Comment by strogonoff | Hacker News Reader

Comment by strogonoff | Hacker News Reader

Parent Full thread

Clipped highlights in digital photography are simply impossible to eliminate in post-processing without conjuring nonexistent information. Even if you shoot raw. Different raw processors use different tricks, color propagation and such, but naturally they are best used when highlights are really small. I would not be surprised if tools like Lightroom invoke ML at the first hint of clipping (because why not, if you all you have is a hammer…).

Pro tip: Digital sensors are much less forgiving than negative film when it comes to exposing highlights. With a bit of foresight they are best tackled at shooting time. Highlights from water/glass reflections are tamed by a fairly cheap polarizing filter, and if you shoot raw you should do the opposite of negative film and always underexpose a scene with bright highlights (especially if highlights are large or are in your subject of interest). Let it be dark, you will have more noise, but noise is manageable without having to invent what doesn’t exist in the parts that are the most noticeable to human eye.

> Clipped highlights in digital photography are simply impossible to eliminate in post-processing without conjuring nonexistent information. Even if you shoot raw.

Huh? This used to be true twenty years ago, but modern sensors in prosumer cameras capture a lot more dynamic range than can be displayed on the screen or conveyed in a standard JPEG. If you shoot raw, you absolutely have the information needed to rescue nominally clipped highlights. You get 2-4 stops of latitude without any real effort.

The problem here is different. If the rest of the scene is exposed correctly, you have to choose one or another. Overexposed highlights or underexposed subject. The workaround is to use tone mapping or local contrast tricks, but these easily give your photos a weird "HDR" look.

You may be confused on at least two counts:

— If you think there is “clipping” and the somehow different “nominal clipping”.

Clipping is binary. Either you camera’s pixel clipped (i.e., accumulated enough photons to get fully saturated, and therefore offers no useful information for debayering), or it did not. If it clipped, then you lost colour data in that part of the image. If you lost colour data, and you want a colour photo, then in almost all cases one way or another you will have to rescue it by conjuring information up for the photo to look authentic and aesthetically good. (A lot of raw development software does it for you by default.) If the clipped bit is small, it is easy to do subtly. If it is big, which is a real danger with sun reflection bokeh, then… whoopsie.

— If you think modern sensors are safe from clipping.

This applies to any digital camera, from consumer to top professional models. Sensor bit depth is not even remotely enough to expose a scene with extreme highlights (e.g. sun or its reflections) without taking special measures regarding exposure at capture time. If you don’t tame highlights with a polarizing filter, you must dramatically underexpose the scene or they will clip.

They specifically clarified that they were talking about the JPEG/picture on screen, which has less dynamic range than the sensor and can clip without losing pixel information.

We are specifically discussing scene-referred image data—the raw pixel values obtained from a digital camera sensor. That is where unrecoverable information loss occurs at capture time. Parsing that data is when LR et al. engage various highlight reconstruction algorithms that fill in the missing information. My point is that a lot of that loss is often avoidable at capture time.

nah-- even in the supplied jpg the histogram shows they're not clipping the highlights, and if you crank down the brightness, you can see the detail.

You may be confusing display-referred and scene-referred image data.

1. Clipping occurs in the scene-referred raw camera sensor pixel data at capture time.

2. Raw processing software fills in the highlight by supplying the missing colour information. Often this happens by default.

3. When you obtain a display-referred JPEG, the missing information for clipped highlights was already supplied. Tone curves and other adjustments apply already after highlight reconstruction, making the final result look more organic.

In other words, with modern digital photography processing workflows you will never see any clipping on a histogram for the final JPEG (unless either something went really wrong or it was intentionally done for a look), and it is a poor foundation to make assumptions about what clipping had or had not occurred at capture stage.

What you can safely assume, however, is that a normally exposed digital photo featuring any sun reflections from water or glass will almost certainly have blown highlights.

Without the raw image, we won't really know whether there was clipping. However, it is clear that the poster here wanted to remove the highlights and was okay with introducing some information that wasn't in the photo to do it, hence the use of an AI filter. That means that there is no issue with using the clone stamp or the paintbrush tool.

This sounds about right. However, regarding your first sentence, I’d be a bit less cautious: sun reflections from water on a clear day guarantee clipping unless a polarizing filter is used or the scene is dramatically underexposed, and based on the general look of the photo (even knowing it’s after processing) my intuition suggests me that neither took place.

Hence my original advice: don’t leave it up to ML when a little forethought can do wonders in terms of preserving information.

What are you talking about? I see clipped histograms come out of cameras all the time. I’ve been working with digital images and digital photo manipulation for thirty years.

Regardless, it’s not generative AI making up details. The differentiated pixels are there in the camera jpeg even if traditional image processing techniques have to be used to make them visible onscreen. The complete structure of the feathers that isn’t visible in the camera jpeg, for example, is plainly visible just modifying the contrast.

Subject domain knowledge mismatch appears to be too strong for further constructive discussion.

Reply on news.ycombinator.com