Reference-based color extraction, color rebuild, cinematic color transfer, and neural 3D LUT generation
Deep Analog is an AI color reconstruction and film-look extraction system. Given a single reference film scan, it extracts the reference color characteristics, rebuilds the look as an exportable .cube 3D LUT, and applies differentiable tone mapping, film grain, and halation to a target photograph.
Try the interactive Hugging Face demo →
- Reference color extraction — analyzes the palette, black and white points, tonal distribution, shadow tint, and highlight color of an example image.
- Color reconstruction / color rebuild — reconstructs the extracted look as a portable neural 3D LUT rather than a fixed filter.
- Cinematic color transfer — transfers the color language of a reference film scan to a new photograph while preserving image structure.
- AI color grading — predicts tone and color adjustments from an unseen reference instead of selecting from a closed set of presets.
- Film look transfer — adds highlight roll-off, luminance-dependent grain, and halation after the photometric color stage.
- LUT extraction and export — produces an industry-standard
.cubeLUT for Photoshop, DaVinci Resolve, Capture One, and other color workflows.
The network separates photometric reconstruction from optical rendering. StyleLUTNet extracts and rebuilds reference color as a 3D LUT, while FilmAnalyzerNet estimates physical parameters for tone response, grain, and halation.
The processing sequence makes each intermediate stage inspectable: reference conditioning, StyleLUT grading, tone-curve matching, film-tone processing, grain synthesis, and halation rendering.
- A reference encoder extracts the visual characteristics of an unseen film scan.
- A conditional network reconstructs a 3D LUT for global color and tone mapping.
- A film-physics stage models highlight roll-off, luminance-dependent grain, and halation.
- The LUT and physical parameters can be reused as a portable generated preset.
The separation keeps the color transform portable while reserving spatial effects for an image-domain renderer.
Different reference scans produce distinct extracted 3D LUTs and visibly different color reconstructions when applied to the same input scenes.
The current internal evaluation uses a 350-image MIT-Adobe FiveK test split.
| Configuration | PSNR ↑ | SSIM ↑ | LPIPS ↓ |
|---|---|---|---|
| Naive baseline | 21.19 | 0.9148 | 0.5167 |
| StyleLUT color stage | 22.05 | 0.9246 | 0.3080 |
| Full pipeline | 21.72 | 0.9229 | 0.4161 |
Grain and halation intentionally add high-frequency structure, so pixel-level metrics decline slightly even when the rendered result appears more film-like. Quantitative color reconstruction and perceptual film rendering are therefore reported separately.
app/ FastAPI inference application and browser UI
training/models/ Core PyTorch architecture, LUT operator, and physics modules
results/ Evaluation summaries and ablation visualizations
assets/figures/ Public teaser, architecture, and optical-effect figures
assets/demo/ Small demonstration images
The unpublished manuscript, large checkpoints, datasets, experimental notebooks, cluster-specific scripts, user-study answer material, and deployment credentials are intentionally excluded. See MODEL_WEIGHTS.md.
The public source can be inspected locally, but trained checkpoints are not distributed in this research-preview repository.
cd app
python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt
python app.py \
--style-lut /path/to/style_lut_best.pth \
--v7 /path/to/analyzer_best.pthOpen http://localhost:8000. The hosted demo is the supported way to evaluate the current trained model without private checkpoints.
Deep Analog is relevant to AI color matching, reference image color extraction, automatic color grading, cinematic look transfer, photo color reconstruction, film emulation, neural LUT generation, color palette transfer, digital intermediate workflows, and reference-to-preset generation.
This method is the subject of one or more pending U.S. patent applications. Patent pending. Publication of this repository does not grant a patent license, covenant not to sue, or permission to practice any claimed invention.
The repository is a source-available research preview and is not open source. Copyright is reserved; no permission to use, reproduce, modify, distribute, deploy, sublicense, or commercialize the materials is granted without prior written authorization. See LICENSE and PATENTS.md.
Third-party packages, pretrained backbones, datasets, and cited research remain subject to their own terms. No MIT-Adobe FiveK images or third-party model weights are included.
If you discuss the project in academic work, cite the repository metadata in CITATION.cff.
Eton (Yitong) Mu — MSCS Capstone Thesis, Rochester Institute of Technology.