LLM Avalanche: Mihail Eric: Storia

LLM Avalanche: Mihail Eric: Storia

Recording: LLM Avalanche: Mihail Eric: Storia

thank you I am mihail I'm one of the co-founders of Storia so we are an early stage company building uh using genome AI to power and supercharge pre-production for film and television projects I'm gonna do a quick few slides of overview I'm going to give a quick demo and then we'll go from there so I have to do a little bit of advertising because my co-founder will be angry at me here's who we are and a bunch of logos we're technical people uh solving a non-technical problem we also have some really cool investors and uh advisors that are helping us out here I'll get right to the demo okay so as I said we are building a creative assistant for storytelling focusing on pre-production for those people that aren't familiar with pre-production it is the phase in most film and television projects where a director or screenwriter a number of professional creatives will go through the steps of actually planning out a film or television episode or fully linked feature film this includes making casting decisions uh planning out where something is going to be shot where uh like what the Wardrobe is going to look like so a number of different brainstorming decisions all of which happen before you actually go on set and start making footage so what we've built and this is the first wedge into the broader Vision that we have for AI powered pre-production is a storyboarding tool and storyboarding for those that aren't familiar is the centerpiece of pre-production it's basically a document used by professional creatives where where you take a script or screenplay you break it up into units you then feed it into some sort of system and you illustrate every single thing to see what the shot will look like again what you get at the end is a document that is an illustrated version of a scene or a screenplay I'm gonna show what this looks like because it'll probably get more tangible our flow is pretty simple as a creative on our platform you can come on and actually upload your own script or pull in from existing scripts that you might have here I'm gonna pull from one of the ones we've already pre-loaded this year these are all famous movies that some of you may have heard of for the purposes of the demo I will go with top gun Maverick we allow people to pick different styles they can use to actually visualize their script so these are famous directors or famous films for this demo I'll just use Blade Runner because it has a very distinct visual Style but we also enable you to provide your own Styles and uh really support freeform text input at this point we can leverage existing text to image models as well as provides custom models to do some of this image generation I'll use one of the ones that are off the shelf just so we can see what it would look like so at this point what's happening is we're actually parsing the script that has been uploaded which can be well over 100 Pages for the top gun Maverick movie we are now generating breaking this up at the units of scenes and then we're generating an image for every single one of those scenes in different uh in the style of the Blade Runner film and so in a matter of seconds I'm only going to show the first 10 scenes here but in a matter of seconds what we've seen is on the right hand side we have the visualizable units of the scenes so these are all you know different actual scenes from the film in the middle we have the visualized depictions of those scenes so what we've done is we've used LMS to actually interpret the semantics of the scenes and then visualize them in the style of top 10 Maverick you know this is done for many many different scenes and this was done in a few seconds but we at scale we could do the entirety of a script in a matter of minutes now while acknowledges the fact that a scene may not be perfect you know what an AI generate system will generate will not be perfect so we also enable you to go and actually fix some of the shots and this can be done either through hard you know actually prompt engineering or modifying Styles modifying some other parameters like camera angle lighting inspiration images which are just different ways of interacting with your shots to actually get something that's closer to what you have in your in your minds and the real use case here is professional creatives that are trying to really understand and explore what their Creative Vision could look like before they go on on the actual set and start creating it so that's the gist of it and I will leave you all with uh some technical challenges because this is a technical group so I feel like there's a this should resonate with some of you we are we have a number of different things that we're trying to solve to actually make this work at scale and for those who don't know this is a an unreleased Jay-Z album um but but uh an example of some of the things that are difficult about this problem are everything from consistency whether that's character environment stylistic really persisting bad across the entirety of a script character ambiguity and understanding what exactly characters should look like and how you persist that information across time uh carrying over a lot of context for a scene and then of course one of our big goals as a company is to really support a long tail of artistic and cinematic Styles so finding ways to really scale out to the entire Corpus of films and television shows out there is something that is still very much a difficult challenge where do you get the data how do you extrapolate from very few data points so yeah just give you a little bit of teaser uh we're hiring right now looking for strong mle folks and and broadly engineering so if you're interested in any of this come talk to me afterwards thank you