TIFF Recap: Filmmakers Hear “AI” and Think “Slop”

0:00I did my usual 1970s New York City street test with both GPT2 and 2.5. Yeah, let's see. That kind of looks like at first glance there's some gibberish text. To me, 2.5 just felt more richer with detail. This feels like it used more world knowledge and web search to build this. Exactly.

0:22All right, welcome back to Denoised. Addy, how's it going? Joey, what's up? I'm on the road today. You're traveling. I'm back from traveling, back from TIFF. It was good. Yeah, I got a couple of things to do. A couple observations we can want to talk about. I loved your life hack. Like stand on the other side and then mirror image the TIFF logo. That was great. Yeah. And I got some good comments from the like, just use Nada Banana. And it's like, this is old school, crappy Photoshop editing. Not even Photoshop. Just mirror the image. Follow Joey for more life hacks. Yeah. It's crazy. Like as TIFF, they'll shut down one of the main streets in Toronto and it sort of turns into the street festival.

1:01and there's like these vendors that'll give out like kind of the most ridiculous freebies, but the line for them are insane. Like people will just wait in line. These lines are massive for like a free cup of ramen or like a free little piece of chocolate. And people just wait forever for these lines and lines to take a picture in front of the Tiff sign. I don't know why that reminded me of the Costco food sample thing, but yeah. Same. I mean, people wait for free stuff. So yeah, Tiff was great. and a few AI observations. The biggest kind of presence there, I would say, was Kling. They did a panel that was officially part of the market on the first day on Thursday and they were teasing they got some new models coming out.

1:43They couldn't announce them there. So they said in a few weeks there should be some new stuff. There I am in the front texting. There you go. The panel was interesting. I would say a lot of stuff that we talk about. Script and story. Yeah. Previs to final pixel. Editing. iterating on idea. Yeah, I think, oh, that's men from Machine Cinema. Yes. Yeah, I mean, it was around a lot of stuff. And yeah, Machine Cinema had an event too that was also a good party, networking event. Amazing. But yeah, Kling was, Kling was kind of the most present. But I would say overall, not a ton of AI stuff at the festival in general.

2:23And definitely when I would talk to non-AI people, just normal filmmakers and producers, and would explain the AI stuff we're working on and would mention hybrid. The idea of hybrid and hybrid production workflows completely over their head. As soon as I mentioned AI, the assumption goes to... Full synthetic. Microtrauma, AI swamp, and negative association. It's like the equivalent of getting in a Waymo. It's like the minute you go AI, you give up all control. You're just in the backseat. Like that is the metaphor in their head. But it's not that. I think it's more like getting in the front seat, getting in the driver's seat of a Tesla where you have, you can still drive it, but then you can let the car drive when you're tired, if you will.

3:13If your Waymo was a Mario Kart and you don't have control and you're also having a really crappy ride. Sure, we can go with that analogy. Yeah, when you mentioned AI, the assumption is slop. And then once they start explaining more that we're working with real people, we're expanding with workflows, like it still involves production and people and just more of the environment is AI generated or assisted. It makes a little bit more sense, but it's still also hard for them for that. It's still very much not mainstream and it's still very hard for it to click until you start showing examples and showing things.

3:49I think even more so than showing Joey, I think they have to generate it for themselves. It's like you can show them the door, but they have to walk through it kind of a thing. The one sort of... Okay, Morpheus. Okay, AI Morpheus. I know, I'm feeling real... Is it Blue Pill? Real Blue Pill today. So the other thing that I hear a lot from traditional filmmakers is the argument that AI is trained on old stuff, and so it can't make new stuff because it's just able to regurgitate that old stuff in different ways. to that I say you can creatively drive the models to do what you want you just have to push it and pull it into a direction that's completely new and it'll go there but it won't go there on its own and I don't think that was ever the intent no and I think you definitely can get new stuff and I think I have an interview coming out that was finally cleared to post that will elaborate on that from a high-end filmmaker clarifying why he likes AI for creating new visuals in a new language.

4:55Which, yeah, I think, at first I did think of that, but there's so much to explore in the latent space that is new visual languages. Only on this podcast are we talking about exploring the latent space. I love it. Also, to saying that it's a remix of everything, people should watch the really good documentary Everything is a Remix that, what's his name, Kirby Kirby Ferguson. He did this series like years ago and it was basically just like a showing like shots and shots and shots from films that of like you see it now and it's like oh it's based off of like a shot from a 1950s film and that film is based off a shot from a 1930s film and just sort of like how everything we see today or this was 10 years ago is like a build upon everything that came before it.

5:47to sort of this same notion, I was watching a YouTube video yesterday where they're talking about The Matrix, the movie, stole everything, and yet is one of the most original movies ever made. Like everything, all the elements in there, like the fight sequences were Hong Kong style from the 90s. They even got the action choreographer from Hong Kong to do the stunts. And then like Bullet Time was like stolen from something else. And like all of these elements just kind of came together, but then it created something completely unique. Other thing that was happening at TIFF, so I had to leave before this panel happened, but Autodesk was also running a panel and they were more focused on Flow Studio.

6:33So Flow Studio is their AI-centric web tool. Yeah, it's what they built on top of Wonder Dynamics. Right, they acquired Wonder Dynamics, they rolled that into Flow. But then it's been built out. It feels a bit like intangible, like a web-based 3D builder, like a very simplified 3D builder that you can use to block out shots and then run that through a generative pass. That seems to be the trend now is like going from 2D to 3D, doing your camera moves and your character animation or your what have you, your motion, if you will, in 3D and then bringing it back to 2D for generation. Yeah, a lot of interesting stuff with what they've been doing with Wonder Dynamics.

7:16And also some of the original models that they have been training are for training character movement and character animation. So like if you had a animal or dog or person that had a specific kind of walk or movement, you could record some data on your own, use their model to help train it. And then you have like if you're just like, oh, I need this character walk from here to here. It has this training data of this very specific walk. and just little tools like that to help speed up the animation process that's very practical and useful and just sort of helps an animator while still having that level of control.

7:53I think Autodesk sees that their revenue is in sort of Maya licenses and things like that, right? Like their bread and butter is maybe on the animation side more so than the live action side. And so they're trying to capture or recapture that market with sort of the next generation of tooling, which is Flow Studio. It seems like they're very animation focused. The animation and I think also pre-vis. I mean, so I tried it with them. I wasn't around for the conference, but I was able to chat with one of them with Matt at the festival. And he said that they're doing, like they're known for animation.

8:27They're known for Maya. Right. All of the material that they sent me beforehand, absolutely nothing mentioned Maya. It's all Flow Studio. Yeah. And he was like, you know, we're much known in post, but we're looking to get into production. And that's why we're going to like film festivals and stuff like this to educate and just get out more and be known more as a production tool. Yeah, that's a smart strategy. I think they're almost trying to capture the generation that never touched Maya. It's like the younger artists that are now on Blender that just quite frankly won't pay for Maya. Maya is an enterprise expensive tool.

9:02It's also complicated. They're like, yeah, it's a complicated tool. It's powerful, but complicated. How do we grab the Blender demographic and bring them on board to flow. Yeah, or not even the Blender demographic of just like filmmaker demographic that just wants some easy 3D visualization tool, web-based, and it's a deal with software, and you can just spin it up. You know, I mean, Wunder Dynamics was always really good with you can film people doing stuff, and then you can turn that into a... Performance capture, yeah. It's amazing, yeah. I wonder if Nikola is still there. He was one of the co-founders.

9:35I think he's still at Autodesk. I think so. Yeah. So, yeah, I thought that was interesting that they're pushing more into the production flows and sort of how they can bridge this traditional Maya tool set with this new AI-assisted web-based tool set. Like, I think there'll probably be some more merging there. I wonder if we can try it out. Is there, like, a free license thing anywhere? Yeah, I'm sure we could. Oh, yeah. Okay. Maybe we'll... Start for free. Oh, there you go. Yeah, we can test it out. I know it's got integration with World Labs. I did ask if you could import Gaussian Splats. They said not yet, but that shouldn't come.

10:10Yeah. So that's my AI. That's your TIFF impression of my AI summary of TIFF. That's interesting that Kling out of all the AI manufacturers were there, and ByteDance was not there, and some of the American companies were not there. It's a weird venue to decide if you want to go all in on that if you're an AI company. There were a couple other companies that I saw that had booths that were more of like aggregator tool set companies that were focused on film production. But like they're not a company that makes any original models. I forgot their names, but it was like the usual stuff to help like, hey, you can storyboard faster.

10:47And it's just an aggregator tool set that's calling on different models. Okay, okay. It's sort of on the Magnifique Hicks feel kind of. In that pocket, but I would also say definitely not as complicated, not as complex or powerful, not as robust as those. yeah um but cling was the only like original model runway had a satellite event uh where they played some of the films and stuff but uh not like an official presence yeah i'm i'm enjoying the runway ai festival uh short films and they're kind of just dropping them on youtube here and there yeah i mean so they played all of them and a lot of them are great yeah exactly like people are doing good work in the medium i forgot the name of it but the one that looked like vhs found footage Oh, I didn't see that one.

11:32The one I saw was like a full metal fruit metal jacket. So it's like full metal jacket, but with fruits. Oh. Yeah. I don't know if that's part of the festival, but I know what you're talking about. Yes. That one did look good. All right. Next story. So right after our last episode where we pleaded for the AI companies to not release models while we're recording, ChatGPT released Image 2.5 right after we recorded. So good. Hats off to, I mean, And look, we're not going to get into the whole OpenAI thing, but from an image model perspective, GPT 2.5, I think to date, is probably the most capable model.

12:10I would say, yeah, most capable for sure. The edit features in this are probably one of the strongest attributes. I like the preservation, which, so for me, it was like, you know, if you put Joey or Addy in there, it knows how to keep our likeness intact while changing everything about us. Even putting makeup or hair on us, you still won't lose the likeness, which is amazing. It's completely generating that. On the left, you've got image two where it's modifying the image, but everything with the person is kind of the same, but they jump around and they kind of shift. And then you have the comparison with image 2.5 where the person's clothing, the background's changing, but the person is staying rock solid, like nothing is changing about them.

12:54Yep. So I did my usual 1970s New York City street test with both GPT-2 and 2.5, and Joey was once again like, hey, I don't see a difference. I'm just kidding. I'm just kidding. Okay, so this is too, let's see. I mean, yeah, it kind of looks like at first glance there's some gibberish text. Okay, but to me, 2.5 just felt more richer with detail. So it's almost like... This feels like it used more world knowledge and web search to build this image. Exactly. The Godfather poster, Burger King logo that's period accurate, Coca-Cola logo, JAWS. And I think the reason for that is GPT-6 Astra. I think underneath the prompt, it's probably going in and doing a bunch of prompt rewriting or prompt boosting with GPT-6.

13:46and then coming back with just like a gazillion things to add to the prompt. Yeah. I mean, we still have issues where like the Empire State Building looks like it was squished. Yeah. Everything else. I mean, that's probably the biggest standout issue in this. Everything else, yeah, theoretically. Yeah, I think. Vime-wise, it feels like 1970s. If you look at it just on its own, it won't feel any different. You're like, okay, man, I don't see the difference between this and Nana Banana or whatnot. But then when you look, go back to 2 and then come back to 2.5, yeah, you'll notice there's just so much more things in that world, so much little details like the signs and the people and their outfits and the variation in cars and so on.

14:32So, yeah, again, it's just like one notch up in incremental reality. Yeah, the sharpness in the text and in the people feels a lot better in this. and also you texted this to me and I think I'm looking at a compressed image so when I'm scaling up, keep that in mind this feels like it's extra compressed the edit stuff, so let me go back to this so the modifying the person nothing changes solid, this one is interesting, test Higgsfield did that they gave it it's sort of like that echo test where they give it a starting image and then had a prompt that just says create this exact replica of this image don't change anything image one is obviously a joke but image two it starts getting easier and then image 2.5 it looks just solid it's crazy pretty much exactly the same so now there are two variants of image 2.5 there is sunburst and flare so flare is the faster cheaper model that they're saying is more for just like ideation creative work don't need the accuracy and then Sunburst is the higher, the pro higher end version of the model.

15:43That's what's running these image edits and modifications and it's a bit pricier, but really good at doing very specific edits to the image and not changing anything else. Yeah, the interesting thing about Sunburst is when you pick Sunburst, which is the higher quality model, there is even an option on the API to go up a notch in quality. So it'll go from Sunburst mid quality to high to max to ultra max. There's tiers of quality. Yeah, there are these quality. I get these quality settings. Go to additional settings. Higher settings, increase detail, latency, and token usage auto lets the model choose.

16:19And I think the resolution is higher than 4K. It doesn't say here. Oh, here. Pricing. Yeah, it's got 4K listed. 4K is the highest. Seems like it. Yeah. Okay, I was wrong. 4K is the highest. 4K at max is about 40 cents an image. Yeah. That's expensive for one single generation. Yeah. It has been not crazy for what it could do. Right. Especially if you consider that it has amazing prompt adherence and likeness preservation and things like that where you won't need to generate again and again. Right. So it could be worth spending more on that one shot. Yeah. That's a good point too. If it does your thing in one generation, then that saves you from having to re-spin and ultimately costing more than 40 cents.

17:08The other thing, I haven't fully tested, but apparently it also has really good world understanding or positional understanding. This test, and I haven't replicated it to see if it's actually true, but this is a Rubik's Cube in the mirror test to see how well it can replicate reflections and if the patterns are correct, like of what it should be with the jumbled Rubik's Cube. And so this is saying this post from Danders. It feels wrong. The reflection is wrong. Does it feel wrong? Yeah, so the line should match. reflection itself should be in the same plane as the object and it's like skewed by 10 degrees.

17:57Oh, you mean like the slight you're saying like the angle of the like the angle should be the angle is off, however, just looking at the colors and how it mirrored the colors. So if you're looking at the bottom row, the real object is yellow, red, blue, and the mirror is yellow, red, blue like that is matching. I think the bigger test is what we're seeing here in the reflection on the side that we can't see of the quote real rubik's cube correct are these patterns consistent i don't know rubik's cube well enough to actually verify i know imagine if we yeah if we had the world champion rubrics guy yeah yeah he would be like oh that looks right yeah right the bigger i think the bigger question is are those patterns real and accurate to the rubik's cube and the reflection correct yeah they say it I'll take their word for it.

18:47I'll take their word. Impressive. So it's cool. Yeah, I want to mess around more with, I think, oh, did we test it with, was it Image 2? I should have brought those images back up with the movie theater, like the reverse angle. The reverse angle is always the test and struggle point with a lot of these models. I think we did that when Nano Banana 2 came out. Remember? Oh, maybe it's Nano Banana 2. Yeah. All right. I'll have to pull up those same images and do another retest for reverse shot because I feel like this might be pretty good at that. Okay. Game on. All right. We'll show that next time.

19:18All right. New model. Interesting model. I'm going to try to explain this. Okay. Yeah, because I didn't understand it. So please explain to me like I don't know. Okay. Diojo Almeida. So they've had a new company called TypeSafe. He was one of the original developers, co-inventors of ChatGPT, left, started this other company. And they've just released their first model called Jev. The headlines are it's super fast, super cheap, doesn't hallucinate, but it's not a chat model. So my takeaway from this, it's a model that is built to deal with structured data. So I think we talked about this before where if you're like working with ChatGPT or Claude and you want to do something like with Airtable or spreadsheets and you're trying to just kind of have it organize a lot of data.

20:09When it does that, it's just kind of burning through tokens. it's either going to be burning through tokens and it's not very efficient or accurate because it could hallucinate and it's still predictive of organizing the data or doing large amounts of data organization or it'll just kind of write a script and then have the script do the stuff. And then you're basically just having it be a Python coder and you're going back to like old school tech and managing this data. Jeff works with structured prompts and structured data. So if you needed to classify a bunch of stuff or you needed to kind of figure out stuff, you sort of give the input prompt as some structured data of like, if it's this, then do this.

20:49If this, then do this. And it takes the input and then starts figuring that stuff out. And it just does it at a super fast, super cheap rate. It's like$45 per billion tokens. Whoa, that's cheap. That's for input tokens. Output tokens are free because they're too cheap to meter with their new architecture. Got it. And it's kind of a predictive model tool too. So you put in the inputs and then the outputs are like percentage predictions of if it's true or false. Okay. Based on the criteria you set. I think I understand now. It feels to me like this AI model is made to be used by other AI models, not necessarily humans.

21:38Yeah. They see kind of a hybrid of using code and AI. And they're very clear, this is not a chat model. No. I just got access to the beta thing, so I'm trying to figure it out. But the first thing on the thing with the questionnaire is like, can you chat with Jev? And you had to say no. No. To verify you understand. Your front end to use Jev will be a Vibe-coded app that you built that spits out the information that Jev can accept in a manner that it can accept. And then whatever Jev spits out, outputs, will get turned into a thing that a human will understand through your Vibe-coded app. So the Vibe-coded app is like the wrapper around Jev.

22:24And that's how you're supposed to use it. Yeah, and when you want sort of this intelligent decision maker under the hood, the routes before were like, okay, maybe use a cheaper model or something like Haiku. Exactly. Yeah, or run it. I don't know if you can run it local. Like, this would be perfect. Oh, Jeff? I don't know. I mean, maybe in the future. But this is sort of that middle ground where it's like, okay, I need to do some kind of predictable things or figure out things that are more efficient with AI, but I don't want to run the risk of hallucination or the cost of doing that with traditional AI models.

22:57This seems like it fits in that middle ground. Yeah, 100%. And with sort of the cost being so low, you can then open it up to doing things that you otherwise would have never really pushed AI to do, right? just 100 calls a minute or some incredible amount of throw put that you need previously. Yeah, something where you would have been like, oh, we need just Python scripts for this or something. Exactly, it was too cost prohibitive. Or unreliable. Yeah. So it feels like a utilitarian machine-only model, basically. Yeah. Here's an example, too, of a traditional LLM prompt versus Jeff where if a common example is someone types in a prompt And then the model router tries to figure out, like, what model to route it at if you have an auto setting on.

23:44Like, is it a complex thing? You need to send it to, like, Astra? Or is it an easy thing that, you know, 5.6 could handle? And so the traditional LLM is more of, like, a text thing. You're a model router. Use fast mode for simple edits, blah, blah, blah. And then the output is what route it should go. And then with JEV, it's more structured data, type, choice, instructions, which mode. And you kind of define your criteria. and then it gives you probability of like fast mode 0.97, full agent 0.03. So taking structured data and giving you probability data. Yeah, and that probability data goes into like a front-facing model, an interface model that then figures out what to do with that probability, turn it perhaps into a code or lines of code that then executes.

24:33Exactly, yeah. I don't think you're doing anything fully with Jev. You're just using it as this intelligence. This brain. Yeah. Yeah. Yeah. So it just came out today. I'm dabbling with it. I'm kind of trying to figure out where does this start making sense to do things. Good luck explaining this to the TIFF people you met. Yeah, that's not it. Because we barely understand it. This is so five years from now. I think this is super ahead of its time. yeah yeah i'm very curious what what this can unlock so i'm gonna dabble with it some more in some of the apps and stuff and see what this could help with that i think traditionally would have been deemed too unreliable or cost prohibitive with with lms or yeah with lms yeah very cool all right let's go back to the creative side to talk about uh quick updates adobe premiere had a couple updates.

25:29It's more integration. They released a new generative media tool in Adobe Premiere, just the beta version right now. And they've done some really good stuff with just integrating Gen.AI into the traditional tool sets. So this is a new tool on your sidebar. It's a generative media tool. And then in your timeline, if you have a gap where you're like, I need to generate a clip here, you can click and drag. It creates a Gen.AI placeholder clip. And then you get like a little prompt box and you can use whatever models integrated with Premiere. Like they have Cling, they have VO, they have... Oh, so did it just take a start and end frame from the clips next to it?

26:10So you can. You can either pick the start and end frame from the neighboring clips or it gives you this eyedropper tool and you can pick on any frame in your timeline. You could pick any frame in your media pool. Sorry, they don't call it media pool. They call it... Media bin? Bin. Media bin. Your media area. or you can click on your playback window. So pretty much anywhere you see video, you can use the eyedropper tool to click and it just automatically grabs the frame. So it's just one of these really clever time saver things where you need to generate a missing clip and you don't have to screenshot your stuff and then download it and then re-upload it.

26:48It just does it all there. And then all the new media that you create, it goes into a new bin inside your project. So it's all very fast, self-contained round trip. while it's generating, you could keep editing. Well, when it's in the package, when it's in Premiere, it's like way, way, way easier than us going out to, you know, FAL or Magnifique. Yeah, it breaks our flow. Yeah, exactly. And it encourages you to use it more and experiment more. And if you don't like it, don't use it, you know? You still have the rest of your timeline. Yeah, it's there if you want to use it, but you don't have to use this at all.

27:19It also, they have audio integration too. So you could drag, use the same tool, you drag it on an audio track, and then you could either make sound effects, make music, or it was like another one was like ambient sounds, and that one just looks at your visuals and then just interprets. Just vibes. Just vibe mixes. But the sound one's cool because you can drag out the audio track and then you can either do a text prompt or you can turn on the mic and use your voice. As scratch audio. As scratch audio. So like I did a test with like car sounds. And so you could like specifically be like, like at the exact moment you want.

28:01And then it turns it into realistic sounding car sounds. Yeah, the fact that it does it in package is amazing. I know Firefly has an experimental, or they have had an audio model on Firefly Home for a while. I'm guessing it's the same probably model that's piped into Premiere with the same functionality. I think so. I think it's all the same model, same thing you would find on Firefly, the website. Just saving you steps and having it built in to Premiere. Yeah. I've used generative extend before where, you know, if you have like a short clip and you just pull on it and it just gets longer, it's worked pretty well.