Are we wasting tokens with Sora like video models? (Luma Dream Machine & Runway Gen 3)

Are we wasting tokens?

Something I've noticed with both @LumaLabsAI and @Runwayml is that they can both give amazing results but at other times they can verge on completely unusable, even when using the same prompt. What I've noticed is my output quality varies on the time of day and day of the week. Why might this be?

Looking back at the video Two Minute Papers did on Sora we see that quality varies drastically depending on compute. What I am assuming is happening in the case with Luma/Runway is that during times of heavy traffic you are loaded into queue on a less powerful server which would equate to reduced quality to the point of getting jumbled messes at times, effectively wasting your precious tokens.

An interesting tidbit is I noticed that img2video with a prompt often gets stuck processing, sometimes for as long as 6-18 hours. What actually might be happening here is that it's forcing your generation to be processed on a more powerful server but the reason it takes so long is likely the fact you are all the way at the back of the queue. From what I know there is no way to pick which server you are processed on but there might be a few hidden rules that dictate it like img2vid, current traffic and complexity of prompt.