Due to several reasons its possible that the generation of a thumbnail fails:
The image format is unknown and cannot be loaded by the program
The image format is known but the file is somehow broken and thus cannot be read.
The generation of a thumbnail would take too long, due to the large size of the image.
If the generation of a thumbnail fails due to some of those reasons (or others) then this information should be preserved for the future to avoid trying it again and again. The problem is that the above mentioned issues are often program specific. E.g Nautilus can't read the native Gimp format xcf but of course Gimp can read it and creates thumbnails from them. Or one program uses a broken TIFF implementation which refuses to load an TIFF image but another one uses a correct implementation.
Because of this, its best to save these failures premanently per program. In the Directory Structure section there is a 'fail' directory mentioned, which should be used for this. Every program must create a directory of it's own there with the name of the program appended by the version number (eg. .thumbnails/fail/nautilus-1.0).
For every thumbnail generation failure the program creates an empty PNG file which uses the nameing scheme described in 'Filenaming Scheme' and stores it in the application specific directory. If it's possible to obtain some additional information from the image (see 'Store Additional Information') they should be stored together with the thumbnail too, at least the 'OriginalMTime' key must be set. If this file is stored in the global thumbnail fail dir the directory structure must recreated too.
This method has the advantage that a program can access information about a failure the same way as successfully generated thumbnails.