Build your own infinite AI video stream (no coding)

0:00You may have caught wind of Minimax H3 Max and Max Turbo recently and it is changing the game for AI video generation forever.

0:07Literally forever, nothing will ever be the same again because you can now generate videos faster than you can watch them.

0:14We see builders like Levels.io creating infinite slop.

0:16You see this Diverge project, which is basically a choose-your-own-adventure that generates the video to watch and participate in the story in real time.

0:28So there's a lot of really cool and inspiring projects out there.

0:31And in this video, I'm going to show you how you can run infinite video generation loops

0:35on your computer using the open source Venice video harness.

0:38And don't worry, you don't need to be a technical person to figure this out.

0:41It is easy.

0:42So you're going to need a Venice API key.

0:44You can generate this on the Venice platform.

0:46So what are we going to do here?

0:47Well, find the link below to the video harness and here, copy it.

0:52And you're going to open up OpenCode and paste it in.

0:54If you've never used OpenCode before, this is an open source AI harness. But don't worry,

0:59it has an app for your computer. You can just download it, install it, and give it your Venice

1:04API key here in the settings under providers. And if I'm moving too fast or you're confused,

1:09you can just check out the video dedicated to setting up OpenCode with Venice here on this

1:14YouTube channel, and then come back here. It's actually really simple. But once you have it done,

1:18we're ready to launch the video harness and start our stream. So first I give it the URL to the

1:23GitHub repo so it knows where to find the actual software, the agentic workflow, the harness. Then

1:28I'm saying let's use the latest version of this Venice video harness to create a stream of a 90s

1:33sitcom in the style of Friends or Seinfeld with a laugh track and a cozy comfortable setting.

1:37Now, I'm saying stream here in capital letters because I could say short film, I could say

1:43mini series, I could say almost any format I want, and the harness will basically follow a different

1:50path for production. With stream, we are looking at quick and cheap generations with no reference

1:58images. That's how these streams currently work if you want to get the speed of them generating

2:04faster than it takes them to actually play back. Finally, I'm saying the twist, the main character's

2:09arch nemesis is a robot waitress at a cafe. So that's my basic prompt here. Now I'm using GLM

2:145.3 flash. This is an insanely cheap but powerful open weight model. It is not the best of the best

2:21model. If you want to use the best of the best, it's going to cost you a little more. You could

2:25use Kimmy K3 Fast, which is best of the best open weight models, in my opinion, at the current time

2:30of me recording this. Or you could straight up go Fable 5.1. You could even go Astra if you want.

2:37A good rule of thumb is just to assume that the higher quality model you choose, the higher quality

2:41result you will get. So the agent's going to work for a little while, then it's going to ask you a

2:45question or two. First question here, which writer model should author the stream's beat? So basically

2:50the agent here is going to do a lot of the work, the model we already selected. However, once we're

2:55in stream mode and it reaches the end of the stream, it's going to need to create more prompts

2:59if you're going to keep the stream going. By default, it's going to prepare 15 prompts for

3:03you from the very get-go. But if it needs more, you need to select which model you want to keep

3:07writing. So we have DeepSeq v4 Flash Fast here because it's the fastest and ideally it helps the

3:13stream keep moving fast as possible. So I'll choose DeepSeq there. Next question, what budget cap should

3:19the stream run with? I'll set two dollars at the default and ultimately that's going to create a

3:24long time. I mean this H3 Max and Max Turbo are very affordable. So now it's going to keep going

3:30and in just a minute the user interface is going to show up on my browser and then we can see the

3:34stream in action. Here it is. It's finished. And we get a link to a local host app for the project

3:41called the Decaf Menace. So before we watch it, let's just check out what this whole thing looks

3:47like, because it can look like a lot's going on. First, let's start at the top. Here we see all

3:51our different projects. This is the Decaf Menace, the main project that we're working on right now.

3:56And over here we see in the treatment tab, we're in a 1996 multi-camera sitcom like Friends or

4:02Seinfeld and Greenwich Village Coffeehouse, New York City, blah, blah, blah. And we have a bit of

4:09a workshop here. Now, stream mode basically fast forwards through a lot of this pre-production

4:15stuff that the harness does allow you to do if you want to make a more complex project,

4:20if you want to be more refined, you want a storyboard, for example. Here's a storyboard.

4:24We didn't render anything because this is just an infinite loop. You're casting locations and

4:28reference images, which we actually don't need for a stream, but depending on how you're going to use

4:35the stream feature, this might come in handy. And we'll get back to that in a second. You can have

4:40the Venice Video Hornets edit everything for you, generate music for you, export a timeline for you

4:46to edit it yourself. Finally, here are settings for default intelligence, video generation, image

4:51generation. Now, in this particular project, we are using Minimax H3 Max, not Max Turbo,

5:01because the quality is just not as good, and I think we need to keep the quality better.

5:06And you'll even notice that if there's AI artifacts you don't like, like the quality,

5:10it'll probably be fixed by jumping to a higher resolution. And then meanwhile, over here,

5:15you have the writer. So as it is right now, the harness will generate, I think, 15 prompts.

5:22for the story so far. So once you start the stream, it'll go for 15 prompts. It won't need to write

5:27any more prompts. But if it's still going after that, if it keeps going, then you're going to want

5:31to choose which model is going to write the prompts. So I have the fastest model available.

5:36It's going to be about four seconds per beat, but it'll write the prompt before sending it to the

5:41video generation model. So you have abilities here to make it as fast as possible or to make it as

5:48good as possible. Like here, we're just choosing CDNs 2.5 for example. It's going to take way

5:53longer, but we can go to a higher resolution here and take all these files that we create and

5:59actually make something out of it. So this can be a good tool for not only people trying to just

6:04watch an infinite video for fun, but also for creators. Let's click start stream and it is going

6:11to start generating the next file as I watch. There's our title screen. There's some artifact

6:23there for the model. So now we have the second part of the stream. I pressed pause while that

6:33generated because I had it set to seed dance. Got a little hallucination there.

6:44Char one, I ordered a quad shot espresso, not a decaf apology with a face on it.

6:50Okay. So now I'll continue the stream, but before I do that, I'm going to make sure I

6:56unselected seed dance because that took a while. So I'm going to do H3 max, which is the default

7:01at 480p, and we'll continue the stream right now. So that'll take a minute, hopefully no more than

7:08a minute. And as that runs, I'll show you the rest, which here we see as the story continues,

7:14we see each prompt, and we can see that it used text to video, the first prompt, just a text prompt,

7:20and then it used the last frame, and it will continue to use the last frame of the previous

7:25video as the first frame of the next video to chain these videos together. We can click the

7:30full prompt button to see the whole JSON of the prompt.

7:34And in case we want to use that elsewhere,

7:36and we can copy it.

7:39Or we can straight up download the JSON or markdown file

7:43of the entire story so far.

7:45So you'll start to see that the real point

7:47of the Venice Video Harness is just to assist you

7:49in working on video projects.

7:51This stream tab is just a brand new feature available.

7:56All right, now I got the next one.

7:57Taste it.

8:02look at this she's not even a real barista she's just a chrome mug slinger with a grudge

8:06i can taste it this lata has the absolute audacity of a passive aggressive ex

8:12and you can see that it's keep it keeps going it had to do a reset

8:15and it just keeps going now so we can watch it

8:24did she did she seriously just program a laugh track into the espresso machine

8:28That's not a robot, that's a high-tech heckler with a caffeine button!

8:33Cold-blooded.

8:34You know what? New rule.

8:36From now on, I'm bringing my own thermos.

8:38Because apparently, I need the pure, uncut joy of a real caffeine hit to survive this conversation!

8:54Threat noted.

8:55But your thermos is still in your apartment.

8:57And this latte is already in my processor.

8:59Well, that is logically indisputable, yet deeply hurtful.

9:21Ooh, you think you've got this cup locked down in your processor?

9:25please i've just cracked the deployment code for the saucer okay so we can see now i'm going to

9:30pause the stream we can see now that it's just going to keep going we can obviously also see

9:35that there is some problems with these generations and the truth is that just happens sometimes with

9:41these mini max models like that looks clean but then as we see earlier there's just some some ai

9:48glitchy there so that just comes from the model provider even behind venice the provider serving

9:53the GPUs. And if you're not liking it, obviously just change to a different model and you won't

9:58have to deal with that. Just know that if you do change to a different model, kind of like here,

10:03it reset completely. So it started fresh. Whereas here we were using C-Dance and it

10:10obviously didn't create those artifacts. It just kind of hallucinated some funny physics here.

10:16But the quality is good.

10:20So at that point now, you can just keep going forever.

10:23I can download the full markdown prompt.

10:25I can copy just a single prompt.

10:28I could bring it over to the Venice Studio.

10:30So let's look at this one.

10:32We'll copy it, paste it here, and we could generate it here to fine tune it.

10:37Same thing, choose the model.

10:39What the Venice Video Harness helps with is just creating these shots in bulk.

10:45And in this case, you can watch it as you go.

10:47You can brainstorm.

10:49And of course, if you have ideas for how this could be improved, go ahead and leave a comment.

10:54And we could perhaps add these features.

10:56Remember, this all runs locally on your computer.

10:58With a single API key, you can set a budget, you can tell it not to go past the budget,

11:03and of course, you can have some fun with all the other tools that the harness offers you.

11:07But I hope for now that's a good introduction to the Infinite Stream feature, the Venice Video Harness.

11:11And stay tuned for the new Venice Learn platform where we'll share more tutorials like this,

11:17and you can get help for whatever project you're trying to build with Venice.

11:20Happy creating.