Hi Selden,
First I know how an histogram works.
selden wrote:The dxt3 and dxt5 image files created by nvdxt are identical, but they are not the same as the dxt1c image. This is easily seen in the file sizes, in the color histograms, and by doing a pixel-by-pixel comparison of the images after they've been expanded into a non-compressed format (BMP in this case).
Selden beliefe me the only difference between dxt1c, dxt3 and dxt5 is the alpha channel. All these formats use exact the same compression and need excact 64bit for a 4x4 block for the RGB part of the texture. The alpha channel of dxt3/5 needs additional 64bit for a 4x4 block and therefore has these format the double size of dxt1c.
Therefore these histogramms must look identical:
I made my own test that confirm this.
Now this picture showing the error of dxt compression looks quite dramatic. Anyhow if you compare with you eyes the compressed image with the original one it's hard to see a difference. Let me try a little compare: A man stands beside a loudspeaker in a disco and you stand 10meter (30feet) away. As long as the music is on you can't hear this man speak but if the DJ turns the music of you can hear this man very clear. What I try to explain it's unimportant if there is some disturbance as long as the main signal is strong enought to mask it. Or how about this we can look at the blue scy on day but we can't see as long as the Sun is visible the stars ofcourse they are really there.
Selden, it's really hard to explain.
selden wrote:Celestia cannot restore the original colors (that were in the TGA file) that the dds file doesn't contain. The histogram of the colors of the on-screen display shows that it really doesn't try. That's what the spikes mean: they show that there are many pixels on the screen which have exactly the same color. The low values between those peaks show that there were relatively few pixels with colors in between the colors indicated by the peaks.
Absolute correct what you say.
selden wrote:I'm not so much concerned about how well Celestia reproduces the colors that are in the lossy DXT texture files. What does bother me is that it does not accurately reproduce the colors that are in the lossless U888 and PNG texture files.
Selden the question is: what is the reson for this peaks? It's really very simple.
I've built a test image that has this histogramm. Now Photoshop has not such a nice histogramm tool therefore I show only the red channel.
Now after reducing the contrast about 5% it looks like this:
And if the contrast is reduced about 10%:
Any idea how it would look with 20% or 40% contrast reduction?
But we can also increase the contrast about 10% and get now little gaps and two peacs at the edges:
Now is there any idea where the peaks come from? To reduce the contrast and brightness about 10% of an image each channel of the image is multipicated with 0.9. But this gives a lot of point numbers witch must be rounded to the 8bit per channel and this causes the peaks.
Now I think it's clear that a lower contrast cause a loss of color information.
Ok, I see already the next question: why does Celstia reduce the contrast and brightness?
First Celestia does not simple display textures Celestia render 3D objects with textures. That's a big difference! In order to show all this nice effects like reflections and bumpmapping it's necessary to reduce the contrast and brightness of all textures a bit. Otherwise it would be not possible to show these effects on a white texture!
I hope I could bring a bit light in this topic?
Jens
