The method outlined above currently suffers from several problems. There are
color quantization effects, violations of the assumption that 
is
uniform on the block, and problems with shadow calculations.
The color quantization effects arise from the fact that the images used to store pixel values are only able to discriminate between 256 different values. This introduces severe quantization problems, particularly in dark or light images. One solution is to use images with higher intensity resolution, however such images carry a significant memory and hence performance cost.
The assumption that is constant works well provided a
sufficiently large region of the block falls under the same lighting
conditions. This is not always the case, and small regions reduce the
validity of the assumption. For instance, a striped face will satisfy the
assumption if many stripes are visible. But if only one or two stripes are
visible, the assumption will be invalid. One way to handle this problem
is to improve the ambient intensity recovery process to take into account
a greater number of faces, and to incorporate weights into the solution to
bias it toward large regions.
Finally, the current implementation uses a shadow buffer for determining occlusions. This method is prone to aliasing problems. The alternate approach, a ray casting method, is too slow. Hence a hybrid approach seems appropriate, where the shadow buffer provides a certainty measure, and uncertain values are confirmed using a ray casting approach. Regardless, the shadow calculations are subject to errors in the modeling process. There is no easy way to artifacts from such modeling errors without making further assumptions about local regions of the face texture.