Do you have a dumb question that you're kind of embarrassed to ask in the main thread? Is there something you're just not sure about?
This is your opportunity to ask questions. No question too simple or too silly.
Culture war topics are accepted, and proposals for a better intro post are appreciated.
Jump in the discussion.
No email address required.
Notes -
AI videos. Have you seen any good short films that were made with the use of GenAI tools, in entirety or in large part? I don't mean oneshot from prompt to full movie, it can be any kind of multistep process where the creator may generate character sheet images with one model, then use a video model to make videos, and any other steps with AI models can be involved. I can only find a handful, and they aren't great. Though YouTube's search is notoriously broken anyway.
The ones I could find are very flat storywise and are more like a techdemo, packed with action movie shots. I'd be interested in a more story and character-driven one or just anything where the filmmaker wants to tell a story. Less "I wanted to geek out with AI tools so I have to come up with a story for this movie now", more "I wanted to make a movie and now AI makes it possible and simple to realize my vision".
It seems that the story, the content is the harder part. Which isn't surprising, since having access to pen and paper / printing press / rich word processor software didn't suddenly turn everyone into a book author either. The bottleneck is having something to say.
Or perhaps the human acting performance is just too subpar for creative people to accept it as of yet. I mean that the people / characters in a generated AI video often seem to express emotions in an uncanny way etc.
Or perhaps all creatives who would have story ideas and execution capability don't use it because A) they are strongly anti-AI of the bluesky sort, or B) they anyway have access to friends who can act for their short films and it's more fun to do it with other people for such non-nerd creative types. or perhaps C) the latest generation of AI tools need more time to penetrate the creative spaces because they are still mostly present in tech-geek spaces only, i.e. creative types don't yet know about how good the latest models are and have dismissed them months ago when they were worse. or maybe D) creatives have very low tolerance of deviating from their vision, and current models are too random and too hard to control for them to be a good vessel to carry their vision and ideas.
The stories themselves are pretty easy. There's a lot of people who already had 'stories' in their heads that they wanted to bring to the screen, to the point where script-fic is a little derogatory in fanfic spaces. And we're at the point where you can give an LLM a pretty rough idea and get a coherent story out from them (cw: 30k words AI-generated story, painfully full of obvious spaces for improvement and still about on par with recent Disney output).
The tech's just hard. There's people messing with it, and for short periods you can get human-like emotion and acting, especially if you're open to pretty non-standard definitions of human (cw: sound, hellhound with human teeth). And then the background fades into a dream, or the character suddenly looks subtly different, or things start clipping, or the lipflaps don't match up with the actions.
Some of that's because people are just dabbling hobbyists: if there are people building their own unique voice clones, converting renders or live video into to pose2video work, or doing a lot of layering, there's definitely no one publicly doing all of those things combined. There's probably a way to make it work, by exploiting first-frame-last-frame and controlnet. But that requires a pretty sizable set of storyboards, and a stronger vision, and the ability to maintain focus when you're presented with something close but not quite good enough..
It's also just hard to do at home. Wan2 or LTX kinda work on an nVidia 3090 or better, but it's minutes-per-generation in the simplest cases. Add on that overhead and it can take the better part of an hour to make a single ten-second clip. With the right workflow or tooling you might be able to make that work out - that's still about a week per hour of video - but right now those workflows are things you'd have to build yourself.
I'd expect we'll eventually get some sketch-focused artist or rough animation-focused modeler working directly with them, and it'll have some major benefits, but I couldn't tell if you if that'll be a weeks, months, or years.
Higgsfield is making some series with Seedance 2.0 and Nano Banana:
https://youtube.com/watch?v=LQ-vSa9_H98 https://youtube.com/watch?v=digHr6k38x0
I just can't stand watching this genre (superhero action movie or whatever), more than 30 seconds at a time, but it looks fine on a technical level. It would be great to have something with a bit more interesting story.
Wow, these are really good.
For the longest time ai video generation was really bad with style consistency and character consistency but I didn't notice any problems with these at all. My background is in design/visual art and I think they look really good visually. At first I didn't like the constant cuts which I took as downstream from ai generation but it does seem consistent with the action genre so maybe it was a choice and not a technical issue.
The worst part is that the dialogue and text seems really obviously generated. The lyrics in the song at the end of your second link is fully just copied from another real song ("Ooh I love you so, but why I love you, I never know") but it's way worse in the ai version.
How difficult is it for you to tell apart the human generated ones from AI, given your background?
Sorry I’m not sure if you’re asking about video or text/scripts, I’m guessing you mean video, and I assume “human generated” means like traditional CGI and/or actors being filmed (I’ll call this combination “Hollywood” going forward) so I’ll answer from that angle
I would say for these two Highsfield videos specifically they are really really close to what Hollywood would look like. Especially taken out of context the ability of AI to mimic the CGI portion is extremely good and I think I’d have to be more familiar with CGI itself to identify the tells. On the other hand there are subtle uncanny valley-esque tells that do give away that it’s AI, like the Asian male character in the first video’s face appears much younger in a handful of scenes than in others which is strange. The women in the first video are much more beautiful than most actresses which also makes it seem fake. The monsters and robots and environments/action sequences in both videos are so good that I don’t think I would know they were AI vs trad CGI. Voices were sometimes inconsistent from the same character and more importantly lacked the distinctiveness of real voices.
I could go on, I think the conclusion I’d draw is much the same as the conclusion that I’d draw from ai chat bots or ai still image generation which is that all of these tools are very good at getting you to like 90% of a perfect product but it’s the last 10 crucial percent that it’s almost never able to achieve. In certain domains you can manually bridge the gap with a bit of human intervention but it varies from task to task.
Not as relevant but I also noticed that there is an issue with AI generation in general VS human work: when humans create something, the person making it nearly always does the best they can, or in the least something that they wouldn’t be embarrassed to produce given their level of skill or experience. And we are usually not overly critical of other people’s work, when they do a task for you we tend to accept it as their work. But with AI generation the latter problem is an issue where the person checking the AI generation is willing to approve work that is of a lower quality than what any human would produce because it took nearly zero time to produce and it wasn’t their own work so you might as well let it go through. Besides that correcting the final 10% to get it to perfect could take a massive amount of time, so in the end you might produce 100 times as much content but it’s all 90% as good so you think you’ll get 90 winners but the Pareto principle reveals that the final 10% is crucial to anything being good enough to succeed so you’ll run into rapidly collapsing returns
Disagreed on the 90% figure. The AI can spit forth from its datasets superficially accurate renditions of tropes and imagery from other works that it has seen. What it cannot do is perfectly mold every detail towards the purpose of a given scene. The act of expression is missing, and thus 100% of the relevant criteria for an aesthetic-artistic work is missing. All you have in the end are a variety of statistically likely tropes and images strung together in some pattern of Baudrillardian remove, lacking any signification or meaning.
As a designer and a creator and aesthetic person I would argue that "the act of expression" is nowhere near 100% of the relevant criteria for an aesthetic-artistic work. This is a really postmodern notion, if you're someone who like weeps at Cy Twombly or whatever then sure I guess that's what you think. But the technical detail of say, JC Leyendecker or Maxfield Parrish to me is easily 90% of the value of their work. I actually agree with you on the rest of what you're saying but I think you are devaluing broadly the value and worth of the technicality of craft or whatever which the AI can do in, on average, 90% of the same quality that a talented human can do. The 10 percent of expression in the end is crucial as I tried to lay out but it's not 100 percent of what matters.
Craft would be about strategically wielding the narrative elements in order to achieve affect or to convey meaning. In the second video in the post to which we are replying, when the girlfriend character gets stabbed by the demon, does it register as anything other than bathetic? Or does it even register at all?
Those artists you mention, they achieve affect through the careful arrangement of skillfully created elements. With the AI, there is no cumulative impact, and it doesn’t use skill to recreate things. All we have is a clunky copy-paste machine.
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link