LAM Guides Β· Fundamentals
As more apps ship an "AI enhance" button, the question is worth asking plainly: when an app says your photo looks "better," is it re-grading the colours in the pixels your sensor captured β or is it repainting part of the image with a generative model? The two sound similar. They aren't.
AI filter
Film-style grading (LAM)
A typical "AI enhance" filter runs your photo through a model trained to guess a nicer-looking version of it β smoothing skin, inventing detail in soft areas, sometimes reconstructing whole regions that were underexposed. The model has no idea what the light in your room actually looked like; it only knows what millions of other photos look "good," and tries to nudge yours toward that. The result can look pleasing, but part of the detail in the frame was invented by the model, not recorded by the lens.
A film-emulation engine works on a completely different principle: it applies a fixed colour transform β usually packaged as a 3D lookup table (LUT) β to the exact RGB values of the pixels you already captured. No pixel gets repainted, no detail gets guessed. A shadow is still real shadow data, just remapped along the contrast curve of whichever film stock is being emulated.
A quick way to tell the difference: shoot the same scene twice in a row. An AI filter can return two slightly different results in the detail areas (because the model "guesses" a little differently each time). Film-style grading always returns the same result for the same input β it's a fixed mathematical transform, not a new generation each time.
Many photographers instinctively shoot with the default camera app on their iPhone or Android, then import the photo into editing apps to slap on a film preset or LUT. It feels like the "safe" workflow. But from an imaging pipeline perspective, this is precisely the path that introduces the most degradation and destroys the organic look of film.
The root problem: your photo undergoes two conflicting, destructive processing passes (double distortion).
Default Camera + Post-Filter
Real-time Film Engine (LAM)
1. Tone Curve Clash: Computational photography algorithms in modern smartphones (Apple Photonic Engine, Deep Fusion, Smart HDR and Google HDR+) are tuned to please mainstream consumers: they strive to expose every corner by aggressively crushing highlights and unnaturally lifting shadows. This flattens the optical depth of the scene. Meanwhile, the very soul of chemical film is natural contrast (rich, deep shadows and smooth highlight shoulders). When you paste a film preset on an already flattened HDR file, the film curve tries to pull down areas that the phone artificially brightened. The result: shadows collapse into murky, grayish mud (muddy blacks), and highlights turn chalky and dirty.
2. Edge Halos vs Organic Grain: To make photos appear "sharp" on tiny screens, default camera pipelines apply heavy unsharp masking, producing halo artifacts around hair and high-contrast edges, paired with aggressive noise reduction that smears skin texture into a flat watercolor look. When a post-processing app sprinkles synthetic film grain on top of this, the grain clusters harshly around these artificial halo edges and sits awkwardly on waxy, smoothed skin. Instead of the delicate, organic texture of silver halide crystals, you get a gritty, digital mess.
3. Two grading layers that know nothing about each other: This is the real breakage, and it is about the number of passes, not the number of bits. The first layer is the phone's algorithm, decided before you could look. The second is the preset you paste on hours later, designed for a photograph that had not been bent yet. Neither layer knows what the other did to the shadows and highlights, so the result is the sum of two conflicting decisions rather than one correct one.
Stated plainly: LAM does not read RAW, and does not switch off your phone's HDR.
LAM receives its frame from the phone's ISP like any other camera app. On iPhone it deliberately asks for Deep Fusion and Smart HDR at the highest setting the device allows, because measured on real hardware that frame is cleaner in the shadows and wider in range β turning it off makes the picture worse, not better.
What makes LAM different is not "reading the raw sensor". It is three things we can measure: (1) there is one grading layer instead of two stacked ones; (2) you see that layer in the viewfinder before the shutter, so a wrong choice is changed on the spot rather than discovered at home; (3) on Android, LAM picks the capture mode that skips CameraX's extra multi-frame noise-reduction pass β A/B'd on device, flat areas came out 17% cleaner at equal detail. We would rather state three things that are true than one that sounds impressive.
The LAM Approach: Rather than shooting first and then trying to "repair" the result with a cosmetic filter, LAM decides the colour before the shutter. The film's sensitometric curve, halation and grain run on the live preview stream, so what you see in the viewfinder is the file you are about to get β and a wrong choice is changed on the spot instead of discovered at home. The photograph leaves with one grading layer on it, not one stacked on an earlier one it knows nothing about.
The same colour logic runs for the still at the shutter, for video while recording, and for the
.cube you export to Resolve or Premiere β one source of colour, not three that merely
resemble each other. That is the part no post-filter can match, because by the time it runs the
photograph is already settled.
π Read more: Deep dive: The "double-processing" trap of shooting default camera then applying a filter.
Not because AI or editing apps are inherently bad β but because the processing order dictates final optical quality. If you want to keep the moment you actually captured β a meal, an evening with family, a trip β you usually want the photo to stay faithful to what the lens saw, carrying true optical depth rather than having a default camera artificially homogenize the light, only for a post-filter to distort it a second time.
In LAM this rule is kept consistent end to end: the frame you see in the viewfinder while shooting has already gone through the same grading engine that saves the final photo or video β there's no separate "AI pass" bolted on afterward. What you frame is what you get.