Low-level Color Guider (LCG) Techniques

• LCG is a set of techniques that embed precise, local color cues into images and videos, supporting applications such as steganography, diffusion modeling, and reference-based colorization.
• It operates by encoding pixel-level or patch-wise color information and coupling it with higher-level semantic or guidance signals to ensure robustness and imperceptibility.
• LCG applications span digital steganography, diffusion-based image generation, reference video colorization, and unobtrusive VR/AR guidance, demonstrating versatile impact on visual processing.

The Low-level Color Guider (LCG) encompasses a set of techniques and architectural mechanisms for guiding, modulating, or embedding fine-grained color information in images and video. LCGs are used in diverse domains: digital-image steganography, generative diffusion modeling, reference-based colorization, and perceptual-guidance systems. Central to all LCG variants is the encoding or injection of color cues at a spatially local, often patch-wise, scale—complementing or controlling higher-level semantics, channel assignments, or temporal dynamics.

1. Mechanism and Core Principles

In all LCG frameworks, local color cues either determine how underlying information is processed (e.g., controlling channel embedding in steganography or dictating attention in neural networks) or are physically presented to modulate human perception. The operational details depend on application domain but share the following traits:

• Encoding local color state or reference with precision, typically per pixel or per latent patch.
• Coupling low-level color information with higher-level (semantic or global) context, either through channel coordination, transformer attention, gradient-based guidance, or external guidance signals.
• Providing mechanisms to ensure robustness and imperceptibility—whether statistical (for steganography), perceptual (for human users), or in the context of model-internal representations.

2. LCG in Color Image Steganography

In color image steganography, LCG governs the dynamic allocation of embedding capacity among RGB planes on a per-pixel basis (Amirtharajan et al., 2010). The key workflow is:

1. Guiding-channel selection: For every pixel, designate one channel (typically Red, or user/cyclically chosen) as the guiding (“indicator”) channel, $P_1$.
2. LSB excess-3 mapping: Extract two least significant bits ($b_1$, $b_0$) from $P_1$, interpret their value $b=b_1b_0$, and compute $n = \text{Decimal}(b) + 3$, yielding $n \in \{3,4,5,6\}$.
3. Embedding allocation: Distribute $n$ payload bits across the remaining channels $(P_2, P_3)$:
   • If $n$ is even, embed $b_1$0 bits in each.
   • If $b_1$1 is odd, embed $b_1$2 in $b_1$3 and $b_1$4 in $b_1$5. This randomized allocation avoids fixed statistical signatures and enables adaptive balancing among channels.

Three embedding strategies are compared:

• Default R-guide: Red always acts as $b_1$6, leading to distortion (MSE) localization in $b_1$7, $b_1$8.
• User-selectable guide: $b_1$9 can be any channel, chosen per embedding process.
• Cyclic guide: The guiding channel rotates $b_0$0 for each pixel, which achieves uniform MSE distribution.

The output stego image applies an Optimal Pixel Adjustment Process (OPAP), which locally minimizes embedding distortion under LSB constancy constraints. OPAP adjusts a stego pixel $b_0$1 (after embedding $b_0$2 bits) by: $b_0$3 where $b_0$4 is the original pixel value.

3. LCG in Diffusion Models and Color Guidance

LCG also refers to a precise method for guiding the generative outputs of diffusion models toward a prescribed low-level color distribution (Bordin et al., 2024). This LCG:

• Projects the target image or condition $b_0$5 onto its $b_0$6 lowest 2D-DCT coefficients via a matrix $b_0$7: $b_0$8.
• Defines the color guidance operator as the exact gradient of the log-likelihood of $b_0$9 given the (denoised) intermediate image prediction, yielding: $P_1$0 with $P_1$1. In latent diffusion models, an additional mean-shift correction is included.

This guidance is applied at every denoising step without retraining, using empirically measured variance $P_1$2 per timestep, with the critical property that the guidance scale $P_1$3 remains high through all timesteps—improving the faithfulness of color transfer, especially under low bit-rate compression constraints.

4. LCG in Reference-based Animation Colorization

In the context of video diffusion transformers for reference-based colorization, LCG is an architectural module that injects patch-level color information from reference frames directly into the transformer backbone (Zhang et al., 27 Jul 2025). The process is:

• Encode a reference image $P_1$4 with a VAE to a latent $P_1$5; flatten and project to $P_1$6 vision tokens.
• The LCG module, parameterized as a full transformer stack, concatenates its vision tokens with the main model’s text and vision tokens plus high-level color tokens from an HCE (High-Level Color Extractor).
• At each layer, self-attention operates on the concatenated sequence: $P_1$7 where $P_1$8 represents text and vision tokens, $P_1$9 the fixed HCE tokens.
• The output is sliced back to update the main model's tokens, effectively letting each patch attend to local color in $b=b_1b_0$0 at every layer and diffusion step.

LCG training in this context uses standard diffusion denoising losses with all other modules frozen, enforcing that the colorization benefits arise entirely from low-level reference injection.

5. LCG for Unobtrusive Visual Modulation

A distinct application of LCG lies in visual guidance by chromatic temporal modulation, notably used for imperceptible gaze guidance in VR/AR (Tosa et al., 2024). In this domain:

• LCG means alternating ROI colors $b=b_1b_0$1 at 25–60 Hz such that users are not consciously aware of flicker but experience subtle, measurable guidance toward the modulated region.
• The pair $b=b_1b_0$2 is defined by traversing the major axis of the nearest MacAdam ellipse in xyY space:

$b=b_1b_0$3

$b=b_1b_0$4

with $b=b_1b_0$5 selected using psychometric thresholds on awareness/flicker.

• Application at runtime uses explicit pseudocode for color assignment per frame, spatial masking, and careful blending in CIELAB or CIE xyY.

Empirically, such LCG-induced modulation can speed task completion by 30–45% under unobtrusive guidance, with median naturalness scores of $b=b_1b_0$6 (unobtrusive) and minimal obtrusiveness reported. Guidance effectiveness is achieved without explicit overlays, leveraging bottom-up chromatic saliency.

6. Comparative Performance and Design Trade-offs

A numerical comparison in the steganographic setting is instructive:

Guiding Approach	Mean BPP	MSE (max)	PSNR (min, dB)	Distortion Localization	OPAP ΦPSNR
Default R-guide	~0.75	~4.5	~41.6	2 channels	+1–2 dB
User-selectable guide	~0.76	~4.5	~41.6	2 channels	+1–2 dB
Cyclic guide	0.50–0.53	<0.8	>49.2	all 3 channels	+1–2 dB

• Uniform allocation (cyclic) evens out MSE at the cost of lower per-pixel capacity, yielding best imperceptibility and robustness.
• OPAP universally reduces distortion by up to 50%, delivering PSNR gains of 1–2 dB.

In diffusion-based LCG, the derived scaling avoids decay of color guidance in late timesteps (contrasting universal guidance), securing consistently high color fidelity at extremely low bitrates. In VR/AR, LCG settings that modulate $b=b_1b_0$7 in $b=b_1b_0$8–$b=b_1b_0$9 ROIs at $n = \text{Decimal}(b) + 3$0 Hz are effective for subtle, rapid, and robust visual guidance.

7. Domain-Specific Applications and Extensions

• Image steganography: Adaptive LCG mechanisms ensure payload anonymity and even error spread (Amirtharajan et al., 2010).
• Diffusion-based generation and compression: LCG enables non-destructive, reference-driven color conditioning; exact guidance scaling generalizes to any linear constraint (e.g., segmentation, sketch) (Bordin et al., 2024).
• Reference-based video colorization: LCG augments diffusion-transformer architectures for temporally coherent, fine-grained animation colorization (Zhang et al., 27 Jul 2025).
• Perceptual guidance: LCG exploits early visual system properties to enable unobtrusive, real-time gaze guidance in interactive environments (Tosa et al., 2024).

A plausible implication is that LCG techniques will see further convergence, crossing over between representation learning and perceptual science, as all rely on the notion that fine-grained, local color structure is a powerful and versatile guiding signal—whether for machines or for humans.