Abstract
While digital photography offers technical precision, it often lacks the aesthetic "soul" of analog film, characterized by organic grain, non-linear color responses, and unique textures. This paper introduces FilmSet, the first large-scale dataset specifically designed for film stylization, containing over 5,000 high-resolution images across three iconic film styles. To effectively utilize this data, we propose FilmNet, a multi-frequency framework that employs Laplacian Pyramid decomposition to independently refine color and texture. Our approach significantly outperforms state-of-the-art image enhancement methods, achieving a PSNR of 40.07 on the Cinema style and demonstrating superior generalization on standard benchmarks like MIT-Adobe FiveK.
1. Introduction
Simulating the appearance of analog film is a complex task that extends beyond simple color grading. Film aesthetics are the result of chemical processes that introduce specific graininess and non-linear color shifts across different image frequencies. Traditional digital enhancement methods, such as 3D Look-Up Tables (LUTs), often struggle to replicate these high-frequency details, resulting in images that appear either muddy or overly artificial.
Two primary obstacles have hindered progress in this field: the scarcity of large-scale datasets dedicated to specific film aesthetics and the failure of existing architectures to account for the multi-frequency nature of film grain and color. This work addresses both by providing a comprehensive dataset and a novel frequency-driven architecture.
2. The FilmSet Dataset
The authors present FilmSet, a high-quality dataset featuring 5,285 images. Each image is processed into three distinct "recipes" based on world-class film simulations:
- Cinema: Designed for a dramatic, cinematic aesthetic.
- Classical Negative (ClassNeg): A nostalgic, high-contrast look.
- Velvia: Optimized for vibrant, high-saturation landscape photography.
This dataset provides the necessary ground truth for training models to understand the nuanced relationship between digital input and authentic film output.
3. Methodology: FilmNet
The proposed FilmNet architecture operates on a "divide and conquer" principle, separating image components by frequency to allow for specialized processing.
3.1 Laplacian Pyramid (LP) Decomposition
FilmNet utilizes Laplacian Pyramids to decompose the input image into high-frequency bands (representing edges and textures) and low-frequency bands (representing global color and lighting). This separation is critical because film style is inherently multi-frequency: color is a low-frequency property, while grain is a high-frequency property.
3.2 Nonlinear Stylization Remapping (NSR)
The low-frequency components are processed through the NSR module, a lightweight, UNet-like structure. This module handles precise color transfer without the computational overhead of heavy activation functions, ensuring efficient global tone mapping.
3.3 Multi-Scale Reconstruction Module (MSRM)
To handle high-frequency details, the MSRM uses "Global-Aware Convolutions." This allows the model to learn how to apply film-like textures and grain appropriately across both highlights and shadows, avoiding the uniform "overlay" look of traditional filters.
3.4 Triple Trilinear Regulator (TTR)
In the final stage, the model employs a Style-Aware Adjuster to predict weights for three different 3D LUTs. This dynamic weighting allows for more flexible and accurate stylization than a single, rigid LUT could provide.
4. Experimental Results
FilmNet was evaluated against state-of-the-art methods on both the FilmSet dataset and standard benchmarks.
- Performance on FilmSet: FilmNet achieved a PSNR of 40.07 on the Cinema style, significantly outperforming the next best competitor, LPTN (36.55).
- Color Accuracy: The model achieved a $\Delta E$ (color error) of 1.47 for the ClassNeg style, indicating high fidelity to the target film aesthetic.
- Generalization: On the MIT-Adobe FiveK dataset, FilmNet outperformed specialized models like SepLUT and 3D-LUT, achieving a PSNR of 25.20 and an SSIM of 0.906.
Qualitative analysis confirms that FilmNet maintains deep blacks and organic textures, avoiding the "washed out" appearance common in other digital stylization techniques.
5. Conclusion and Future Work
FilmNet and FilmSet represent a significant step forward in digital film simulation. By treating color and texture as distinct frequency-dependent tasks, the model achieves a level of realism previously unattainable with global transforms.
Future work may address current hardware limitations, such as the CUDA memory constraints that limited the depth of the Laplacian Pyramid in this study. Additionally, extending the framework to ensure temporal consistency in video sequences would allow for the application of these film styles to motion pictures, where consistent grain behavior is paramount.
Summary updated