00:00So, there's been some interesting back and forth this week on prompting techniques with VO3, namely that of text versus JSON prompting

00:00So, today I figured we take a deep dive in to see if there really is a best prompting strategy, uh, plus some tips and tricks along the way

00:00I've also got a first look at a brand new cling workflow

00:00Yeah, you guys are definitely going to want to check this out

00:00And then we'll round out with, well, some good midJourney news

00:00So, kicking off with some fairly minor VO3 updates before we dive into our main course

00:00This does mostly involve the flow platform

00:00So, if you're using V3 via API, uh not a lot of this will necessarily relate to you

00:00Namely, it looks like they have made some improvements to the scene builder function in that it now autosaves your work

00:00That actually has burned me in the past

00:00The other kind of minor updates is that text to video in V3 will now generate uh voices for a subject of all ages

00:00I guess that is not exclusive to FL

00:00That is going to be available via API

00:00I can't believe we can finally talk like this in videos

00:00I know

00:00It's like magic

00:00Now everyone can hear our real voices

00:00Overall, pretty minor updates, but it is good to see that they are continuing to work on the flow platform as I do think that there is a ton of potential there

00:00That out of the way, let's bite into our big kahuna burger

00:00Uh there has been some pretty interesting back and forth this week on prompting techniques in V3, namely again text versus JSON

00:00So just to bring everyone up to speed on JSON prompting in case you're not aware, uh JSON stands for JavaScript object notation

00:00You can kind of think of it as a way to organize information into categories that makes it much easier for machines to understand, but is still legible to, you know, the human eye

00:00The JSON format was created by Douglas Crockford, who developed the idea over the course of a summer at Camp Crystal Lake

00:00Only one of those things is true

00:00In terms of the origin of JSON prompting in VO3, well, I mean, to be honest, there really is no clear answer, and I've gotten like digging

00:00Instead, the idea seems to have sprung up organically through a lot of community experimentation, which I I don't know

00:00I kind of think makes it cooler

00:00Now, it is important to note that within JSON prompting, there is no standardized formula

00:00For example, friend of the channel, Stealthy the Time Traveler, recently posted up a JSON prompt that uh he utilized

00:00And I know if you're looking at this and you're not familiar with JSON prompting, you're like, "That just looks like a wall of crazy text." But when you break it down, you can start seeing the categories uh such as scene and action here

00:00Uh we do have camera angle, lighting, aspect ratio is here, and then, you know, action and motion down here

00:00The end result is a video that, well, let's face it, uh, makes anyone who has ever built anything from IKEA super jealous

00:00Whereas the formula that I like to use has these as kind of our category buckets

00:00As always, if you're interested in getting the GBT gem builder that I utilize to, you know, put this all together, uh, that is available totally for free over on Gumroad

00:00Now, as far as what JSON actually has to offer, well, it ultimately comes down to more control

00:00Uh for example, if we just use natural language of a fast red sports car, that leaves a lot of options for the model to decide what does fast mean, what does red mean, what does sports car mean

00:00Whereas with the JSON prompt, we can go as far as to give it a color hex code for what color red that is and even like how fast the car is going

00:00Interestingly, when it comes to efficiency, at least for simple prompts, uh natural language actually is the winner there

00:00uh something like a cinematic shot of a red rose only runs seven to eight tokens whereas the JSON output of that same prompt uh it gets up to 20 to 25 tokens but conversely with longer more descriptive prompts which is something that I think we all tend to favor u that is actually where JSON ends up coming out on top namely due to the fact that it cuts out glue words or articles prepositions and descriptive phasing a in a from with etc etc all that said the debate four natural language prompts

00:00I believe kicked off with a tweet from Jason Zada, who you might remember as the filmmaker who created the V2 short film uh The Getaway, as well as the Wuang video for Mandingo, which is awesome

00:00By the way, Zada called out JSON prompting as not intuitive or expressive actually

00:00Uh it's like trying to write a love letter using Excel

00:00That's pretty that's pretty funny

00:00Instead, he advocates for using natural language with a structure, which he does outline here

00:00Uh, and I do want to note that he's not saying don't use an LLM

00:00He's just saying, you know, use natural language and an LLM with a structure like this

00:00Lots of debate back and forth on this leading to the question, well, is there a difference in the output? Well, let's go burn some credits and find out

00:00So, checking in with our man in the green tuxedo, who you may remember was last seen uh being left at the altar by the woman in the red dress

00:00Here he is in uh natural text language prompting

00:00She left me, but I'll have my revenge

00:00I'm going to find that man in a blue business suit and make him pay

00:00Oo, it looks like we have the origin of our villain

00:00Overall, pretty solid output

00:00I did utilize the prompt structure that Jason had outlined earlier run through uh Gemini 2.5

00:00Um, yeah, solid output here

00:00Little bit of a garble text in the background, but I mean, at the same time, you know, that's it's just AI video

00:00Now, for that exact same base prompt, this time run through a JSON prompt

00:00She left me, but I'll have my revenge

00:00I'm going to find that man in a blue business suit and make him pay

00:00Overall, also a pretty solid output

00:00I mean, back to back, they're very, you know, similar, but I do want to point out that in the natural language version, we did call out for the camera to begin on a medium shot and slowly push in, tightening on his face as he delivers his line

00:00Uh, that did not actually happen here

00:00Instead, our actor actually walked towards the camera

00:00Whereas in the JSON prompt, it called out for a single continuous slow push in starting in a medium shot and then ending, you know, on a tight close-up, which indeed we did get

00:00Another quick text to video test

00:00This time, uh, kind of incorporating a 70s crime kind of taxi driver vibe

00:00Yeah, that's pretty solid

00:00No complaints there

00:00Okay, let's try that out

00:00The exact same prompt again with JSON prompting

00:00Now, in terms of the actual shots, I think that both are pretty much on par with one another

00:00Uh, that said, obviously the JSON uh music cue was, I mean, kind of lacking

00:00It was likely part of the prompt of a tense, minimalist score with a funky baseline and quiet symbols

00:00Um, none of which we actually got there

00:00But again, uh, when you're queuing for music in V3, it's it's always a bit of a dice roll

00:00One last text to video test before heading over to image to video

00:00You know, that whole love letter in Excel thing really did get me thinking like how would VO react with kind of more abstract prompts

00:00So, I took the opening lines to TS Elliot's The Wasteland and just ran those as just a straight text prompt

00:00Uh, and this is what we got

00:00And for sure that output is picking up things like April's the crulest month, the breeding lilacs, dead land, um, dull roots, and spring rain

00:00So it was really interesting to compare that to the JSON output which looked like this

00:00And while I'll admit that does look like a bit of a mess, particularly like the whole bit with the woman speaking German, uh when you dig into it, what you can see that what it's doing is it's actually picking up a lot more of the poem from the lilacs uh to the woman saying that she is not Russian, she's from Lithuania, uh even down to um the kids out on the sled

00:00Now granted, we are going through an additional level of thinking by taking this through an LLM to output the JSON prompt, but I do think it also showcases that utilizing the JSON format, you can get a lot more granular

00:00So on the image to video side, I ended up utilizing this scene

00:00I wanted to have the two characters in the foreground there, uh, discussing the bunker and the end of the world

00:00Uh, but to have a, you know, hold that as an establishing shot and then cut to essentially close-ups of them

00:00Um, this is something that V3 is very good at

00:00So, uh, our initial just straight natural language text prompt yielded this

00:00Please, you have to listen to me

00:00We need to get to the bunker

00:00It's the only way

00:00I know, but is it truly safe? I mean, that is pretty good

00:00Squint your eyes a bit, and that's that's an HBO show

00:00Now, I will say that Vio did the thing that it can do from time to time, which is jump from that first frame uh into essentially the same location

00:00It just gives us completely different shots

00:00There is a way that we can utilize this to our advantage that we'll take a look at later on

00:00Uh but yeah, we did not get that establishing opening shot

00:00So, rerunning this as a JSON prompt, it does not come out perfect, but it does actually lead to a pretty interesting trick

00:00The bunker is our only chance

00:00A chance to what? To be the last ones left

00:00So, yeah, pretty solid other than the fact that we did get a fairly jarring jump cut there

00:00By the way, is that the bunker that's right behind them? Why are they looking for the bunker? It's right there

00:00So, one thing you can do, and I've been utilizing Gemini for this, mostly because, you know, I have the ultra plan, so I might as well use it, is just upload the video back to Gemini and say like, "Hey, this is weird jump cut

00:00Uh, can you give me a prompt that fixes that?" The bunker is our only chance

00:00A chance to what? To be the last ones left

00:00And it did indeed fix that cut

00:00Now, granted, it did not give us another cut

00:00Instead, it gave us a push in

00:00and also put a weird like howling wolf in the background who is clearly going to eat everyone in the bunker by the end of this movie

00:00But I did want to highlight that is a way to work with an LLM when you're not getting a desired output

00:00You know, just ask

00:00Now, I do want to point out that JSON prompting can sometimes get frustrating, especially when you're trying to diagnose a problem

00:00Uh, take for example our, you know, hard-nosed nor detective here

00:00Uh, running this with a natural language prompt to get a twofer out of it

00:00We end up with this

00:00It was a Tuesday when she walked in

00:00Nothing good ever happens on a Tuesday

00:00Overall, a pretty solid output

00:00I am a little concerned about our guy's liver considering that he's on his third glass now

00:00And we can tell by the lighting

00:00The sun hasn't even gone down yet

00:00Um, yeah, our guy is putting the hard and hard boiled

00:00Now, the problem with our JSON output of the same prompt was, well, I'll I'll let you check it out

00:00It was a Tuesday when she walked in

00:00Nothing good ever happens on a Tuesday

00:00I mean, is crime noir, country, musical a thing? Uh, I apparently I guess we just made it and I got to say it was a bit of a hassle to figure out

00:00Gemini actually couldn't figure it out

00:00Uh, it turns out that it was something between the voice over uh the grally and tired and then the background melody playing continuously

00:00Something about those combinations for some reason triggers country music

00:00Now, one additional benefit, uh, kind of a hack that you can utilize with JSON prompting is something that I've documented a few times on the channel

00:00If you use an image prompt, you can use a prompt like immediately cut to a completely different situation

00:00And a lot of times the attributes of that image will then carry over

00:00Now, this does not work all the time, and a lot of times, especially with characters, you'll just have a character kind of hanging out in that initial environment

00:00That said, with JSON prompting, we can take an image like this and immediately have her move over into like a cyberpunk setting

00:00[Music] Where can I find the dark wizard? Or for some strange reason, she's picked up a gig as an office temp

00:00Where can I find the dark wizard? I love the fact that Bill from accounting here is like, I got to talk to HR about this lady

00:00But yeah, by utilizing a JSON prompt of something like scene description, then timing it at zero to zero seconds um and then just completely changing the prompt, you do seem to have a very high success rate

00:00So ultimately, when it comes down to natural language prompting or JSON prompting, you know, I I honestly I don't think that there's a clear winner either way

00:00As I always say, there is no right or wrong way to prompt

00:00You know, experiment around, see which one you gravitate towards or which one you like more

00:00So, definitely let me know down in the comments which you've been having more success with

00:00Uh, okay, let's go check in and see what's going on with Cling's new workflow

00:00Uh, but first, a quick word from today's sponsor

00:00If there's one thing I really love about the community of folks around this channel is that we all come into AI content creation from different backgrounds

00:00We've obviously got a lot of video creators, musicians, uh, illustrators, 3D folks, but there are also folks here very knowledgeable in world history, and I've definitely picked up a recipe or two down in the comments section

00:00So, even if you're just getting started in AI or you're like a comfy node wizard, I'd wager a good dollar that all of you are an expert in something else as well

00:00And a great place to share that knowledge with others is Teachable, who were kind enough to sponsor today's video

00:00Now, I'm sure you're all aware of Teachable

00:00They have been a staple of the online education space for quite some time

00:00But what you might not know is that they've recently launched an AI course curriculum generator uh to help you build and outline your course

00:00So for example, we can begin just by describing a course

00:00In this case, we'll go with a midjourney prompting course uh taking students from the basics of image generation and midjourney all the way through the more advanced commands and then finally into video

00:00Uh we simply hit continue here

00:00And as you can see just from that description, we already have an entire like course outline

00:00Now it is important to note Teachable does not actually create your course for you

00:00That is still your job

00:00Um but you can see that it makes it very easy

00:00We simply add in some content either text, images, video, even audio

00:00Uh and what's kind of cool is that you can actually generate quizzes as well uh or even just autogenerate quizzes for each individual section

00:00There are also design templates available to you to give your course a professional look

00:00Uh let's take a look at this one

00:00And although this course does not have any information in it yet, we can see that professor man in a blue business suit is now taking attendance

00:00From there, there are a whole host of other tools available to you to manage unlimited amounts of students

00:00Uh they do have an app hub here that will allow you to connect from everything from calendarry

00:00Uh Zapier is in here

00:00I did see that

00:00And Mailchimp as well

00:00So yeah, just a lot of, you know, integrations and options for you

00:00Teachable also pays you out very quickly for any courses sold via Teachable Pay, which is available globally

00:00So basically, Teachable has everything you need, you know, all right on platform

00:00So if you've been thinking about taking your expertise and well, I mean, anything from AI video, Comfy UI, or I mean, again, anything really, like I'm still trying to figure out how to finish this train wreck of an experiment

00:00And you know what? Teachable does have a course on that, which I do think does go to show that whatever your skills, Teachable is a great place to turn it into a digital product

00:00And in good news, Teachable has given us an exclusive 30-day extended free trial

00:00Hit the link down in the description to start getting paid to share what you know

00:00Thanks to Teachable for sponsoring today's video

00:00So, sliding over to Cling

00:00They have introduced an entirely new workflow

00:00Now, I do have to say right off the top, this is super early access

00:00From what they've told me, at least if you're watching this video early on, we're looking at at least 2 weeks before this goes into wide release

00:00But I am genuinely curious to see how you guys are going to react to this because uh well, as it turns out, Clling is going full murder board

00:00This will be a new feature called Cling Lab, which will be located down here presumably at least

00:00Uh and yeah, what this will lead to is uh yeah, a node-based workflow

00:00Now, I do want to point out I have not had a ton of time to play with this

00:00In fact, this just dropped into my inbox a few hours ago, but I did want to give you guys at least a first look at it

00:00Uh I will be back later on with like a full deep dive into all of this

00:00Uh but yeah, as you can see, we can essentially begin with a text box that can go over to uh their image generator colors

00:00Uh and then we can out that node over to uh image previews and then out to video

00:00So, um, yeah, and then we have each of our videos here

00:00Beyond that, it does look like they're adding in a few new features as well, such as co-pilot, which acts as kind of a creative assistant

00:00Um, interesting one in here, uh, is that it can, uh, analyze image and video content

00:00So, that's kind of cool

00:00And then as an additional bonus to using this asset node system, uh, it does appear that you can do things like combine multiple assets uh, and then create an output from that

00:00That is something that I will definitely be digging into uh in a video next week

00:00Look, I know that nodes, spaghetti noodles, and murder boards are not everyone's favorite thing, but for me, I've always found it to be a pretty interesting working style that can sometimes lead you into some uh interesting results in places that you did not expect

00:00Now, I will say that Cling Lab is currently a bit on the limited side

00:00This is like day one, so I'm not knocking them too hard

00:00Uh but again this is something that I had mentioned when Kyber first switched over to their nodebased workflow and even Flora uh when they first got started in that um you know there are a lot of outs

00:00Let me see if I can grab that node

00:00Uh there are a lot of outs but there are not a lot of ins

00:00Um so that is something that I presume will develop as more features are added in and you know once you guys all get access to it um you guys start telling them what you want to see

00:00But in the meantime, yeah, I'm pretty excited to start putting Cling Labs through its paces

00:00Now, to note, if like you hate nodebased workflows and like all of this is giving you a headache, that's okay

00:00Cling like regular is not going anywhere

00:00You can still use uh the standard Clling interface

00:00This is just a different way of working and I I hope as time goes on uh will evolve into its own thing

00:00I'll be back with a deeper look at this sometime hopefully next week

00:00uh basically when I have longer than like 15 minutes to play with it

00:00Sliding over to some midjourney news

00:00Uh it looks like we'll be able to start generating in 720p for midjourney video very soon

00:00This was spotted by friend of the channel Christopher Fryan

00:00This was mentioned in yesterday's office hours

00:00So not only uh a 720p mode

00:00Uh potentially also a turbo mode is going to be coming our way

00:00Uh and then additional features uh will be a start and end frame and a looping feature as well

00:00All of which is really great news

00:00I've really been uh endlessly entertained by the midJourney video model

00:00It's really good

00:00Although, of course, the downside being that 480 raster size

00:00You know, it's it's pretty small

00:00So, it'll be nice having a 720 output

00:00At least that will give your uh upscaler or even like Topaz's uh Astra, the creative video upscaler, a little more to chew on uh when it starts blowing things up

00:00And it does look like the midjourney API is actually soon going to be a thing

00:00Finally, one little like midjourney hack trick that I've been meaning to work into a video is uh is this um you know how midjourney can generate in like weird aspect ratios? Uh like this is 118

00:00Well, apparently you can also generate video in those weird aspect ratios

00:00Uh something like this, I mean, what are you going to use it for? I don't know, but it's it's kind of cool looking

00:00Uh stack a couple of these together, maybe as an installation piece or something

00:00Uh I don't know

00:00It's just neat

00:00Um, I can't think of any other video generator that will generate in like super bizarre aspect ratios

00:00So, um, yeah, just plug some numbers in, see what happens

00:00It's fun

00:00So, yeah, I guess that's it for today

00:00Uh, I'm going to go hop over to Clling and see how many nodes I can wrap myself around

00:00Uh, I will definitely be back very soon for a much deeper look at Cling Labs

00:00Uh, in the meantime, I thank you for watching

00:00My name is Tim

00:00