AI Video Tools

[CDXbow]

A couple of days ago I started trying out some of the Video AI's that offer generation and other capabilities. I have tried about a dozen and 3 or 4 of them have reasonable plans. The vids are paleoart themed. You can see them here:
https://teamxbow.web.fc2.com/site/ai-ex ... index.html

I'll walk through some of the problems with them. The first 4 are videos of still graphics that have been animated by Pixverse ai. The first is Peter Truslers megafauna picture that was the basis of a series of stamps many years ago. Peter was tickled pink when I sent it to him. The ai missed animating Megalania (the lizard), I think it failed to recognise it as there is a log which looks similar nearby. The roos that should hop don't. The second is from an picture I did of short faced Kangaroos, these guys were thought to be walkers not hoppers. The ai gave them a dino like gate that looks good until you notice at the 4s mark one of the legs disappears on the far roo. The next is of thylacoleo defending its kill. It worked great until the thylacines attack, a bit of meat appears then one one of the thylacines just disappears. The last Pixverse one is an Ordovician diorama of Ostracoderms and this one seem perfect, no limbs and simple propulsion make fish scene good candidates.

The 5th video was done by the Crea ai after I gave it the text prompt - 'a herd of triceratops walking through the desert' t looks pretty good till you count the horns. Some of these are quadraceratops! The toe number is wrong and in the last frame you get a glimpse of a bifid tail. When the leading dino gets close you can also see what looks like teeth and part of a mouth underneath the jaw. What is it with these LLM's and altering the flesh? Have they been trained on too many David Cronenburg movies?.

It's been an interesting experiment and I will probably subscribe to one of the services. They increase the range and scope of an artists work immensely, as well enabling greater efficiency and volume. I think it's the same with coding and I hope the clever coders in these forums embrace AI fully. I also hope Fred and the development team will find clever ways to incorporate AI into PB because if they don't, I don't think PB has much of a future. That feeling is not specific to PB. Of course it could continue as an even greater niche product, PHB. Pure Human Basic.

[idle]

It's pretty good despite the glitches, given another 5 years we probably won't be able to tell the difference.

[CDXbow]

It's pretty good despite the glitches, given another 5 years we probably won't be able to tell the difference.

Yes it is, I was very impressed especially with it's results of Peters painting. In that example it's had to move multiple animals, with different position, direction and gaits. Sure it missed Megalainia and got the Roos wrong but it's impressive that it can get so much right. I still find it bizarre that the LLM doing graphics have problems with numbers of limb and digits. There isn't some directive (Four shalt be the number you count, not 3, etc) that terrestrial vertebrates have four limbs and and 5 digits (or derivatives of 5) and critters don't loose them in the casual fashion that the ai's dispose of them. I've done a bit of stuff with the image generators over the year and to me that it their main unresolved problem. The Pixverse site has a dancing kangaroo vid on it's home page, and in the still image one of the roos has a third arm. For some reason it's a difficult problem to solve.

[BarryG]

given another 5 years we probably won't be able to tell the difference

I'd say it's already started. I've seen videos on TikTok that I didn't realise were AI-generated at first. Totally fooled me.

[idle]

given another 5 years we probably won't be able to tell the difference

I'd say it's already started. I've seen videos on TikTok that I didn't realise were AI-generated at first. Totally fooled me.

just as well I prefer to read.
Though I'm looking forward for the movies that you can become the protagonist, Top gun Idle!

[HeX0R]

Maveridle

[Bitblazer]

Trilobites on ocean and dinosaurs both done with the kling pro 1.6 AI video generator.

Veo2 did a great job too, but for this task, kling pro did well (IMHO) and it allows slightly longer videos (in one session).

If you want to play with creating ai videos too, i used and recommend freepik which currently has 7 AI video models and plenty AI image tools.

[CDXbow]

Trilobites on ocean and dinosaurs both done with the kling pro 1.6 AI video generator.

Veo2 did a great job too, but for this task, kling pro did well (IMHO) and it allows slightly longer videos (in one session).

If you want to play with creating ai videos too, i used and recommend freepik which currently has 7 AI video models and plenty AI image tools.

Those are not very real trilobites and excessively horned triceratops

[Bitblazer]

Those are not very real trilobites and excessively horned triceratops

That is probably true, i have no clue about either of the animals. In my experience, google veo2 produces correct videos while kling pro is "just" good to create pretty pictures.

ps: I hope kids in the near future will learn about these animals by also being able to look at (ai) videos and in those "living book" products, the movies will hopefully be correct (to the current knowledge of the creation time).

Who is up for a new "children school book" project and willing and able to create the required videos?

[Nituvious]

Did they really use a chat bot to design a text editor and then complain about it? Lol

[CDXbow]

I've put together an animation of some of Peter Trusler's artworks on youtube. All these are still images of his animated by AI. If you like Dino's and birds you'll enjoy it, go and give it a few likes. It's very much unfinished, mainly because I've run into a problem with the video editor, I've marked a bit to play and now I can't unmark it. Yes, I am that ignorant, I confess I haven't done video editing for 30 years.

https://www.youtube.com/watch?v=nP3bFIo ... nel=CDXbow

If you would like to see more of Peter's works and how he approaches paleoart you can watch this:
https://www.youtube.com/watch?v=kUWw6OO ... Steven2156

[idle]

it still does my head in how it can animate from images with very little context.

[CDXbow]

it still does my head in how it can animate from images with very little context.

Yes, quite extraordinary. Pretty close to magic. The ai has to build a model internally that includes lighting and some form of rigging to animate the model. Some times they fail at very simple things, one I have not been able to get right is a classic wagon wheel space station that rotates to provide pseudo gravity. I must have tried more than a dozen times with text prompts and image2text and all have failed, often miserably.

I've been using meshy.ai more to make, rig and animate 3d models from 2D pictures, though I have done some with just a text prompt, I have found seeding with an image is more predictable. Have a look at the link below to see some of the models I have made - you will notice a classic space station, straight from Blakes7.
https://www.meshy.ai/discover?search=cdxbow
Their free plan is pretty good with a few ways to get credits.
They have exporters to platforms like Blender, Max, Unity and Unreal.
A toolchain that saves hundreds of hour per model. Game development has changed forever,

[tj1010]

I'd actually wager a large sum of fiat currency that in one-hundred years there will still be hallucination in neural networks used in the large transformation models that do NLP>ViT and NLP>Code.

BTW I love that Google's Veo 3 is going to shake up the AI scene meanwhile time-limited photorealistic video generation has been around a while. AMD Amuse is 100% free and local has been doing it since April with the same error rate..

[CDXbow]

I'd actually wager a large sum of fiat currency that in one-hundred years there will still be hallucination in neural networks used in the large transformation models that do NLP>ViT and NLP>Code.

BTW I love that Google's Veo 3 is going to shake up the AI scene meanwhile time-limited photorealistic video generation has been around a while. AMD Amuse is 100% free and local has been doing it since April with the same error rate..

Thanks for the tip about AMD Amuse, I haven't tried it but I will.