Color Space and Gamma Reference for Video/Image Pipelines¶

Practical reference for color management in video/photo processing pipelines. Covers color spaces, gamma transfer functions, RAW workflows, and Color Management System (CMS) architecture.

Architecture Overview¶

Camera sensor (linear light)
  → OETF (Opto-Electronic Transfer Function) applied at recording
  → File (with color space + gamma metadata)
  → EOTF (Electro-Optical Transfer Function) applied at display
  → Monitor output

Most display devices apply Rec.709 gamma regardless of file metadata. Log-encoded material must be explicitly transformed before output.

Dynamic Range by Format¶

Color Space Standards¶

Display Standards (ascending gamut size)¶

Rec.709 ≈ sRGB  <  DCI-P3  <  Rec.2020

Camera-Specific Wide Gamut¶

Gamma Transfer Functions¶

Standard Gamma (SDR)¶

• Rec.709 Gamma: gamma 2.4, standard for broadcast
• sRGB: piecewise function ≈ gamma 2.2, for monitors
• 90% of displays automatically apply Rec.709 decoding to all input regardless of file gamma

Log Gamma¶

Logarithmic curve compressing ~14 stops into standard bit depth. Each manufacturer has a different curve — not interchangeable:

# Conceptual: log-encoded value from linear light
import numpy as np
# Generic approximation (NOT S-Log or V-Log specific):
log_val = np.log(linear_val + 1) / np.log(max_scene_luminance + 1)

Critical: Displaying Log on an SDR monitor without transform = washed-out image (monitor applies Rec.709 EOTF to Log data — wrong).

HLG (Hybrid Log Gamma)¶

Hybrid curve with two segments: - Below reference white: standard Rec.709 gamma (backward-compatible with SDR) - Above reference white: logarithmic extension (used on HDR displays)

Result: single file plays correctly on both SDR and HDR displays without separate versioning.

RAW Format¶

Linear photon counts from sensor, no gamma or color space applied.

Advantages: - Maximum dynamic range for the camera - Maximum color gamut - Post-process can apply any color space / gamma

Processing: Requires debayering (demosaicing) + explicit gamma + color space assignment. In NLEs this is the Camera RAW step.

Common mistake: Importing RAW without setting correct Input Color Space — NLE assumes Rec.709 and clips dynamic range.

Bit Depth vs. Chroma Subsampling¶

Bit Depth¶

Publishing to social platforms at 8-bit: gradients in backgrounds become visible bands. Mitigations: add subtle grain, avoid large smooth-tone backgrounds, publish in 4K (platform applies softer compression).

Chroma Subsampling (YUV)¶

Reduces color resolution while preserving luma:

Noise in camera footage concentrates in chroma channels. Denoising: apply to chroma (UV) channels only, not luma.

CMS Workflows¶

ColorSpace Transform (CST) — Explicit Pipeline¶

Input clip → CST node (Log+WideGamut → Rec.709+Gamma2.4)
           → Color corrections (in Rec.709)
           → Output

Most explicit and understandable. Balance/exposure corrections should be applied before CST to preserve full dynamic range during correction.

DaVinci Wide Gamut (Internal Working Space)¶

Input → Interpreted as camera color space (auto or manual)
      → All corrections in DaVinci Wide Gamut (~ACES-like container)
      → Output Color Space = Rec.709 (applied at render only)

Advantage over CST: corrections in wide gamut, no clipping risk during grading. Single compression step at output.

ACES¶

• Input Device Transform (IDT): camera color space → ACES
• All work in AP0/AP1 linear
• Output Device Transform (ODT): ACES → target display

ACES is linear — VFX/CG compositing integrates perfectly. Expensive to configure per-camera, Lift/Gamma controls behave non-intuitively.

Color Analysis Scopes¶

Parade (RGB/YRGB)¶

Separate R, G, B waveforms stacked. Neutral color = all channels at same height and parallel.

# Detect color cast programmatically
r_mean = img[:,:,0].mean()
g_mean = img[:,:,1].mean()
b_mean = img[:,:,2].mean()
avg = (r_mean + g_mean + b_mean) / 3
threshold = 0.05

cast = {c: (v - avg)/avg for c, v in zip('RGB', [r_mean, g_mean, b_mean])}
# cast['R'] > threshold → warm/red cast

Vectorscope¶

Plots hue angle vs. saturation radius. Center = neutral (no saturation). Skin tones fall on a specific diagonal line regardless of ethnicity — use skin tone indicator for white balance.

Skin tone hue angle: ~20-25° (reddish-orange sector)
Neutral white balance: distribution centroid near (0,0) in UV space

Broadcast legal limit: saturation must not exceed outer box markers on vectorscope.

Waveform¶

Horizontal pixel position mapped to brightness value (0–1023 in 10-bit). Useful for exposure: clip checking, IRE targets (18% grey = ~512).

Practical Input Interpretation Rules¶

White balance and exposure corrections: apply before transform (in log/wide-gamut space) to leverage full bit depth.

Gotchas¶

• Log on display without transform: 90% of displays apply Rec.709 EOTF to any input. Log footage shows as desaturated/flat unless explicitly converted. NLEs often silently apply wrong gamma.
• ARRI ProRes container ≠ Rec.709: ARRI frequently records LogC3 encoded video into ProRes .mov containers. Container format says nothing about color space — check camera metadata or ask the shooter.
• DaVinci Histogram axis reversal: in Resolve, histogram left = brighter, right = darker — opposite of Photoshop/Lightroom. Applying Lightroom histogram intuition causes incorrect exposure adjustments.
• HLG auto-detected as Rec.709: auto color managed workflows often misidentify HLG-encoded clips as Rec.709. Verify manually: HLG footage has a characteristic toe in dark areas that looks wrong under Rec.709 gamma.
• Chroma denoising target: camera noise is predominantly in UV channels, not luma. Applying spatial noise reduction equally to all channels sharpens luma artifacts while leaving chroma noise. Target the chroma channel specifically.
• Wide-gamut monitor + sRGB signal: consumer monitors with P3-wide panels often display Rec.709 signal too saturated without explicit color management. Calibration + ICC profile in OS is required for accurate monitoring.

See Also¶

• color theory for ml - perceptual color theory, Volkov local vs radiation color
• color checker and white balance - automated calibration pipeline
• intrinsic decomposition - separating albedo from illumination
• diffusion lora training - color space considerations for training data