| --- |
| library_name: pytorch |
| pipeline_tag: image-to-image |
| license: apache-2.0 |
| tags: |
| - image-to-image |
| - color-grading |
| - 3d-lut |
| - computer-vision |
| - pytorch |
| --- |
| |
| # Nolanizer V1 |
|
|
| Nolanizer V1 is a scene-adaptive image-to-image model for cinematic global |
| color grading. It preserves the input composition and predicts a color |
| transformation designed to produce a restrained, large-format cinematic look |
| associated with Christopher Nolan-inspired visual language. |
|
|
| The model analyzes each image, predicts a mixture of eight learned `33³` 3D LUT |
| bases, and applies bounded exposure, contrast, saturation, temperature, and tint |
| adjustments. A single intensity control blends continuously between the |
| original image and the full grade. |
|
|
| > Nolanizer is an independent research project. It is not affiliated with, |
| > endorsed by, or an official product of Christopher Nolan, any |
| > cinematographer, colorist, or film studio. The output is an algorithmic |
| > interpretation of a cinematic visual style, not a claim about how a specific |
| > person would grade an image. |
|
|
| ## Model details |
|
|
| | Property | Value | |
| | --- | --- | |
| | Task | Scene-adaptive global image color grading | |
| | Framework | PyTorch | |
| | Architecture | CNN condition encoder with a learnable 3D LUT mixture | |
| | LUT basis | 8 learned LUTs, each `33 × 33 × 33` | |
| | Global controls | Exposure, contrast, saturation, temperature, and tint | |
| | Effect range | `0.0` to `1.0` | |
| | Recommended weights | EMA | |
| | Release checkpoint | Epoch 7 | |
|
|
| The transformation is global: every output pixel is produced from its input |
| color and the scene-conditioned grade. Spatial structure, objects, and |
| composition are therefore left unchanged. |
|
|
| ## Files |
|
|
| | File | Purpose | |
| | --- | --- | |
| | `nolanizer_v1.pt` | Frozen epoch-7 checkpoint; use EMA weights | |
| | `config.json` | Architecture and inference contract | |
| | `training_config.yaml` | Resolved optimization configuration | |
| | `checkpoint_manifest.json` | Checkpoint checksum and release metrics | |
| | `requirements.txt` | Runtime Python dependencies | |
|
|
|
|
| ## Usage |
|
|
| The checkpoint requires the Nolanizer Python package or a Nolanizer source |
| checkout. Install the project, then run: |
|
|
| ```bash |
| python -m pip install -e . |
| |
| python -m nolanizer.inference \ |
| --checkpoint huggingface/nolanizer-v1/nolanizer_v1.pt \ |
| --input path/to/input.jpg \ |
| --output path/to/output.jpg \ |
| --intensity 0.8 \ |
| --weights ema |
| ``` |
|
|
| Supported inputs include standard RGB images such as JPEG, PNG, and WebP. |
|
|
| ### Intensity |
|
|
| - `0.0`: exact identity output |
| - `0.6–0.9`: recommended range for most images |
| - `1.0`: full predicted grade |
|
|
| The inference command saves the graded image and can also expose the predicted |
| LUT mixture and global adjustment parameters for inspection. |
|
|
| ## Evaluation |
|
|
| The release checkpoint passed the project's frozen color, content-preservation, |
| clipping, LUT-basis health, and exact-identity gates. |
|
|
| | Metric | Value | |
| | --- | ---: | |
| | CIEDE2000 | `5.7666` | |
| | Lab MAE | `3.6339` | |
| | Luminance SSIM | `0.8617` | |
| | Edge correlation | `0.9837` | |
| | Clipped-pixel ratio | `0.0017` | |
|
|
| These are internal release metrics rather than a perceptual preference score. |
| They should not be interpreted as a guarantee of equivalent performance on |
| every image domain. |
|
|
| ## Intended use |
|
|
| - Interactive grading of photographs and still images |
| - Research and education in global cinematic color grading |
| - Analysis of scene-conditioned LUT mixtures |
| - Non-destructive look exploration through the intensity control |
|
|
| ## Limitations |
|
|
| - The renderer applies global color and tone transformations only; it does not |
| perform local masking or selective relighting. |
| - It cannot change geometry, composition, production design, lens |
| characteristics, film grain, bloom, or halation. |
| - Extremely clipped highlights, near-black images, unusual color spaces, and |
| heavily compressed inputs may produce unstable or subtle results. |
| - Skin-tone behavior has not been assessed with a dedicated demographic |
| benchmark. |
| - Output quality is subjective and depends on exposure, white balance, scene |
| content, and the selected intensity. |
|
|
| ## Responsible use |
|
|
| Use the model only with images you have the right to process. Do not present |
| generated results as an official Christopher Nolan look, endorsement, or |
| creative decision. Keep the original image when provenance or faithful color |
| reproduction matters. |
|
|
