # compare colorization models
$PY $S/boya.py uygula  film.mp4 --model ddcolor_modelscope
$PY $S/renk.py onizle  video.mp4 --esitle              # preview all looks
$PY $S/renk.py uygula  video.mp4 --gorunum film_baski --esitle --yogunluk 0.8
$PY $S/renk.py uygula  video.mp4 --referans ref.mp4 --gorunum dogal
$PY $S/renk.py lut     sinematik -o sinematik.cube
$PY $S/renk.py gorunumler                              # list looks

Main options: --esitle (match shots), --referans (reference matching), --gorunum (look), --yogunluk (look strength), --sicaklik (temperature), --doygunluk (saturation), --kontrast (contrast), --pozlama (exposure in stops), --gren (grain), --vinyet (vignette), --sure (render only the first N seconds), --olcek (downscale). The keys for your own looks are documented in skills/huemancer/reference/looks.md.

Honest limitations

• Colorization invents plausible colors. Clothes, cars and walls may not be historically accurate.
• On very long shots the colorized background can drift slowly. A higher --yumusatma (temporal smoothing) helps.
• Output is H.264 8-bit BT.709 SDR. This is for finishing and social media, not a replacement for a professional HDR grading pipeline.

---

Türkçe

Huemancer, bir Claude Code skill'i ve arkasındaki Python araçlarıdır. Claude ile konuşursun ("bu eski filmi renklendir", "sinematik yap", "şu klibin renklerine benzet"). Claude araçları çalıştırır, sana önce/sonra sayfalarını gösterir ve sonucu çıkarır. Her şey kendi bilgisayarında, ücretsiz çalışır.

Ne yapar

Örnek görseller için yukarıdaki Gallery bölümüne bak.

• Onarır: kötü kopyaları (eski TV, VHS, YouTube) temizler ve Real-ESRGAN ile HD'ye büyütür. Temiz girdi çok daha iyi renklendirme demektir.
• Siyah-beyazı renklendirir: DDColor yapay zekâ modeliyle. Önceki karenin rengi optik akışla taşındığı için titremez ve her kesmede sıfırlanır. Siyah bantlar siyah kalır.
• Otomatik düzeltme: beyaz dengesi, siyah/beyaz noktası, orta tonlar ve iPhone HDR → SDR dönüşümü.
• Çekimleri eşitler: kesmeleri bulur, çekimler arasındaki renk zıplamalarını giderir.
• Referansa eşleştirir: bir videonun ya da fotoğrafın renk karakterini kendi videona taşır.
• 11 sinematik görünüm sunar: teal-orange, sıcak film, film baskı, soğuk gerilim, bleach bypass, pastel, 70'ler, gece mavisi, noir, sepya, doğal. Kendi görünümünü JSON dosyasıyla ekleyebilirsin.
• Ön izleme sayfası: her görünüm bir sütun. Kareye tıklayıp orijinalle karşılaştırırsın.
• LUT verir: her görünüm DaVinci Resolve, Premiere ve CapCut için .cube olarak dışa verilir.

Kurulum

Claude Code'da:

/plugin marketplace add thecemozen/huemancer
/plugin install huemancer@huemancer

Ya da skills/huemancer klasörünü ~/.claude/skills/ içine kopyala. Sonra Claude'a "şu videoyu renklendir" demen yeterli. İlk kullanımda Claude kurulumu kendisi yapar: ~/Huemancer içine Python ortamını kurar, ekran kartına uygun PyTorch'u seçer ve DDColor'u indirir.

Gerekenler: Python 3.10–3.13, ffmpeg (Windows: winget install Gyan.FFmpeg) ve renklendirme için tercihen bir NVIDIA ekran kartı.

Örnekler

C:\filmler\eski_film.mp4 dosyasını renklendir, modelleri yan yana göster.

dugun.mov videosuna sıcak film görünümü ver, çekimleri birbirine eşitle.

Videom referans.mp4 gibi görünsün, biraz daha az doygun olsun.

Sonuçlar ~/Huemancer/Renkli/<video adı>/ klasörüne yazılır. Kaynak dosyalara hiç dokunulmaz.

---

License

Huemancer's own code is released under the MIT license. DDColor (Apache-2.0) is downloaded during setup and is not bundled in this repository; see THIRD_PARTY_NOTICES.md.