How do you define looking good ?

I once read an article/presentation/blog (can't find it at the moment, sry) in which the author sugguested, that the tone-mapping should represent a close physical mapping (not what the human eye sees) and then apply additional LDR color correction to make it look good. Therefor he sugguested to use even Reinhard over a filmic tonemapper, because the latter already included some kind of color correction.

I think, that the basic idea is to map a scene to LDR in a stable and (physical) color true way and leave all the 'making it pretty' to the color correction mapping.

Well if you're going with physically based stuff a filmic tonemapper: http://mynameismjp.wordpress.com/2010/04/30/a-closer-look-at-tone-mapping/

They're all, supposedly if the name means anything, based off the tonemapping for actual film, which deals obviously with physically base "real world" stuff anyway. There was this huge spiel with data, a pdf, I read months ago explaining exactly why film exposes in that way and how it suits real world scenarios, including real world under/over exposure clipping and light ranges and etc. If anyone knows what I'm babbling about and has it bookmarked then that's the most helpful thing I can think of. Real world light intensity of everything from a moonless night to the middle of a bright day and etc.

One of the specific I can remember was a reason for the toes on either end spreading out in an S shape, and that was to avoid sharp clipping in either over or under exposure, as well as ensuring there's still some saturation in both relatively dark shadows and bright highlights.

It sounds like you're talking about this presentation, which is from the SIGGRAPH 2010 course about color enhancement and rendering.

This probably isn't helpful to anyone at all but it's as good of an excuse as any to post my thoughts, I guess.

I'm not using a tone-mapping technique at all, at least not by most definitions. I'm ensuring that my lighting (by which I mostly mean lightmaps) is in the 0-1 range, and then applying an RGB curve (using a texture lookup as described here) to my rendered image. I'm not overly worried about color quantization so far, as I'm not really planning to stretch the values a great deal.

The obvious disadvantages to using an RGB LUT is that a) it assumes independence in the red, green, and blue channel (meaning it's impossible to, say, desaturate the image by this method alone) and b) that it doesn't make any attempt to model a real physical process.

One could, at least in theory, get past a) by using a 3D LUT, but so far I've found that if I grade my image with an actual color grading program, I don't usually violate the independence assumption too heavily. That means I can just apply my grade to a simple gradient and generate a 1-dimensional RGB LUT that more or less captures the same look. I think I will probably add a couple of other features to my algorithm to, for instance, crush saturation at high luminosity, but I haven't even made it that far yet.

As for b), I tend to think that many attempts to model film (or whatever) physically are wasted, because all of my color data is stored as RGB values rather than full spectral information. This is why things like FilmConvert, which presumably uses something very much like a 3D LUT, work well -- but not perfectly: they still only have RGB data as input. For an example of why this isn't always sufficient, consider that for every sensor (or film stock, or even eyes, excluding those fancy people with tetrachromatic vision), there's some color that includes spectral yellow that will be indistinguishable from a color that includes red and blue, but no spectral yellow. The challenge comes from the fact that, between sensors, which colors are indistinguishable can vary, so if we only store RGB information, it's always possible to lose something important.

It'd be fascinating if we could practically work with color that is represented by more than just three channels, and I know that this is not uncommon in some offline renderers, but it would mean we'd need to be able to work with spectral data in every intermediate step. For instance, a lot of extra work (or at least equipment) would be needed for texture acquisition from real-life (not to mention creation of textures not based on reality) to make use of it. This would need to be taken into account for every light bounce, etc. all the way up until we eventually get to simulating what happens once the light actually reaches the camera.

The way I see it, since I'm doomed from the start anyway, I'm better off just doing something simple and tweaking it to taste rather than worrying about simulating certain things accurately when I'll always be missing pieces.

I'm sure none of this is new to a lot of people in this thread, and I doubt my decision to ignore almost everything just because I am forced to ignore something is satisfactory to most people, but maybe it'll be helpful to someone anyway.

Thanks for the replies, links and app suggestions

How do you define looking good ?

except for the light sources themselves, which saturate to white, bloom out and burn into my retina

...but yes, this is a very subjective thing: I want to use physical values and end up with a rendition that's artistically appealing.