Color images may be stored in any format on disk as the color space is up to the whatever is writing the file. However, the more general use case or encountered case is that color images are saved in a non-linear color space as it allows for tricks such as allocating more dynamic range to certain parts of the image. With that said, the general texture samplers will assume that the data its sampling is linear( as is ). This means that if you loaded one of the textures with non-linear data, your texture sample will also be non-linear. You will have to tell the sampler that the source data is non-linear so that when the texture sampling unit fetch a value from the texture, it will apply the inverse of the non-linear function to convert the value to linear. I haven't touched DirectX in a while, but the premise is somewhat the same as OpenGL. You'll have to flag/tell the texture/texture sampler unit that the associated texture data is non-linear and needs to be converted.

There are several links at the bottom of that page. I read the top one and it was a great introduction as well.

What they say in these tutorials/articles is that in an image editing program or 3D modelling software the artist would be doing gamma correction to make it look physically realistic on a screen, resulting in the texture/model/image that comes into your application is sRBG. But the problem is that the algorithms used for different effects operate in linear space. So if you apply them to the data directly without linearizing the data first, it will look quite bad. You can several examples of this in the top link.

So as cgrant said, you should instruct your graphics library that the data coming in is in sRGB space (or do it manually in a shader) so that it can translate it to linear space. Then at the end of your pipeline, in a post-process fragment shader for example, you again do the gamma correction to make it look better. If you didn't take this into account when importing, you would have performed gamma correction twice.