Here I'll explain one method to expose manually with the Canon 5D mark II when shooting fairly static scenes from a tripod, for example landscapes and architecture. It should also apply to other recent Canon models such as the 7D, all my tests have been performed on the 5D mark II though. I use this method myself.
The goal when exposing digitally is to gather as much light as possible without clipping important highlights, since more light means less noise and thus better image quality. This is sometimes referred to as "Expose To The Right" (ETTR), however the term can be quite confusing since the auto exposure function may in many situations choose to clip more highlights (expose more to the right) than you do when you expose manually. Anyway, the message is that one should concentrate on the highlights, the brighter one can make them the better exposed is the raw file. However, if you clip a highlight it is permanently lost -- digital is actually less forgiving than film -- while film will compress the highlights when nearing overexposure it is just clipped right off in digital (usually one channel at a time though). Carelessy executed ETTR technique leads to lost highlights, and therefore it has got a quite poor reputation despite that it is a perfectly valid concept.
Traditionally a light meter was used to measure incoming light and based on that suitable exposure parameters were set, but getting it right could be difficult and you could not see the result until back in the darkroom. In modern digital cameras with large LCDs (like the 5D mark II) the use of a light meter is rare, instead the auto exposure function is used as a starting point, a test picture is taken and verified in the review mode and adjustments are made accordingly. A quick loop of trial and error to tune the exposure parameters. To support this way to work the review mode provides tools to evaluate the exposure.
On the 5D mark II there is a histogram per channel (RGB histogram setting must be enabled), so you can see if any of the red, green or blue channels are clipped. Unfortunately the histograms are small and it is sometimes hard to see if there is clipping or just nearly so (I sometimes use a magnifying glass). Not only the small size makes it difficult, the histograms are dark grey on a black background so in strong light it can be hard to see where the histogram ends (clips). The other tool is highlight alert "blinkies" (not enabled per default), but the camera shows only the luminance channel (meaning that an individual channel may clip while luminance does not, so you still need to watch the histograms) and you cannot see the blinkies when zooming in.
The user interface for evaluating exposure could thus be better, and hopefully future camera models will show improvements. However, there is another problem that is much worse: the histograms and blinkies are based on the embedded JPEG and not on the raw data itself! The JPEG is embedded in the raw file in full resolution (but quite heavily compressed), it is this JPEG you see when inspecting and zooming the image in the review mode, so even if you shoot "raw only" you actually look at a JPEG. Obviously the camera user interface is intended for photographers that use JPEG directly out of the camera, and not for still life photographers that only shoot raw and aim for the best possible technical quality.
That the JPEG is shown in the review mode means that the histograms will not show the correct information, it may be similar to what this in the raw file but it may also differ significantly. So what to do? One approach is not to care, expose using the JPEG histograms in standard picture style and a white balance that looks good. This does work quite well -- with the standard settings the camera will make conservative JPEG renditions meaning that there is very little risk that the raw file will clip without the JPEG doing so. Instead there will be typically some headroom left, often around 2/3 stops, but sometimes much more. This approach will probably be good enough for 99% of anybody's pictures, but if you really want to you can apply some workarounds to make the histograms match the raw file a little bit better, or rather mismatch a bit less.
The first step is to assign "universal white balance" (uniwb) to the custom white balance. Make a long exposure with a wide open lens in a bright scene to make a 100% clipped picture (all white, all channels saturated on all pixels), then set custom white balance from that (ignore the warning). I know this trick works on 5D mark II and 7D, but may not work on all Canon cameras. If you have done it right, shooting a picture with this white balance will give the preview a green tint (see attached picture). This is the universal white balance, meaning that the RGB channels will have the same sensitivity relations as on the sensor, and since the green channel is the most sensitive, pictures will get a green tint. Unfortunately, this will not work perfectly because the maximum saturation for the channels is different on the sensor than what can be represented in the JPEG. It is a good idea to change colour space to AdobeRGB which is a bit closer to what the sensor has than the default and smaller sRGB.
Another problem is that the "picture style" setting leads to color transformations and dynamic range compression which further reduces the similarity with a true raw histogram. The least bad picture style is the "neutral", I prefer making a user-defined version of that with increased sharpening so I can more easily review sharpness too. Do not change the contrast or color settings. Some guides do recommend to reduce contrast to get a more linear histogram which in theory would match the raw file better (which is linear), but my tests showed that it is still non-linear and you also get a significant risk that the JPEG histogram will not clip when the raw file actually does. The default contrast setting seems to be the best.
So let us summarise the settings:
- Enable RGB histograms
- Enable highlight alert ("blinkies")
- Assign uniwb to custom white balance
- Set the colour space to AdobeRGB
- Set the picture style to "Neutral" (leave contrast, saturation and colour tone at default)
Now we have histograms and blinkies that are as close as possible to show clipping as in the actual raw data. Unfortunately, there may still be mismatches in some situations. A worst case is a scene with a strong colour component causing say green and blue channels be at almost equal levels, the JPEG may then "clip" without green or blue going up to maximum value, and thus no clipping in histograms despite clipping in the raw file. The best cases seem to be when the green channel is clearly ahead of both red and blue, which fortunately is almost always the case due to the more sensitive green channel.
I usually switch to universal white balance when tuning the exposure, and then switch back to a natural-looking white balance when taking the final exposure (switching white balance is quick on cameras with dedicated white balance button). In many cases it will be an overly bright JPEG with some clipping, but I prefer that as embedded JPEG rather than a green one.
If uncertain about when clipping occurs there is a safe method: instead of adjusting the exposure to just below clipping, adjust to one stop down (the first vertical line below clipping in the histogram). If the histogram was perfectly linear there would be one stop headroom in the raw file, but since JPEGs are dynamically compressed by the camera in varying degrees depending on image content the actual headroom is (according to my tests) one to two stops. More specifically, at least one stop and at most two stops, a typical headroom seems to be 1 1/3. This means that when increasing one stop from that starting point the exposure is safe, and probably only 1/3 to 2/3 stops underexposed. This is a good base exposure.
Only go for those last fractions of a stop if you really need it. Doing it every time just leads to occasional overexposures you will regret and very rarely visible quality gain in the pictures. A typical example when it may be worthwhile though is a landscape scene with a bright sky where you know that you will push the shadows in postprocessing, a scene when you would use a graduated filter if shooting on film. Fortunately those scenes often have a strong green component (as seen by the camera) making clipping indications in the histograms more reliable.
What about tiny highlights? Some highlights are so small that they do not show up in the histograms, and thus you will not see if you clip them. There is no good workaround for this, but fortunately clipping those is often exactly what you want to do, often too bright to keep anyway and too small to be meaningful. If I have many of those small highlights, I use the "safe method" described above and thus leave a little bit of headroom to keep a little bit more around those highlights.
What about larger bright highlights outside reasonable range, such as the sun? You will then have to clip, and since the highlight is large it will show up in the histogram making it more difficult to see if you clip other highlights that you want to keep. Again, aiming at one stop down and raise one stop from there can make it easier to read the histograms, since the highlights are more separated.
If the situation allows and you feel you are risking overexposure don't be afraid to bracket, especially in cases when you have to clip parts of the picture. Sometimes you need the raw converter to be able to see what is the optimal amount of clipping, and you might choose to go HDR.
What about intentional clipping of one or two channels (but not the last), and reconstruct in post-processing? If at least one channel is unclipped, a fairly good rendition of the highlight can usually be made by the raw converter. White clouds for example can do really well with only one channel left. However, the JPEG histograms are not precise enough to make these kind of decisions, when clipping occurs you start see the camera's own highlight reconstruction in the histograms so the similarity with the true raw histogram is strongly reduced. If you by some reason still want to do this, make a precise ETTR exposure and bracket in 1/3 stops up to about +1 stop from there.
What about long exposures? Tuning exposure when having to wait a minute or so per test shot is tedious. The trick there is to increase ISO and open the lens wide, tune with those settings and then increase shutter speed accordingly at target ISO and f/stop. So if you got 1/4 sec at ISO3200 and f/2.8 and want to shoot at f/8 and ISO100 that is 3 stops aperture and 5 stops ISO, 8 stops total meaning 64 seconds. Be aware that vignetting at wide apertures and high ISO can reduce reliability in this method. I don't aim to get the last fraction of a stop in long exposures.
What if you cannot shoot at base ISO, you need to increase it to make shutter speed shorter (moving objects, windy conditions etc)? The whole ETTR concept then becomes meaningless, just set the shutter speed you need and increase ISO until you have a reasonable exposure. Open up the aperture if you can and your DOF intention allows. The difference of say ISO3200 near clipping and ISO1600 one stop underexposed (raised in postprocessing) is minimal so aim for a safe exposure.
A note about the 7D: the 7D has (unlike the 5D mark II) live view histograms which can be helpful as a starting point, but it seems that these are less predictable than the JPEG histograms, so I still prefer to use them.)
