Okay, what if I render the image twice, once with regular sRGB and once converting to linear before blending and back to sRGB after blending, then use the maximum value of the two images so that the sRGB one can smooth out the darker parts? Eh, if I have a big alpha gradient blending between exactly black and white, which have the same binary representation in both versions, the linear version will rise fast enough that the sRGB version won't catch up before the stair steps are noticeable. In fact, they'll both cross the threshold from one value to the next in the same places, just the linear version will have bigger steps in the darker areas. BUT at least I'd be able to preserve flat areas of dark color that lose their precision in the conversion to linear but could be maintained in the sRGB version. All this might be moot though if I want to have more advanced light/color management though, like combining multiple light sources with an underlying albedo color (plus specular!). The browser's built-in art tools just won't work for this use-case anyway. Actually... I could combine all the diffuse light layers with additive mixing, then a multiplicative albedo layer on top of all that, and an additive specular layer on top of that. The only 3 blending modes I need are add, multiply, and occlude. And masking. Blue noise dithering helps a lot! But if I crank up my screen brightness I can still clearly see the noise.