Abstract
This paper introduces FilmSet, the first large-scale, high-quality dataset specifically designed for film-style color grading, featuring over 5,000 images across three distinct film aesthetics. To accompany the data, the authors propose FilmNet, a deep learning framework that uses a Laplacian Pyramid to separately enhance image features across different frequency bands. This approach significantly outperforms state-of-the-art methods in both accuracy and visual appeal, bridging the gap between digital convenience and analog aesthetics.
Introduction
While digital photography offers unparalleled convenience, many photographers still seek the unique "soul" of analog film—characterized by specific color science, grain, and a timeless aesthetic. Currently, achieving this look digitally relies on Look-Up Tables (LUTs) or labor-intensive manual editing. These methods often result in artifacts because they apply global changes without considering local image details.
A major hurdle in automating this process has been the lack of dedicated, large-scale datasets for film stylization. Existing benchmarks like MIT-Adobe FiveK focus on general enhancement rather than specific stylization. Furthermore, standard deep learning models often struggle to balance global color shifts with the preservation of fine textures and edges.
Methods
The authors propose FilmNet, a novel architecture based on the principle of multi-frequency decomposition. The model operates in several key stages:
- Laplacian Pyramid Decomposition: FilmNet breaks the image into different frequency bands, allowing it to treat color (low frequency) and texture/edges (high frequency) as separate tasks.
- Nonlinear Stylization Remapping (NSR): Low-frequency components are processed through a lightweight UNet-style architecture to handle complex color transformations.
- High-frequency Refinement: A Multi-Scale Reconstruction Module (MSRM) uses global-aware convolutions and attention mechanisms to protect important textures and edges during stylization.
- Triple Trilinear Regulator (TTR): A learned 3D LUT approach performs final global refinement, ensuring the output maintains a consistent tone characteristic of specific film stocks.
Datasets
The primary contribution is FilmSet, a dataset containing 5,285 high-resolution images. Each image is paired with three "ground truth" versions corresponding to Fujifilm-inspired styles: Cinema, Classic Negative, and Velvia. These were generated using professional-grade recipes in Capture One to ensure high fidelity to analog looks.
In addition to FilmSet, the authors utilized the MIT-Adobe FiveK and HDR+ datasets to evaluate the model's general enhancement and high dynamic range capabilities.
Key Results
FilmNet achieves state-of-the-art (SOTA) performance across multiple benchmarks. Quantitatively, on the FilmSet "Cinema" style, FilmNet achieved a PSNR of 40.07, significantly higher than other SOTA methods like LPTN (36.55) or 3D-LUT (35.49). It also consistently demonstrated the lowest $\Delta E$ (color difference) scores, indicating high faithfulness to target film recipes.
Qualitatively, visual comparisons show that FilmNet avoids the "muddy" or over-exposed look common in standard CNN-based enhancers. By separating frequency bands, the model preserves the organic sharpness of the original image while applying sophisticated color grading.
Conclusion and Future Work
FilmNet successfully bridges the gap between modern deep learning and traditional color grading. By creating FilmSet, the authors provide the necessary "vocabulary" for AI to understand analog aesthetics. Future work may involve expanding the dataset to include iconic stocks from Kodak and Agfa, as well as integrating explicit grain synthesis to further enhance the organic feel of high-ISO film simulations. Current limitations include hardware constraints for ultra-high-resolution processing due to the depth of the Laplacian Pyramid.
Summary updated