I'm working on a 2D rendering engine that uses textured quads. Basically, I create a texture, load the image from a file with ARGB data (such as .png), and then draw 2 triangles using the texture coordinates to get the image on the screen. This works great when I render the image pixel-for-pixel. The problem comes when I scale the image.

I'm using linear scaling (device->SetSamplerState(0, D3DSAMP_MAGFILTER, D3DTEXF_LINEAR)), and what appears to be happening is the the fully transparent pixels (alpha 0) are having their actual color values blended with neighboring semi-transparent pixels. As an example, I had a red letter 'T' in a 32x32 image. The T was solid (alpha 1.0) and a single outline of pixels around it were the same red, but alpha 0.8. The transparent pixels were alpha 0 and color black. What happens is, when I render on a full white background (device->Clear() using white), I get a black halo around the edges of the T. If I change the fully transparent pixels to be alpha 0 color white, it looks as expected, but if I then render on a fully black background, there is a white halo around the image.

Now, this seems understandable as it is doing a simple interpolation of the data. However, I noticed that I do not have this problem when scaling with Photoshop or when scaling an image in a C# form using Graphics.DrawImage with bilinear scaling.

Is there any way to make DirectX perform this smarter scaling? I've tried every combiniation I could think of in device->SetRenderState() with no luck.