AI video camera control with GPT-6 Astra, Blender and FLORA

Lightly edited from the recording for readability. Wording is the speakers’ own.

Hello, everyone. My name is DJ, a creative workflow specialist at FLORA. In case you haven't met me, I work at FLORA as a workflow specialist, meaning FLORA power user in-house.

So I teach people how to use FLORA in the best ways. We used to host these livestreams regularly. I was kind of out of it for a month, and now I got back in. I used to co-host these with my colleague, Sasha. He did a cool museum challenge last week, and it was cool. Today, my topic is showing precise camera control with our open AI's newest model, GBT6 Astra. I'm sure you guys know.

I mean, if you scroll like 10 seconds on Twitter, LinkedIn, Instagram, you might see the name Astra all around. For those of you guys who don't know what Astra is, it's just a super goated AI model made by OpenAI. It dropped like last week or something. Yeah, but it's kind of phenomenal. I mean, we live in an age where AI models are dropping like every week, every month, and even every day, but this one is kind of phenomenal. I think it's worth everybody mentioning. We also got on to kind of like this Astra Wave.

Because I come from a video background, my biggest interest was using this to create 3D previews with Blender and plug that into video models in FLORA. I kind of posted some content over the week. I can actually pull this up. 3D previews to see this. So this was the kind of content that I posted This is already out and it already has tutorials on I don't think it's like worth showing this again but It was cool. The concept is making this blender previous Animation in like a white box style very simple low poly one I provided like two or three reference images of the real dominant sugar factory design architectural design and This new model astro just opened up blender and It like made the whole 3d modeling and animation itself of course I kind of drove the direction and I told it to like start from a nest establishing shot and Go through a single window of the building and land inside showing the bottom of the building Yeah, this is actually where we work the floral offices in the dominant sugar refinery show but I potentially kind of made it empty to just make it look like surreal and like historically this was a sugar factory and we built like an office full of like tech companies inside right now, but Yeah, that'll be cool and Yeah, I plugged the blender previous inside With the right reference images Using these two images actually these two images were the exact images that I provided to Astra in order to make the 3d model and I have a couple more here first of all I Visualized based on like screenshots of the previous video so Because I have the accurate camera angle and shot size I could get like very accurate angles of the dominant sugar factory building and Then I plugged both in with the motion reference. Also, that's how I achieve this video Yeah, so I'm gonna try this kind of a pipeline again for today's workflow So actually this live stream was very was posted very last-minute we were kind of busy recently, but I was able to cook this up like just This morning because I realized we didn't have any materials on socials in order to like notify people that I'm going live today I did the same workflow using the Tokyo Hermes store It's really it's a really beautiful building. I never been there, but I just know that there's a really beautiful Hermes building in Tokyo.

I provided the same way I provided three images to Astra and It created this whole blender world and it rendered a 720p low-poly Blender camera movement video and then I used floor MCP to Import the video that was made inside blender and in fact, I use floor MCP for everything I was just kind of like testing because we're really, like Actively building in MCP right now MCP got really better So yeah, I use this low-poly 3d motion Footage here. I did the same thing as the dominant sugar factory one I plug these in into image models and created like the real Hermes store with the real texture, like these should all be glass so That's that's what the MCP actually Prompted so it made it like this photorealistic version and then using these as the first frame and last frame and The motion reference video plugged in together. I Was able to make this like super realistic dolly end footage to the Hermes store so yeah, I'm gonna show something like this today with like some different examples And the interesting point here was I didn't use Seedance for this one I used cling o3, which is I believe it was kind of outdated The MCP chose this because Seedance was kind of failing at that time. Maybe there was a problem with the model provider, but like it was a technical failure though But I thought it was interesting because it still made a good video out of cling o3, which I thought was outdated I think the biggest reason is that I already had this very detailed visual reference, so that's why I was able to use a more lower TO model than CDANs 2.5, but still got very great results. I think that was worth noting. As I always say in these live sessions, all that matters is a clear visual reference, whether it's for motion, or for the visual consistency or whatever. We have everything here, like the first frame and last frame that has the consistent store design, and we also have the camera movement reference here.

I think this is a great workflow, and Astra really does a good job on powering the back end of this pipeline. I see some questions in chat. From your images, or does it research the building as well? Because Astra is a model inside ChatGPT, especially CODEX, it has the ability to do web search. It has its own intelligence. It's an LLM, so it can pull information from the web. I can just drop in the word Tokyo Urban Store, and it will just do its research on its own.

If I just tell it to skip the research and use the images itself, it can do that, but if I want more context, I can direct it to take in more context. Today, I'm planning for a two-hour live stream. I still kind of feel bad that the Twitter live stream is not working, but I'll just proceed. If you guys want to follow along today, you can, but it might be kind of tricky compared to other live stream sessions, because technically, you need some dependencies and apps. First of all, you need CODEX, the desktop app version, or the CODEX version. Oh, someone says that Twitter is working. Thank you so much for letting me know.

That makes me feel so much better. Maybe it was just latency. But anyways, just to recap, today, as it says in the description, I'm going to walk through live of how I built this 3D previs to photorealistic video type of pipeline. So there are some requirements. I mean, first of all, you need to be signed in FLORA. I'm not saying that you should follow along, but yeah, you can just like watch me cook.

But if you want to do something on your own, it requires, first of all, FLORA account. And second of all, you need a chat GPT account signed in order to use the model Astra. I just recommend you download the desktop app. And it's either the desktop app or the command line interface, which is called CLI. I prefer this way because I'll just pull up the window. It kind of looks like this. It's running on the terminal.

But I mean, developers might be familiar with this interface. But for those of you guys who are not familiar with coding, you might be oh, what the hell is going on? It looks too complicated. It's actually not that complicated. It's just a terminal window. And you're doing the same thing as in the Codex desktop app. Yeah, I'm not sure about the total details, but it just feels like I have more control.

And yeah, for like writing code, especially because Blender is an open source app made out of Python, it feels like the controls are a little more faster. But the Codex desktop app is also enough. I also have the Codex desktop app. Yeah, so I'll try both. I'll try both in parallel and see what happens.

So yeah, that being said, if you want to follow along, you can also open up your own harness to run Astra. Also, you need Blender. And because I'm using MCP, you need the blender mcp connected which is inside edit and you can go to preferences the shortcut is command comma and you'll have this window open up you can search mcp here there's only one option coming up mcp for blender and you can just check that option then codex will automatically like recognize the blender mcp server and it will do its job let me quickly check that's what I'm going to show today actually but for short answer I'm using chatgbt to drive the 3d part of this workflow also just to make things easier I kind of made a skill yeah skill and default I'm thinking how I should distribute that in this chat but yeah ideally I can like if any of you guys are curious and you want to try on your own I can send the skill to you guys to like activate it right away inside chatgbt and yeah the skill has the context of driving blender and importing that video into flora and all the way to the end so it's an end to end skill actually have it open here yeah it's skilled on empty file yeah it's called blender previous plus flora what this does is you can just download and unpack the skill on your local environment and yeah it will just do the job because it has the concept on just to like show a quick preview of what this does is it reads the first of all it reads the reference so the first step if you want to activate the skill is you should drop in a couple reference images of the kind of building or architectural design that you want to work with for my case it was just these simple images of the Tokyo Airmen store I found it on Google and I just copied and pasted and drop like three images into chatgbt and then it reads the references then it drives on its own into blender and yeah creates all the scenes and the skill also tells Astra to export a 720p like low res camera movement video something like this so yeah so based on this it can AI can read the context of the motion reference so that it can like read texture on inside flora using either seed hands or clean models so yeah that's the basic overview of this skill yeah you don't have to read everything here okay yeah it's a lot in this pipeline but I'll try to walk through something live starting from scratch okay you can stick around if you're curious of this process I'll see how I can distribute the skill maybe I can like post it on my X or LinkedIn with like a Dropbox power or something but yeah also at flora we're planning to like distribute skills later because like we see a high value in skills it's a very like global thing that you can apply to so okay first of all let's get it started by the way if you guys have any ideas of what kind of drone POV footage I should create or like your favorite architectural structure or like I don't know like Statue of Liberty or something just drop in the chat I'll try to find some inspos on Google or honestly I can just make up like a virtual place okay first of all I need a new project how about CSGO map. I'm sorry, I don't know what a CSGO map is. I'm not much of a gamer, but I definitely know what Counter Strike is. Cool. I think this is a good idea.

I'm not going to use this project for any commercial use, so I hope just copying and pasting from Google Images is okay. In order to stress test, I'm just going to use one image this time. All the previous results, I used more than two images. You have about 60 people watching on Twitter. Good to know. Thank you for telling me that. I just worried nobody might be able to join from Twitter.

I actually never played Counter Strike myself. I only saw my friends play it. Actually, this map kind of looks familiar. I do like shooting games, like first-person shooting games. I played a little bit of Battleground because I grew up in Korea. I also play a lot of mobile shooting games. I think I kind of like this one.

I'm going to try in a new environment, not in the CLI. I'm going to try in the desktop version. First of all, I copied and pasted the image. Someone says, "Does 2 map in CSGO?" I think I know this map. Maybe I've played it. Copy this map. I'll use this one instead.

In order to start from scratch, in order to call a skill, you can just slash command. The skill that I made is named Blender Previs + FLORA. It actually has the short description here. Let me show you. For reference image to 720p Previs and FLORA video.

Then I copied and pasted the image. By the way, I'm working inside a folder that I set. This is mirroring the exact folder that I set on my local laptop. I'm going to organize my projects. I'm also just going to share these organization tips that I use. I don't know if it's helpful, but I like to do this because I also like vibe coding and pushing my stuff to GitHub. It just helps to organize stuff for later.

Blender Previs + FLORA. Let's just start with make a 3D game map in Blender based on the image. I'm kind of debating, should I do this first and then move on to FLORA later? Just write the whole prompt end to end. Let me do this first. I was just saying, I'm just going to cover the 3D part first and then depending on how Codex cooks, then I'm going to decide later whether I should move on or fix some stuff. By the way, I have full assets on, which means letting Codex just drive it all by its own.

It is kind of risky, but I like to do this because I'm lazy. I just hate clicking approve, approve. Let me check the chat. Can we ever have Genjitsu inside FLORA? Genjitsu is something like this. I mean, so the thing is, I know Hixfield is doing a lot of different kind of content. They're like well-surfaced in terms of marketing, but yeah, if you use FLORA, you'll know that everything is possible inside FLORA as well.

I can pretty much confidently say that because I've usually been making videos with AI. I've been in FLORA for like over a year now and usually what I do is, like make previous to video type of workflows or like shot by shot based on a shot list or based on a script type of workflows. Everything is possible inside FLORA is more yeah, just like give you my take on this comparison between two different tools. Hixfield is more like a very, like fancy but messy toolkit that just provides you with a bunch of templates and it just blankets everybody online with, oh, we can do this and we can do that.

But FLORA, you can do everything the same as in Hixfield. Better. I don't know. I'm a power user of FLORA myself.

So anyways, but what I'm saying is that FLORA, because we have like a node based cameras as the basis of our toolset, but we also have like easier tools like Fashion Studio or App Mode technique and stuff like that on top of the layer. You can use these as like building blocks and make your own tools and especially because we have nodes, we can make your own workflows. And not just because I work in FLORA, I think like the infrastructure on the FLORA platform, it's pretty exceptional. And yeah, it makes the experience very smooth. And I personally work most in video making and VFX stuff like that. It's been pretty satisfying. The overall experience is pretty satisfying.

Before I'd been using like comfy and stable diffusion and stuff like that like a while ago. Oh, we have a question. Oh, we are watching the screen. Blender is actually running, creating the whole scene, but we're not seeing it. That's a good question.

So I ran this same thing over and over again. It seems like it is running on the backend. It's not like opening a new Blender project itself and like showing in front of me. As you see, like the project is still pointing to the Hermes building.

But yeah, it's saying that it's running Blender. And it says that it has the lower tunnel upper hill. So yeah, I guess it's running on the backend. But yeah, this is what happened as well for the Hermes store and the Dominature factory as well. Which is kind of interesting. Yeah, just heads up, it might take a while.

So yeah, today is a very chill session. And yeah, in case if you want to like follow along, just let me know in the chat. I'll be constantly checking. In the meantime, because nothing is happening right now. Oh, there's a the loading wheel is running. I guess it's cooking. Oh, shit.

Damn. Okay. I'm always kind of odd. Wow. That's cool. Yeah, by the way, for those of you guys who missed it, I think this whole thing just happened in what, like five minutes or something? Oh, yeah, five minutes says it's working for five minutes and a half.

Good take on a site. I guess this is the site. I haven't played a lot of Counter Strike. What is this thing hovering?

Okay, cool. I think this project is gonna be pretty fun. Because it's a game map, probably I can drop in some human characters and make like a I don't know like a third person shooting scene as well. Does it work on a Mac mini? To be honest, I haven't tried. My laptop is, let me actually check here.

This is my laptop. M4 Macs 48 gigs memory. I haven't updated my Mac OS yet, but yeah, well, I mean, my hardware is not bad. So I haven't had problems pretty much.

But I mean, as far as I know, they launched a new Mac mini, I think recently, right? Yeah, but just speaking about the technical like details under the hood, it should just technically run flora on the browser. And codex the chat GPT desktop app, or the terminal on a codex CLI. And also blender. I guess blender would be the most like computationally heavy process.

So if your computer can stand blender, maybe it can hold it. Is DJ single? Unfortunately, DJ is not single. What's up, Timothy?

Okay. I'm still kind of brainstorming how I should build on top of this project. Because it's a CS map. I think the output should be like a first person shooting footage. Also, I wasn't planning but there are possibilities of developing this into a like a three j s browser based game. I can show you like a preview while Astro is cooking.

So okay, I have this versatile site. They're just built up shit. Okay. Thankfully, it's working. Yeah, anyways, I made a game based on some photos of our office. Actually, this content is going to go out pretty soon and the flora official accounts.

But yeah, these are actually based on photos, like real photos of the flora office. And I made a game out of three j s, which is a library that lets you build 3d stuff on the browser is pretty fun. This is our presentation room. It's kind of far from what it really looks like.

But okay, chive GPT is sending me some notifications. Okay, cool. Yeah, anyways, I'll turn off the sound for a moment. Yeah, just to show how this process was in the meantime, we actually find this project.

Okay, here it is. So yeah, I had like six photos of our office plus our rooftop. You see the flora sticker on CRT minor here. This was my laptop with my earbombate. And this is our bathroom on our rooftop. I made these into like 90s Lucas art style, like pixelated graphics. I kind of liked how they turned out.

And the same process I asked Astro to make blender scenes out of all these. And yeah, I plugged them all into Seedance 2.5 and Minimax H3. It created some great videos. I put them all together like this. Oh shit. Yeah, so something like this. And yeah, I didn't stop at just making videos.

I let Astro just take it away, cook some more into this game. So let me know if the link works. It's on the first cell, so probably if I modify it later, it might not work anymore. But you can play it for now. What Astro is doing. Cool. Nice.

I made this scene. Okay, what kind of scene should I create? I think first of all, we can start with like a simple like drone FPV drone style. Yeah, maybe like from here to I don't know. What would be like an interesting camera movement? I think I might start from here. Actually, what I did was process is very raw.

I just kind of like took two screenshots from blender and I was oh, make a camera movement trajectory. According to these two images. I kind of want something like this. Starting from an establishing shot that shows everything all the way to this little building here. Maybe like this. I don't know if this too complex.

Okay. Now we have three screenshots. I'll just say make an 8-section. I wonder if I should use a skill again.

Okay, I'll just use a skill again just to like give context that it should be a 720p low poly video. A lot of waiting. It's almost like letting someone else cook for you on blender and you only use the you only review the results and use it if it's good and give feedback if it's bad. When I asked to do a little bit of 3D, I always had a mouse. Like because blender it requires that the scroll button click, but at some point I completely ditched the mouse in my creative process.

Now I only use a tripod and I'm kind of like more used to using these buttons. Some people might despise this, but I don't know, it's just more convenient, I guess. Blender is a free software, yes. It's completely open source, and it's free. And open source meaning that you can use their code base and make modifications yourself. And because AI is so good, you can technically vibe code anything you want now.

So Blender has an amazing, huge community with cracked people making tons of different kinds of plugins. And also because it's open source, it's good to use it with an MCP, which is what I'm doing right now. Using AI to drive the 3D modeling and 3D animation, camera work and everything. How much Blender knowledge do you have from 0 to 100? Honestly, I'm not pretty confident. I said that I have a little experience in 3D. It's literally just a little experience.

But I mean, I learned the first 3D software I learned was Maya when I was back in college in my senior year. I studied photography and mostly did like video freelance work. I used to operate the camera in like fashion film sets and music videos and stuff like that was like back in 2017 to like almost when it was the COVID era, I started just working freelance. And you might know like back then, it was a kind of like a 3D boom. And I kind of felt the need of I do need to convert to like fully digital instead of like working on set.

So that's how I got onto like Maya. And then I realized Blender is a more attractive tool in terms of like because it's open source. It's easier to learn and it has a huge community. Started using a little Blender.

And then I moved on to like real time game engine software is like Unreal Engine and Unity. But yeah, like in terms of like 3D modeling and like rigging a character and animating myself, I wouldn't say I'm an expert. It's I don't know, like from 0 to 100 is like 20. To 30.

But that's about doing myself. I do have some like base knowledge of what works and what doesn't work. Like technically what's complicated to accomplish. And I'm more interested in like technical operations. How should I say this? Like optimization, render optimizations and more like the logistics stuff. And I think it's because like I was doing freelance work.

Using Unreal Engine and like real time softwares, you really have to care about optimizations in order to make games or else like if you hand it off to your client, it just breaks and it doesn't work on site. So I really had so many times I had a headache because like the client was saying that I designed the whole thing. I designed like the scene and the characters and the movements and everything, but it just breaks on site. Eventually what matters is like optimization.

So that was the biggest thing that I cared about. Okay. Oh, nice. It created the video inside the Codex harness. It's cool. It actually did really pick up the exact spots that I pinpointed. It's interesting.

This is even more epic than I thought. Okay. Cool. Let me bring this into a project now. Kind of forgot where I was working.

Okay. So I'm going to set this project as my I can name it I don't know, CS Game Map. Whisper Flow is stupid. Live stream, CS Game Map, cool. I'm just gonna copy this URL. By the way, I'm gonna copy this URL. If you want to use FLORA MCP, the way is to click this, open in agents.

You'll see this button on the top right. And you'll see all the agents that you can use, like ChatGPT, Cursor, Claude. My personal favorite was Claude, to be honest, but because of Astra, I'm now using ChatGPT. But yeah, if you click ChatGPT, you can open in codex or open in Word. And if you open in codex, it will actually take you to codex with a prompt populated. It will say something give a link and you should connect to your FLORA project. And if you are connected, it will give like a test prompt or something.

But yeah, if you haven't set up MCP, this is the way to do it. Cursor also has one, Claude also has one. And if you're using other, agents I think we also have like Hermes agent and other kinds of like agent softwares, but the most used ones are these three. So yeah, you can try it yourself. I copied the link here. And let's see if it works.

This is the FLORA project that I'm working on. The rendered video in FLORA and render a photorealistic version of it using the initial reference images I provided here. Okay, and just to maximize the context, I'll tag the skill again. I don't know if this is necessary. I'm not doing anything wrong. Cool. And also, just talking about like usage and price and stuff like that.

Probably the thing that most people are curious about. I didn't set it to max, I just set it to high, which is like the medium level. But it still does a pretty good job. Lower the effort, but yeah. And I kind of wrote in my previous post, when I made the dominant sugar factory video, like this regarding the subscription fees to FLORA and codex, or the cost of the generations alone, like on FLORA, I spent, I did the math. And I spent like $10 in total, I think. Because CDN is 2.5 and 1080p, it costs like four something dollars.

Or the reference images, like in this case, like these kind of, like the reference images that contains the actual style, I made it with nanoman and a pro and it costs if I set it to 2K, it's like $0.18. Anyways, what I'm saying is that the whole like image and video generation, it costs like $10 last time when I made the dominant sugar factory video. So that was pretty interesting. Yeah, it was this one.

So blenders free video generation with CS 2.5, it costs $4.3. Image costs like $0.18 each. Yeah, I kind of deleted all the trials and errors, but yeah, just to be transparent, it only took like a couple times to get it right, because I already have a very clear visual reference. And because I also have a clear reference of the motion, the drone motion. The craziest thing is that the video generation only took two times, only generated two times with CS 2.5. And of course, there's a cost of using Astra inside I'm actually using the $100 tier subscription. Right after I installed, I only used it for this project.

I checked the usage before and after. I only used like 3% of the usage. I thought I would use a lot, but that was worth noting, I think there's a $100 version and a $200 version of the max subscription. I controlled my computer and just typed in the URL in a split second and it took me to another project here. That was really fast. Oh, it's asking me a question. What final resolution do I want?

I think 1080p is enough. 1080p high bit rate. Yeah, it recommends 1080p and high. So I'll just say. The FLORA MCP is populating my canvas automatically. We have a question. Could you make the domino and cling and it will cost less? Yes.

It will cost significantly less actually. CS 2.5 was like $4.3. Clings model. Which is clean 03. It's 0.8. It's not even a dollar. Well, actually it might depend on the resolution.

There's no resolution option here. I think it generates like 1080p size. You can upscale later, but. Anyways, even if you upscale in topaz, I'm pretty sure it wouldn't cost over like $4 like clean Pro reference.

OK, this also doesn't have the control of resolution. So I guess if you want like. If your focus is just solely high quality. I guess you should use Seedance.

But there's also Seedance 2.0. Which is more cheaper than Seedance 2.5. Computer is becoming slow. I guess because MCP is controlling.

OK, and made the workflow here. Seedance 2.5. Seedance 2.0. OK, this one, if I set it to 1080p, not 4K. It's cheaper than 4. It's $2.3, so. Yeah, Cling is definitely way cheaper.

But if you're looking for an even cheaper model, Seedance 2.0 references. Technically, Google Gemini is also capable of running this, but I don't really recommend. I'm sorry, Google, but you lose in this game. I tried a couple rounds with comparing Gemini to Minimax H3 to Seedance to Cling-03. I can confidently say that Gemini didn't really do the job. Maybe it needs more reference images or I don't know, but. Seedance 2.5 was exceptionally better at this.

And especially. Like these kind of footages, it's very complex. As you see, like it starts from outside of the building and then go inside. It's a really long trajectory and yeah. These kind of complex. Type of camera movements, you should just definitely use Seedance, even though if it's expensive, because eventually if you use other like low tier models, you'll just end up wasting a bunch of generations over trials and errors.

So just might just use Seedance 2.5 and achieve it in a less number of generations okay let me see how codex is doing okay if finished wait oh okay so it stopped right before generating it's the maximum total number I see okay so okay this skill is smart enough to not spam your credits it just stopped right before in order to let me authorize it yeah just for stress testing in this session I'm gonna say I'm just gonna go full MCP mode I'll say like let me see if codex can evaluate the results maybe like try until you think the result is good enough generate maximum five times I need to let it give feedback and fix things by itself wondering what I should say generate maximum five times and evaluate the whole process and make necessary fixes cool see how it goes I'll move blender over here okay yeah this is kind of like overview of what happens when you use MCP it just populates the whole workflow by itself connects all the nodes as if somebody is working next to you kind of curious what's going on X I'll see it by myself I just hit the window here okay something's happening sorry what are you doing in the MCP right now so right now I'm letting MCP to dry I'm sorry right now I'm letting codex to drive flora because now we have the blender rendered scene here huh fail to load MCP anyways what I'm doing right now is letting astra drive the floral project based on the generation the renders and made in blender and oh this is pretty shitty pretty much do a good job here but let me see how it goes actually feel like I need to jump in this it's taking too long I'll just copy and paste this workflow over here oh copy paste okay and I'm gonna try building in parallel with the MCP let me actually see how the workflow is going we have the original image here the blender rendered scene once to two screenshots okay so this image-- OK. May I try the same prompt one more time? And see how it does. OK. Actually, because we don't have a bunch of time, it wants to record the screen and audio. Go for it. Chat GBT.

Oh, shit. It should quit and reopen. OK. It should technically pick up where it slept off at. I'll just say the program died. Pick up from the last part.

OK. This one's way better. Actually, let me add some instructions, because I think the background here is very empty. Fill in the background with buildings in the same style. Let me make another one here using this image, 16 by 9. Use the same prompt. Replace the white box materials.

This prompt is very specific. I'm going to try my own prompt. Scene referring to-- I'm going to mention this node. Desert street scene. Stylize the 3D preview scene.

OK. Keep the-- let me see how it does. I kind of like to write my own prompts in a very simple way, instead of providing a whole paragraph here. Sometimes you need AI to write it for you to get it more accurate. Oh, MCP was faster than me.

This is not bad. Oh, this is not bad either. I guess we're almost the same. OK. Not this image. I should feed in this image. And not this one.

Switch the orders, because this should be a first frame and this should be a last frame. Oh, shit. No. Oh, wait. Oh, this was an image node. Never mind then. Got to do a video node and plug this one in.

We have the motion reference video here and the first photo-realistic scene and the last photo-realistic scene. One video, two images. Let me try. You look more correct. Thank you, that's good to know. I don't really have the eye to judge whether this CS map is correct or not because I don't remember a thing about playing CS, but Okay, let me well MCP is also running some evaluations.

So I hope it pulls some context of the real Counter-Strike game But yeah, it's almost like a race now. Let me generate with first of all CS references CS 2.0 excited to high What prompt should I put on Drone POV footage Refer to the camera Movement of the 3d previous video Thanks for covering this DJ never thought to try it with codex course, I always enjoy trying new things and Letting people know what's good and what's not good I'll try this one with CS 2.0 Because this is a live stream also Demonstrate other models as well. Wait, It's most important thing. This is eight seconds.

So I need to match the timings. Oh fuck already did five seconds It should be eight seconds because it should technically match the duration of this video Eight seconds on I'll generate this again. I feel so bad at wasting my credits, but Yeah CS 2.5 this one duration auto Okay, we have an auto here. So I hope it picks up eight seconds CS 2.5 Third model I'm gonna try h3 I'm gonna skip Gemini because it looks too shit all the time Unfortunately h3 doesn't go up to 10 AP It's only 760 AP. It's a very specific number It has this interesting parameter called prompt expansion mode I'm not sure what this is.

So I'm just gonna leave it as default Oh, we had a question are you going to keep the video game aesthetics or make it photorealistic I Don't know I'll try this one first and let's see how it goes. Oh I have an idea for the live stream maybe For the next generation if this goes well, I can try the same thing with a different texture Like I mean like a whole different, like a night scene or something H3 is so fast. Yes. I agree. It's so fast. I don't know how they do that, but it's phenomenal Okay, MCP is rushing I wonder why the generations keep failing. Oh It's done Opening and ending images inputs high workflow remaining your project you always should feel it also have a log, okay It gave me this log Reference to edit a woman.

I wonder where the output video is Your floor project. This is my project. Oh Shit, okay. I can do the same thing here. Oh Wait, I wonder where the video is you didn't generate a video Okay, I guess I can just run it on my own CDS 2.0, oh, wait, it kind of did the same thing. That's interesting.

So, MCP also is trying both CDS 2.0 and CDS 2.5, which is kind of interesting. I didn't ask for running for all different models, but maybe that was part of the evaluation process. And we can see that it's pretty smart enough to select duration to eight seconds because the video was eight seconds. It didn't leave it as default five seconds. That's pretty nice. It's set to high bitrate, as I asked for. Let me generate both at the same time.

Oh, it actually did try generating, but it failed. I wonder why. Let me check my generations. Of course, H3 is the fastest. It's not bad. The only thing is that these were blue, and they turned white. And I'm pretty sure it's because the images were like this.

Oh. I just noticed I missed one step. Okay, so when I'm creating this image, because this is in a sequence from this image, I technically should plug this image in as well. Actually, yeah, I don't need desert street scene because this was the initial reference.

This was enough for a reference for this image, it doesn't need the initial image anymore. I can just break the connection here. And for this image, all I need is the context of the establishing shot. So referring to replace white box. It's a weird name, but the name of the node is replace the white box.

So I'm just going to do this. Let me see how it goes. Yeah, other than that, it looks pretty promising. But I actually just realized these trees are still low poly. They should be realistic looking. Oh, they are actually realistic looking trees, but yeah, maybe H3 just didn't pick it up, I think. That's funny.

Okay. No, this tree is still some freaking Lego block. And these pillars are still different. Let me try one more time. And actually, I can add a line saying, stay accurate to the previous scene. Keep the accurate structures and colors based on the previous scene. Cool.

Yeah. So because this image was wrong, that means these two generations are also going to come out wrong. So eventually I wasted three generations, but I learned my lesson. Okay. I really don't like this tree right here.

Okay. What I'm going to do is generate from this image because this one almost got it, right? Except this part here. What I'm going to do is this kind of a workflow that I enjoy doing for very granular changes. I'm going to draw on this image and let me double click and select the brush. Type to red, and I'm going to select this part. And make this image.

So now I have an image that I have a circle on, plug this in. And yeah, what I'm doing is I plugged in the original image and the sketched image. The reason I'm doing this is because I only want to change this part and leave everything else the same. I can just say, replace the toy tree that is isolated in a red circle into a real tree. Refer to the tree design on the right side. I wanted to aspire enough to understand this. And let me check this footage.

OK. There's no way it's going to get it right because the previous image was wrong. I guess it's kind of hard to get it right. OK. I think this one might have been better. Looks a little more closer.

OK. Let me actually use this one. Change the prompt. Apologies if I'm getting too much into the weeds, but this is what I have to do to get this right. Replace the toy tree. That is isolated in the red circle into a real palm tree. Actually, I think an easier way would be just removing the tree in the isolated red circle.

Let me see how these turned out. OK. That's a lot of visual errors. Let's see how cdns 2.5 does it.

OK. Maybe I picked a very complex task. Well, OK. I guess the last image it created, it is kind of accurate. Let me try this one. Copy this node. Disconnect these images and plug this one in.

Plug this one in. Do the same thing. Seedance 2.5 auto. The same thing here. Copy this node. Seedance 2.0. References. 8 seconds. 1080p.

High. OK. I'll do the same thing for minimax. 8 seconds. Yup.

OK. Let's see how it goes. I think this one will come out a little better. It will come out a little better. Because technically I have a more accurate image. Actually I can check side by side. Oh, by the way, we have this amazing new feature where you can select two images and compare side by side.

In case you missed that, I'll show it again. I don't think a lot of people are watching right now, but I'll still show. You select one image. You select another image. I'm pressing down command. And on top of the toolbar, there's going to be compare image.

And then the great thing about this is the zoom in is syncing on the other side as well. So you can zoom into the same exact part on the image. Yeah, let me see this part, how accurate it is. So technically I can compare how accurate it is looking at the wide shot and the close up shot. I guess it is somewhat correct, because I see the fence here. I see another fence here. There should be a tree here, but I see the shadow here.

So I guess it is valid. And I see these little blue tiles on the wall. It's not 100% accurate, but it'll work. Let me see what minimax did. It still had these toy trees, which is kind of annoying, but still works. Well, other than the toy trees, I think it's not bad. Yeah, there's no glitching or visual errors here.

Let me know if you guys spot anything. It seems not bad to me. Hello, my fellow vibe coders. Welcome. I love vibe coding. Why all of Reddit anti-AI even in the AI side? I'm not sure what you're referring to.

I haven't seen Reddit anti-AI. Well, today I still got a lot to cook. So I'm just going to focus on finishing this project. Let me see this one. C9s 2.5 result. Oh, that's actually not that bad. Oh, it did kind of show a little glitch on the end.

But OK, I guess I can plug this into Timeline Editor. There's a hack here. I don't always need to use the full eight seconds. You removed the tree and one of your generations.

So in the video will show the tree being there and then disappearing. Yeah, that's true. But I actually just leaned on the fact that it has a shadow here. So I hoped it can get closer. Nothing. A quick clip cut can't fix. Yeah, that's true.

I mean, the goal here is not to make a whole CS game trailer or something. I can just cut out the last part where it glitches. Let me see where it happens. Here it is. And because this footage is like eight seconds long, it's pretty lengthy. I can cut it out. You'll get way better generations if you generate another image in between.

Yeah, that's true. Right. I totally agree with that. I was just being lazy.

But yeah, like this camera movement is kind of complex. So I believe I need another frame in the middle in order to make it more accurate. Okay, I'll stop being fixated on this. We'll move on because today the focus is not finishing a video. It's to show the power of Astro.

So okay, I think sorry about that. I think this is enough for showing the proof of concept of blender previous to FLORA students video. Then I'll move on. Just for efficiency. I'll make a new channel. I mean a new chat thread making say copy and paste this project here.

And then I'll say maybe use the videos that I marked read on the canvas. No, my bad. I'm trying to make a game here. So okay, I guess I can say using the best images here and this video here. Oh, actually, I actually do have a blender project.

So well, this next step might not be like super FLORA related, but I'm just going to try making a web based game that I can like play as like a first person POV and walking around in this CS map. Yeah. Make a game that's playable on the browser based on the blender scene and the video that I created on this project. I'll see if I can find the best video. Not this one. Yeah, I forgot to tell it 3gs. Make a 3gs game.

Tell it to not be stupidly recursive agents read through about it. That's good to know. I'll try to check out the Reddit thread kind of curious what's going on there. Okay, CS 2.5. Well, I guess it's inevitable to have these toy like looking trees because the images are already like that.

But yeah, I don't care. I just want to make a game out of this. And honestly, for making the game, I just need to look at how the world looks like. So I guess the video doesn't really have to be perfect. I marked the best generation in red on the canvas. Make a 3JS game that's playable. I'll say first person game that's playable on the browser based on Blender.

Actually, okay. I don't know if it will have time, but I also wanted to add a character to make it playable in like first person shooting game type stuff. Okay, let me actually send this request first and then on the side, I'll let it make some characters. Oh, I actually have an idea. Wait, I'm trying to think how to creatively integrate a character inside this game. Oh, we have a question. Did Astra make scenes in Blender from your screenshots?

Yeah, it actually did. I'll show the process here. I dropped in a single screenshot from Google images, which was the real CS game map and then based on that it created this scene inside Blender and it came all the way to creating a low res render 720p and 24 frames per second. I used that rendered video, dropped it inside FLORA and that's how I got this far.

Now I'm going to make a game out of that. I found the original Desert Quarter Blender map, including the uphill street floor or tunnel. Okay. Alright, let's see how it does. It has two windows popped up here. It's pretty cool. In the meantime.

So we got like 20 minutes left. I wonder what I should do either. Create a character that's playable inside this game or make another scene, another drone POV with Blender and make another video. Okay. It says I can read the project's videos, but connector almost counts color markings. It's interesting. When I was using the CLI, it was able to read it.

This top app can't. What's the title? Continually Blender map. The title. I'll just change the title of this node. Selected video. Selected.

Whisper flow is stupid. Maybe the mic is designated to a different mic here. Okay. So does Astra have access to your whole computer and everything on it? You make a YOLO mode. Yes. I think you're referring to full axis as YOLO mode, but yeah, like if, as far as I understand, if I give full axis inside this session, it just has all the access to instead of me hitting approval every time, all the access to the steps that it decides.

I think there's also, inside the settings, the computer kind of froze here. Oh, here it is. Yeah, like inside the settings, I think there's, you can also change it here. Right now it's full axis turned on.

So yeah, technically it does have access to almost everything, like deleting files, modifying files. Let me just quickly check if I have enough usage left. Okay, 41%, it's enough. I started with like 56 and now it's 41%. I sometimes add to the chat that I'm using Whisper and tell it just so it sounds crazy. That's funny, but it sounds helpful. Oh, technically LLMs should be, able to understand giver-ish and like make auto corrections.

So, well, I guess that does also help because it might have context of Whisper flow. So that's helpful to know. It's writing Python exporter. Actually, one of the reasons I prefer to CLI is that in the desktop app, it just shows like a one-liner. I mean, you can drop down and like see what's going on here, but the CLI just like transparently shows everything.

So just for me personally, I kind of like that. Oh, cool. So I'll be there. Okay, let me try the game. It says desert quarter. A fake counter-strike game, desert quarter, signal run, begin signal run free export.

Okay, these controls don't work right now. Oh, it just saw the video. Oh, I guess it just made the landing page first. Sound quality high.

Okay. Oh, I just remembered what I was going to try. So because I made this type of texture, I'm going to re-texture this into a different like nighttime. Or actually, no, on second thought, that's technically not using Astro.

So I'll think about a different one. I think it's best to use Astro at this point to create some characters. I'll create the character assets inside FLORA. So let me just say I want to make characters in this scene. Pull from the internet. Pull inspo from the internet. And make a character sheet.

Okay, let me do that. Actually never tried making a character sheet in Astro, but see how it goes. Okay, still not playable. All right. Actually, in the meantime, I think I know what to do. Until this time, I'll hand off this skill that makes a Blender scene and takes it all the way to FLORA to make a video. I haven't covered the gaming part here yet, but if you still want to create a FPV drone style video, this skill should be enough.

I gotta find where this skill is. Oh, here, this Blender previous with FLORA. Okay. I'll put this in my drive. Cool. Skills.

Okay. Oh, shit. Seems like it's ready. Blender previous and FLORA. Share words.

Okay. I'll drop it in the chat. Let me know if this works for everybody. Cool. Drag to look and use arrow keys.

Okay. I'm playing Counter Strike in my browser. I can also jump around. Oh, shit. What is this?

Okay. Drag to look. Use the arrow keys. Okay. I don't like it. I don't like that it's dragging. I'll make some updates.

Okay. All right. I'll give some feedback. Oh, it's playing by itself. Maybe it's evaluating right now. Wait. Let me check how the characters are working.

Character sheets. I'll say yes for now. Oh, wait. It's generating inside chat GPT. No, I want it. I want it inside FLORA.

Okay. Stop this. Okay. Since we have 10 minutes, I'll just give up on the characters. I'll try to pick it up after this live stream and maybe post about it later. Let me just get this game done inside for us. Oh, it also has a timer here, okay.

I guess this is my gun. It's cute. All right, I'll just send this command because we don't have time. Oh, I kind of understand why it's dragging. Wait, I'm still thinking like how I should replicate the shooting and the moving around interaction separately. Maybe if, maybe I need a mouse. And have the viewfinder follow the cursor, but when I click, it should shoot.

Okay, I'll make the view turn as you move the mouse without holding the button. Okay. I'll send some more feedback. Make the click, left click interaction. Shoot.

Okay, well, because we only have like eight minutes left. Thanks for joining guys who are watching so far. It was such short notice today, but yeah, I'm glad that people have joined and interacted with me in the chat. Some things didn't work out as I thought, but still, I think it was a fun session. I'll try to do more of these. And if you're still sticking around, please let me know in the chat what kind of what kind of live streams you guys want to see soon.

This is for hours. I'm glad to know. I swear things were better when I was trying on my own, but as soon as I go live, things start breaking. Honestly, the Domino sugar factory building, I finished everything from feeding in the reference images to getting a perfect video, a perfect sedance video under, like under two hours. That was my first try. Actually, the second try was the Hermes Tokyo Hermes store that I just showed previously today. That one was also fast.

I mean, this one was pretty easy. It was just very basic Dolly. And I guess this one was pretty complicated because I started here and I ended up super close up. And these trees are just toys. And it also has texts here. Like full of complicated assets. It makes sense that things can break here.

But anyways, okay. Codex just finished cooking. I can now move around. Oh, cool. Oh, but the problem is if my cursor escapes the window, it just. Oh, it's kind of jank, but still I can walk around. All right.

I'll send some more. Oh, if I get out of focus, it just automatically pauses. That's pretty smart. Make the click interaction. Shoot the gun. And what else was going to say? Oh, prevent the cursor.

From escaping the window. Set ESC or. Releasing focus. Yeah, and these are full screen borderless option. Well. I might work on this a little more and deploy it soon. With playable characters and shooting.

But yeah, let's see how far it goes. We got like five minutes left. And I still got a decent amount of usage left, I think. 38%. Yeah, I still need to cook on some other stuff with Astra, so I gotta save a little bit. I wonder if anybody like directly compared Astra to I don't know, like Fable 5.1 doing the same things because... Yeah. I haven't tried these same projects with Fable.

I do also have a cloud subscription and my favorite model used to be Opus 5. I barely use it nowadays. I also recently onboarded to Grogbot. And it seems like it's a different type of agents and still figuring out what the best tasks to let Grogbot do are.

But yeah, it seems pretty interesting. I'm kind of liking it so far. I'm kind of telling Grogbot to do more like organization and operations type projects I was kind of scared to give my own credentials to Grogbot because in case you don't know, Grogbot has a virtual environment inside. It has its own Linux computer, so you can technically provide your own credentials and let it just use its own computer. I'm still exploring though.

Okay, it's going to add gun fires and lock the mouse. Cool. Can't wait to play this game. Oh, the arrow keys actually... Are for navigating. Oh, that's cool.

So I can use WASD to walk around and the arrow key to turn the head around. I wonder what this... Oh, hold to restore relay. Forward passage, connection restored. I don't know what this is actually. Yeah, for those of you who are well versed in Counter Strike, let me know what this is. Oh, maybe this is like the objective of playing the game, like how to find all these restore points.

I don't know. Okay, I gotta wait till this updates. I just can't play it like this. The mouse keeps escaping the window. Click fire now has visible sign.

Okay. Switch to free explorer. Oh, it also has minimap here. It's dope.

Here's a performance mode. Okay. Maybe it's like rendering optimization? Not sure.

Okay, we got a minute left actually. Let me see how fast enough is Codex. Today's artifacts, we have our video that we made, which is not that great. And also we have a game. And to be honest, the game was pretty impressive than I thought.

But I'll have to see how far it goes after all my requests. Well anyways, thanks for watching guys Really appreciate you five people who are still sticking around Yeah, I Don't mind how many people are watching my Improvising livestream, I just had fun Up I can play around with Astra a little more. Yeah, please Keep sharing stuff Or like trying flora and trying cool stuff with Astra and stuff like that on socials and Tag us if you want. We like seeing people cook with flora So, okay. Luckily We're one man past but I think Astra made some updates here Resume exploration Let me refresh this Left-click fire Huh Mouse capture unavailable. That's weird.

Okay. I'll solve this later Anyways, thanks for watching guys. This was kind of like a very experimental stream I think I use flora for only like half of it, but still it was fun Anyways, this Astra wave is just highlighting the model Astra and using like other AI tools for I don't know like orchestrating and Like acid creation and stuff like that but I kind of did enjoy the process of using flora MCP as well because I mean as you guys know like there's there are a ton of AI tools out there, but I think flora is still good at like having the building blocks to build your own workflow and if things don't work like you can quickly make fixes like Having a draw node and like isolating parts and in paint that and I could have went further of like putting this into a layer editor and cropping this image Part of that image and compositing with the other one that has like the right free image But I mean that's too much of diving into the weeds. I didn't want to do that So but anyways, today was a good proof of concept that we can do this We can make a game with just chat GPT and blender and flora.

So That was cool. I hope you guys get to use the skill that I made User was interesting. Yeah, I actually never I should never did a race with MCP but yeah for the first time I Felt like I was kind of threatened by the speed of MCP. I bet it's gonna get better but yeah, Thanks for watching guys. See you next time. I can't guarantee when the next time would be but Yeah, I promise I'm gonna Be alive again and keep this up Yeah next time I'll try to send out a post or notice Before I do a livestream today was like almost like less than an hour before but yeah, I'm grateful that you guys joined Thank you. Peace You you