From Code to Kodachrome: Film Emulation from Scratch

The input to this algorithm is not actually an image, but a video. It takes the first frame of the video and treats it as a photograph. This is due to my utilization of the Apple ProRes Log video, which the iPhone 15 and 16 Pro/Pro Max models can record in.

I found LUTs by Tobia Montanari which converts from Apple ProRes Log to Rec.709 with filmic colors, modified from the LUTs from DaVinci Resolve (Montanari 2025).

Using Log video probably wasn't completely necessary: a RAW photo would have provided a lot of data. But Apple ProRes Log applies less processing than a RAW photo since RAW photos in iPhones still apply quite a bit of processing. Using a DSLR camera would avoid this problem, but I chose not to pursue that route for this project. With a little more effort, you can convert the current algorithm to input a raw image from a DSLR camera, though.

Here's what a Apple ProRes Log "photo" looks like compared to the photo with the filmic LUT applied.

Lastly, I convert from Rec.709 to linear color space, since linear better represents the physical radiance of light. This is useful for any effect we are applying since it gives it more of a physical basis. Here's the image in Rec.709 compared to linear color space.

For the rest of the article, assume all processing is done in linear color space unless otherwise specified. All photos will be displayed in Rec.709 color space from now on since it shows details in the shadows.

Bloom is an effect where bright areas of an image bleed over into darker areas. Graphics programmers have adopted the term "bloom," but the effect is predominantly due to two physical phenomena. Bloom is present in any optical system, including the human eye. However, it is exaggerated in film due to film emulsion, especially in small film stocks like 8mm (Dehancer Blog).

Bloom is not a part of my implementation because I was unaware of how much film contributes to the effect at the time of writing my program. If I were to revisit this project, bloom will be on the list of features to add, but it will require special consideration to avoid overwhelming the halation effect.

Ironically, the first step to adding grain is to remove noise. Having both noise from the digital camera as well as the noise of film grain can be distracting. I use OpenCV's cv::xphoto::dctDenoising function with the parameters sigma=7.5 and psize=8. At times, the denoising is aggressive and creates a blurry image, but this is masked by the addition of film grain. This may also be a beneficial artifact, since film images are softer than a digital photograph.

Since denoising is subtle in the example image, only the cropped image is shown.

I use a modified technique from a paper titled "Realistic Film Grain Rendering" published in the Image Processing On Line journal (Newson et al. 2017). It uses a Monte Carlo technique to synthesize film grain by simulating the literal grains of film, which takes roughly three minutes to compute on my M4 MacBook Pro using all CPU cores.

The technique in Newson 2017 is for a single color channel only. The paper proposes a solution:

However, my testing disagrees with this conclusion: applying film grain to each channel independently creates an undesirable color noise effect, not seen in most film photography. An interesting exception are some images taken by a friend of mine on a Fujifilm QuickSnap Flash 400 disposable camera, where color noise is clearly present. Notice how the film grain appears to distort the color.

Photo credit: Natalia Builes

In my film emulation software, do the following steps to apply film grain:

In real film, there is little grain in the deep shadows of an image (Sudhakaran 2025). Applying film grain to a linear image exaggerates the grain in shadows when converting back to Rec.709. Converting to Rec.709 before applying film grain is a simple fix (step 1).

The application of film grain to the monochrome image (step 3) and adding the film grain texture to the original image (step 5) required a lot of trial and error with parameter adjustment. Here's what works:

Step 3: I have an average grain radius $\mu R=0.14$ and the standard deviation of the grain radius equal to $\sigma R=0.075 \cdot \mu R$. I use $\sigma_{filter} = 0.875$, which removes high-frequency noise from the image. I also use $N=350$ Monte Carlo iterations, which balances performance and quality.

Step 5: I use the constant 0.525, which adds film grain without it being too exaggerated. This parameter could be reduced some more, too.

To summarize, the full pipeline is as follows:

After I run my algorithm, it is usually necessary to make minor adjustments to the colors to improve the image.