Compilation - a litany of fails

More about Compilation - a litany of fails

This compilation I believe includes all the failed clips I've produced, with no doubt a few repeats, plus some of the successes and partial successes mixed in cuz they were part of compilations I'd already made & thus easier to include. I figure someone might find it enjoyable. There are a fair number of clips I never published. <shrug> It helps pass the time while I wait for enough energy (both compute and personal) to try again.

In the past 2 weeks, I've generated by my count 32 5 second Provideo clips, 7 10 second Provideo clips, 20 images, and 5 times taken a keyframe and converted it to a text prompt in an attempt to guide my efforts. A total of over 6,000 energy expended, and I lost count how many sanity points. Two of those clips were for other images not related to this project. Of the remaining 37 clips, there are 4x5 sec & 2x10 sec I'd call successful and worth splicing together into a coherent story (see "Tramma in her school uniform" in my dreams), and 2 more 5 second clips which add a least a little to the story so I included them, for a total of nominally 40 seconds.

The other 29 clips totaling right around 160 seconds of video are for the most part useless to me, if interesting in their own right sometimes. So I'm getting a yield of approximately 1 usable second for every 5 I generate. That ratio is slightly more favorable if I download the clips and do some judicious editing and adding transitions in a local video editor, but that kind of defeats the purpose of sharing the results with the DDG community even though it works for my personal use.

Even with that, though, I doubt the yield goes even as high as 25%. There are just too many glaring continuity errors like doors opening the wrong direction, or wildly different furniture or item placement or the elements of clothing. I find that the more prompting you put in to try to control everything, the more likely the AI seems to be to go berserk and ignore what you're telling it to do, probably because of the way it tokenizes and parses prompts I suppose.

I don't claim to understand how the model works, that would be hard since it is for all intents undocumented. I just know it is maddening to work with if you're trying for coherence and consistency to build an extended video. I've not managed to create what I consider a usable clip to continue the story in 5 days now. Depending on the whims of the random results you get when you feed the same prompt repeatedly to the dratted AI, that might change eventually. I believe I've nailed down a starting image and prompt which SHOULD get what I want for the next 10 seconds IF the AI pays attention for a change and finally produces the results I've seen in previous iterations again.