LLF-LUT++: Efficient Image & Video Filtering

• LLF-LUT++ is a family of lightweight, real-time image and video enhancement frameworks that use learned, compact LUTs to balance quality and resource efficiency.
• It employs SVD-based factorization, bilateral grids, and Laplacian pyramid decomposition to achieve spatial adaptivity, reducing LUT parameters by up to 88% while preserving fidelity.
• LLF-LUT++ demonstrates competitive performance in 4K photo processing and in-loop video filtering, significantly lowering computational overhead and storage requirements.

LLF-LUT++ refers to a family of lightweight, real-time image and video enhancement frameworks built on the principle of learned, highly compact look-up tables (LUTs), which are designed to replace or augment conventional deep neural network (DNN)-based or hand-engineered methods in both image enhancement and in-loop filtering in video coding. Distinguished by their aggressive factorization, adaptive parameterization, and efficient interpolation strategies, LLF-LUT++ systems consistently demonstrate competitive perceptual quality and resource utilization across diverse vision tasks, including photographic retouching, tone mapping, and artifact suppression in video codecs (Kim et al., 22 Aug 2025, Zhang et al., 13 Oct 2025, Li et al., 11 Sep 2025, Li et al., 2024).

1. Core Design Philosophy and Application Scope

LLF-LUT++ addresses the longstanding tension between the visual performance of DNN-based enhancement systems and the speed, deployability, and memory footprint constraints typical of edge devices, mobile hardware, and embedded video decoders. Classical pointwise 3D LUT approaches offer rapid color transformations but lack context sensitivity, while spatially-aware methods relying on U-Nets or similar architectures incur excessive runtime and parameter counts. LLF-LUT++ frameworks close this gap by restoring spatial adaptivity at minimal marginal cost: they exploit the redundancy of classical LUTs through SVD-based factorization, combine global and local operators (often using Laplacian pyramids or bilateral grids), and leverage tailored interpolation mechanisms to preserve accuracy while keeping memory, compute, and DRAM traffic tightly bounded (Kim et al., 22 Aug 2025, Zhang et al., 13 Oct 2025).

Crucially, LLF-LUT++ sees use both in high-fidelity, real-time image enhancement (including 4K photo processing) (Zhang et al., 13 Oct 2025), and as a practical “lookup-tables instead of CNNs” technology for in-loop filtering (ILF) within modern video coding pipelines such as Versatile Video Coding (VVC) (Li et al., 11 Sep 2025, Li et al., 2024).

2. Mathematical Formulation and LUT Factorization

Traditional LUT-based systems rely on dense 3D tables mapping input triplets (e.g., RGB) to output values: for each output channel $c$, a full LUT is $$t^c_{rgb} \in \mathbb{R}^{D_t \times D_t \times D_t}$$. LLF-LUT++ replaces these monolithic structures by linear sums of low-dimensional, efficiently parameterized entities:

• 3D-to-2D LUT Decomposition: Each 3D LUT is replaced by a blend of three 2D LUTs per output channel,

$$t^c_{rgb}(r,g,b) \rightarrow w^c_{rg}\, t^c_{rg}(r,g) + w^c_{rb}\, t^c_{rb}(r,b) + w^c_{gb}\, t^c_{gb}(g,b) + b^c$$

where weights and bias are regressed per image (or per spatial region) by a small MLP (Kim et al., 22 Aug 2025).

• SVD Factorization of 2D LUTs: Each 2D LUT

$$t^c_{\alpha\beta} = U^c_{\alpha\beta} S^c_{\alpha\beta} (V^c_{\alpha\beta})^\top$$

where $U, V \in \mathbb{R}^{D_t \times k}$, $S \in \mathbb{R}^{k \times k}$, and $k \ll D_t$. The SVD rank $k$ is chosen to minimize parameter count: e.g., for $D_t = 33,\,k = 8$, this yields an 88% reduction in LUT parameters while maintaining PSNR within 0.02 dB of the full-rank solution. Components are predicted directly from image features via MLPs and reconstructed during inference (Kim et al., 22 Aug 2025).

• Image-Adaptive Blending: LLF-LUT++ learns both global and per-pixel blending weights (biases), parameterized by image-wide statistics (e.g., from a Laplacian pyramid’s coarsest level, or a global context vector extracted by a lightweight convnet or transformer) (Zhang et al., 13 Oct 2025).

3. Spatial Adaptivity and Feature Fusion

LLF-LUT++ achieves spatial adaptivity while preserving efficiency through several mechanisms:

• Bilateral Grid and Feature Slicing: Spatially varying features are generated via low-rank decomposed bilateral grids. Slicing at pixel $(x,y)$ and local value $$t^c_{rgb} \in \mathbb{R}^{D_t \times D_t \times D_t}$$0 forms spatially-adaptive channels, which are linearly combined with LUT outputs to produce a per-pixel map. By interleaving slicing and LUT lookup, LLF-LUT++ reduces DRAM traffic and avoids large intermediate tensors (Kim et al., 22 Aug 2025).
• Laplacian Pyramid Decomposition and Local Laplacian Filters: In image enhancement, a closed-form Laplacian pyramid isolates low- and high-frequency content. Image-adaptive 3D LUTs perform coarse, global tone mapping on low-pass representations, while learned local Laplacian filters—parameterized by small CNNs—remap each high-frequency pyramid level, enabling edge-preserving detail and artifact suppression. Pyramid synthesis reconstructs the final output without loss (Zhang et al., 13 Oct 2025).
• Spatial-Frequency Transformer Weighting: Weight fusion modules employ jointly spatial and frequency-attentive transformers (e.g., Fourier-Swin) to compute blending weights for both basis LUTs and Laplacian branch fusion. This enables global and local enhancement strategies to coexist in a single architecture (Zhang et al., 13 Oct 2025).

4. Practical Implementation for In-Loop Filtering

In video coding, LLF-LUT++—also referred to as LUT-ILF—replaces computationally intensive neural in-loop filters with aggressively quantized, interpolated LUT cascades (Li et al., 11 Sep 2025, Li et al., 2024):

• LUT Construction: A compact DNN is trained for reference-range-restricted mapping (e.g., a 4-input CNN for a 2×2 patch). The DNN’s outputs are enumerated across sampled (MSB-only) index points, cached into the LUT, and then fine-tuned (via the same loss) with the entire LUT treated as trainable parameters.
• Hierarchical and Cross-Component Indexing: Multiple LUTs—spatial (4D), progressive (cascaded to expand receptive field from 5×5 to 17×17), and channel-wise (for YUV)—operate in parallel, their outputs fused via learned weights. Cross-channel coupling for chroma planes is realized through joint LUTs, and diagonally dominant subspaces are compacted for memory efficiency (Li et al., 11 Sep 2025).
• Efficient Interpolation: Due to coarse MSB-only sampling, runtime lookup of LUT entries employs high-dimensional trilinear or barycentric (4-simplex) interpolation. Pseudocode for practical deployment in VVC is provided in (Li et al., 11 Sep 2025).
• Complexity/Storage Profile: On standard settings, LUT-ILF U/V/F modes achieve 0.13/0.40/0.93 kMACs/pixel and 164 KB–1.15 MB storage, with ≤108% decoder time (vs. 4717% for NNVC-HOP). In typical conditions, BD-rate reductions are 0.13–0.51% (AI) and up to 0.39% (RA) (Li et al., 2024).

5. Empirical Performance: Image Enhancement and Video Coding

LLF-LUT++ systems have been benchmarked across photographic enhancement and video coding tasks, consistently achieving competitive fidelity within resource-constrained regimes.

Image Enhancement — FiveK Photo-Retouch (480p):

Method	Params	PSNR	SSIM	ΔE_ab
3DLUT	593K	25.29	0.923	7.55
SepLUT	120K	25.47	0.921	7.54
AdaInt	620K	25.49	0.926	7.47
SAB-LUT	464K	25.66	0.930	7.29
LLF-LUT++	161K	25.76	0.931	7.26

At higher resolutions and on HDR+:

• 4K, HDR+: LLF-LUT++ achieves 27.96 dB PSNR, 0.882 SSIM, 13.5 ms per 4K frame, with +2.64 dB gain over LLF-LUT and an order-of-magnitude speedup compared to heavy DNNs (Zhang et al., 13 Oct 2025).

In-Loop Filtering — VVC / VTM-11.0:

Method	BD-rate AI	MACs/pix	Storage	Dec Time
NNVC-HOP	−7.9%	477k	2.8–7MB	4717%
LUT-ILF Fast	−0.51%	0.93k	1.1 MB	108%

LLF-LUT++ and variants thus deliver most of the coding gain of neural ILFs at a fraction (≈1/200×) of the memory and computational cost, enabling real-time deployment on CPUs (Li et al., 2024, Li et al., 11 Sep 2025).

6. Ablation Studies and Architectural Insights

Systematic ablations highlight the significance of each architectural component:

• SVD factorization of LUTs can reduce parameterization by up to 88% while maintaining PSNR within 0.02 dB,
• Number of fused LUTs (three vs. two) in color transforms affects quality by ~0.1 dB,
• Spatial fusion via bilateral grids provides ~0.27 dB gain,
• Removal or replacement of Laplacian modules or localized filters leads to 0.5–1.3 dB performance degradation in enhancement (Kim et al., 22 Aug 2025, Zhang et al., 13 Oct 2025).

LUT-ILF ablations confirm that progressive indexing, learnable weights, and CTU-level rate-distortion optimization are all essential for maximizing BD-rate reduction (Li et al., 2024).

7. Limitations and Future Research Directions

While LLF-LUT++ sets new Pareto-optimal benchmarks, several limitations remain:

• On extremely resource-limited (single-core) CPUs, real-time may remain elusive; possible mitigations include smaller backbones or minimal LUTs (e.g., NILUT).
• Storage scales linearly with the number and order of patterns/LUTs, suggesting further factorization or compression opportunities.
• Adaptive SVD rank selection conditional on input may yield additional savings.
• The decomposition, factorization, and cache-fusion techniques pioneered here are plausibly extensible to video restoration, neural implicit operators, or domain-general hardware-efficient vision modules (Kim et al., 22 Aug 2025, Zhang et al., 13 Oct 2025, Li et al., 2024).

LLF-LUT++ thus constitutes a versatile, resource-optimal class of learned LUT architectures bridging statistical optimality and hardware practicability in modern image and video processing.