3 ms·
This approach still doesn't fix what I feel is the worst part of "portrait" mode images, which is that the transition to the out of focus areas looks incredibly
by starky 5y ago
This approach still doesn't fix what I feel is the worst part of "portrait" mode images, which is that the transition to the out of focus areas looks incredibly unrealistic. If your background is so far out of focus, then the entire subject shouldn't be sharp, areas outside of the focus plane need to be slightly out of focus. There is also a strong uncanny valley effect because the Bokeh doesn't have any of the character that comes from a real lens.
Personally, I think we are quite a long time away from being able to computationally blur photo backgrounds acceptably well.
- notorious_IPD 5y ago[article author here] The issues you're seeing here comes down to quantization errors in the depth maps. The effective range of a depth map for most human-scale scenes is about 20m; what you then have is 8 bits (i.e. 256 levels) to try to encode everything from right up close to the camera out to 20m. Even if there's in interpretation function on top of this to distribute the depth more meaningfully (e.g. storing 1/depth to prioritise detail on closer objects) you still, in the end, have banding. When we developed this approach we validated it against existing techniques and I'd still like to think we have the smoothest, least prone-to-banding technique for simulated defocus out there; the maths was tweaked to specifically address this. Of course, there are limits - aliasing can only be mitigated so much. But that's why we're still working on it and this kind of feedback is enormously useful (thank-you!) in helping us move closer to a virtual camera that feels more like a real one!