Labrador Retriever Mix Watersporting in Two Ways at the Same Time (SFW, Videos)
Labrador Retriever Mix: "Is this radical enough for you, dawg?"
Look at her go, watersporting in more ways than one. She's surfing, drinking from the tube, all at the same time. Yes, she thinks drinking is a type of sport relating to water.
Decided to give Seedance 2 a second chance after calling it off for a month. I had accumulated lots of credits for an app by letting ads run and scrolling through ads on the .com version of Civitai. However, the problems crept up again. The app that I use (which works really well for images and image editing) was causing errors from timeout. I had no successful generations after at least 6 attempts (that's how bitchy it is). Like last time, it did not give back refunds from errors caused by timeouts (again, like the bitch it is), which is very stupid. It did however generate Grok Imagine Videos successfully (so the API is less likely the problem), but the results were bad and not worthy for show (almost guaranteed that it won't be approved on e6ai).
I had to use blue buzz to make this work, and oh boy it was pricey:
-First attempt (second video of the post): I was pretty excited when I first watched it, the flip, the sun effects, they were good, until a random straw appeared and she began to drink from it, which does not make sense.
-Second attempt (third video of the post): The flip was a bit too crazy that the quality degraded from it... and did she just turned into an anthro?
-Third attempt (fourth video of the post): I adjusted the prompt to see if it knows how to get her to use the tube attached from the headwear. This is where the tube issue got bigger. There is an additional long-ass tube waving as she surfs, which is not good as it is clearly visible. This one had (in my opinion) the best animation in terms of the jump, tricks, and the transitioning with the camera change. Of course the tube-drinking is still an issue.
-Fourth attempt (fifth video of the post): Adjusted the prompt to make sure that it understands that the tube is a drinking tube. Long tube issue happened again, boring jump, no tricks, it did got the drinking part right though.
-Fifth attempt (sixth video of the post): The jump and the trick is cool, but the shark tooth pendant changed significantly, and a random can of soda with a straw appeared on her paw (which is fine), her satisfied "Mmmmm" ended up with a weird sound.
-Sixth attempt (first video of the post): Changed the prompt again to have her have a satisfied sigh. Big jump, not really a trick, has fucky physics, which does bother me, but let's just say that it must've been the wind or something, but has (more) proper drinking from the tube. This one has the least major errors, I think. My new banner image came from this video.
-Seventh attempt (seventh video of the post): Lots of noticeable errors, like that bloody long tube, some random thing appeared on her paw as she jumped, did a small neat trick with the surfboard, not going to comment on the drinking.
Overall, I'm still disappointed with the results. I expected better given that Seedance 2.0 is still the top of the image to video leaderboard (with audio) according to Artificial Analysis. My prompting can't be that bad as I did adjust for almost each iteration. Maybe it's the drinking helmet can holder hat that the model does not understand, I mean, they're not that used, which could be why the model (and probably the rest of them) struggle with this concept. Wasted thousands of credits from the app and 14,000 blue buzz to get these. Pretty anxious that one of them that I post might not be approved on e6ai. Likely won't touch video generation models again until they get more better than this and that I can run them locally. I might need to get a M5 MacBook with more ram because my M3 MacBook Air with 16GB of ram can't do shit for video generation. Or...
Maybe you can suggest me another app that also uses an ad credit model that properly refund credits back for failed generations. Not willing to pay with real money.