@loftwah@jtregunna I faced similar issues processing petabytes of data with a legacy algorithm that involves cluster resizing. Since allocation is region-based, clusters can fail to provision when capacity is unavailable in that region, disrupting daily workflows. A solid infra-level issue (GCP)
These tokens are globally connected,therefore, when a prompt is provided, modifying one token can influence other dependent tokens. As a result, changes may appear in unintended regions of the image rather than being confined to the targeted area.
Ever wondered why AI image models can’t truly “edit just one thing”?
In order to answer this question, let us begin with the basics of how Vision Transformers work. Vision Transformers are an image-based extension of the model proposed in the paper “Attention Is All You Need.” They divide an image into fixed-size patches, which are then tokenised.
@linkedlist64@sauravk87 Similar approach but incrementing it by 1..3..6 is better .. there can be edge cases where val at index 121 can be 1 .. the approach you proposed just keeps incrementing.
@sauravk87 Limitation to this might be .. if I reverse traverse from the index that has value 1 … index-1…index-3 and so on and so forth .. I might end up finding another index with value 1 .. but approach might work if this limitation is addressed.
@sauravk87 More than the problem itself, I wondered why only 145 queries. sum of n natural numbers 1…145 >10k. Incrementing indices like 1, 3, 6… eventually we get an index with value 1. Then we know the solution lies between last and current index and search backward. Math is great 😌