Burnout, as described in #34, can sometimes happen when there are particularly many feedback images or for certain kinds of feedback images. It's not quite clear to me what exactly causes these results, but I've found a way to mostly prevent this from happening.
Burnout protection can be enabled by selecting the corresponding checkbox in the UI. It's recommended to do this when using a large number of feedback images or when FABRIC produces low-quality results.
How it works:
The feedback is dynamically reduced (by 0.5 in each step) if the difference between conditional/unconditional forward pass is too large. If the difference is small enough, burnout is unlikely and we increase the feedback weight again (by a factor of 1.5 in each step).
Additionally, to prevent the generated image to drift too far out-of-distribution, we center the output of the model by subtracting the channel-wise mean. This prevents the generated image from getting too dark/bright.
Burnout, as described in #34, can sometimes happen when there are particularly many feedback images or for certain kinds of feedback images. It's not quite clear to me what exactly causes these results, but I've found a way to mostly prevent this from happening.
Burnout protection can be enabled by selecting the corresponding checkbox in the UI. It's recommended to do this when using a large number of feedback images or when FABRIC produces low-quality results.
How it works: