Generative AI Videos using Bedrock, Polly, iMovie - FakeTube #2

We are building FakeTube - a YouTube clone using cloud-native and web technologies. In the last episode we defined initial project requirements. We are almost ready for the development, but before we jump into writing code, we need one more thing - a video content. Due to potential copyright issues, using existing YouTube videos is not an option. Instead, we can consider using stock video footage from services such as Shutterstock or creating our own content with a mobile phone. There is however a better way in my opinion. We are living in a Generative AI era, so why not just use some foundation models to generate not only ideas for the videos, but also the videos themselves. Content Requirements FakeTube is all about videos and we need at least a few of them to implement and test home page features such as responsive grid layout, infinite scroll or inline video player. How many? For the infinite scroll we will need a lot of videos. This is because YouTube loads videos in batches (which is actually a page size) of 24. So we will need more than this to accurately illustrate the feature. Also, having a home page full of videos, especially on the larger screens will look much better, so let's assume our target will be to generate 32 videos (which is BTW also a perfect number from the computer science perspective: 25). What topics? Should we generate some random videos or do we have some specific topics or categories in mind? It turns out that YouTube can help us with that. When you upload a video using YouTube Studio, in the "Details" step there is a hidden "Paid promotion, tags, subtitles, and more" section, which you can expand by clicking the "Show more" button. One of the subsections there is "Category" ("Add your video to a category so viewers can find it more easily"). There are 15 categories in total. They are very generic, but it will help us to create more structured and consistent video content. For now let's select 8 categories and we will generate 4 videos per category, which will be 32 videos in total: Autos & Vehicles Education Gaming Howto & Style Nonprofits & Activism Pets & Animals Sports Travel & Events Generative AI With a clearer understanding of our desired content, the next crucial step is to identify a Generative AI service and foundation model capable of generating a video script and footage. I personally have a lot of experience with AWS (Amazon Web Services), so it's a natural choice to use it for this project. At some point I really would like compare it with other cloud providers like Microsoft Azure or Google Cloud or even independent services, but for now we will use AWS as a foundation and a baseline for FakeTube. Amazon Bedrock Luckily for us, we don't have to spend a lot of time on research. Amazon Bedrock is the service dedicated for the generative AI applications in AWS. It's one of the fastest growing AWS services, which gives you access to a wide range of Foundation Models (FM) from leading AI companies like AI21 Labs, Cohere, Stability AI, Anthropic, Meta, Mistral AI, DeepSeek and of course Amazon themselves. Amazon Nova One of the models or actually a family of models which can be a perfect fit for our use case is Amazon Nova, because as stated in the documentation, it includes: creative content generation models that accept text and image inputs and produce image or video outputs. They are designed to deliver customizable high-quality images and videos for visual content generation. We will be interested in 3 models from the Amazon Nova family: Amazon Nova Reel state-of-the-art video generation model that allows customers to easily create high quality video from text and images. Amazon Nova Canvas state-of-the-art image generation model that creates professional-grade images from text or images provided in prompts. Amazon Nova Micro text-only model that delivers the lowest latency responses at very low cost. It is highly performant at language understanding, translation, reasoning, code completion, brainstorming, and mathematical problem-solving. Limitations There are currently some important limitations of the Amazon Nova Reel, which can be difficult to accept or affect our content generation workflow: Video duration is limited Amazon Nova Reel serves a wide range of potential application domains and offers the following core capabilities: Generate videos of up to 6 seconds given a text prompt. Generate videos of up to 6 seconds given a reference image and a text prompt. There is no audio Modalities Amazon Nova Reel does not currently support audio or 3D content. The first limitation is not a big deal at the moment. We can accept that we will have short video examples. The second one is more problematic. One of the features of the video player is the "Unmute/mute button", which will be kind of useless when a video doesn

Mar 31, 2025 - 17:13

Generative AI Videos using Bedrock, Polly, iMovie - FakeTube #2

We are building FakeTube - a YouTube clone using cloud-native and web technologies. In the last episode we defined initial project requirements. We are almost ready for the development, but before we jump into writing code, we need one more thing - a video content.

Due to potential copyright issues, using existing YouTube videos is not an option. Instead, we can consider using stock video footage from services such as Shutterstock or creating our own content with a mobile phone.

There is however a better way in my opinion. We are living in a Generative AI era, so why not just use some foundation models to generate not only ideas for the videos, but also the videos themselves.

Content Requirements

FakeTube is all about videos and we need at least a few of them to implement and test home page features such as responsive grid layout, infinite scroll or inline video player.

How many?

For the infinite scroll we will need a lot of videos. This is because YouTube loads videos in batches (which is actually a page size) of 24. So we will need more than this to accurately illustrate the feature.

Also, having a home page full of videos, especially on the larger screens will look much better, so let's assume our target will be to generate 32 videos (which is BTW also a perfect number from the computer science perspective: 2⁵).

What topics?

Should we generate some random videos or do we have some specific topics or categories in mind?

It turns out that YouTube can help us with that. When you upload a video using YouTube Studio, in the "Details" step there is a hidden "Paid promotion, tags, subtitles, and more" section, which you can expand by clicking the "Show more" button.

One of the subsections there is "Category" ("Add your video to a category so viewers can find it more easily").

There are 15 categories in total. They are very generic, but it will help us to create more structured and consistent video content.

For now let's select 8 categories and we will generate 4 videos per category, which will be 32 videos in total:

Autos & Vehicles
Education
Gaming
Howto & Style
Nonprofits & Activism
Pets & Animals
Sports
Travel & Events

Generative AI

With a clearer understanding of our desired content, the next crucial step is to identify a Generative AI service and foundation model capable of generating a video script and footage.

I personally have a lot of experience with AWS (Amazon Web Services), so it's a natural choice to use it for this project. At some point I really would like compare it with other cloud providers like Microsoft Azure or Google Cloud or even independent services, but for now we will use AWS as a foundation and a baseline for FakeTube.

Amazon Bedrock

Luckily for us, we don't have to spend a lot of time on research.

Amazon Bedrock is the service dedicated for the generative AI applications in AWS.

It's one of the fastest growing AWS services, which gives you access to a wide range of Foundation Models (FM) from leading AI companies like AI21 Labs, Cohere, Stability AI, Anthropic, Meta, Mistral AI, DeepSeek and of course Amazon themselves.

Amazon Nova

One of the models or actually a family of models which can be a perfect fit for our use case is Amazon Nova, because as stated in the documentation, it includes:

creative content generation models that accept text and image inputs and produce image or video outputs. They are designed to deliver customizable high-quality images and videos for visual content generation.

We will be interested in 3 models from the Amazon Nova family:

Amazon Nova Reel

state-of-the-art video generation model that allows customers to easily create high quality video from text and images.

Amazon Nova Canvas

state-of-the-art image generation model that creates professional-grade images from text or images provided in prompts.

Amazon Nova Micro

text-only model that delivers the lowest latency responses at very low cost. It is highly performant at language understanding, translation, reasoning, code completion, brainstorming, and mathematical problem-solving.

Limitations

There are currently some important limitations of the Amazon Nova Reel, which can be difficult to accept or affect our content generation workflow:

Video duration is limited

Amazon Nova Reel serves a wide range of potential application domains and offers the following core capabilities:

Generate videos of up to 6 seconds given a text prompt.

Generate videos of up to 6 seconds given a reference image and a text prompt.

There is no audio

Modalities

Amazon Nova Reel does not currently support audio or 3D content.

The first limitation is not a big deal at the moment. We can accept that we will have short video examples.

The second one is more problematic. One of the features of the video player is the "Unmute/mute button", which will be kind of useless when a video doesn't have an audio track.

Solution would be to generate audio using some other foundation model or service and merge it with the video using for example Apple iMovie, which I already use for video editing.

There are two ideas about an audio track

Generate background music
Generate voice-over narration

Background music is a very nice option, but unfortunately I didn't find a foundation model in Amazon Bedrock which is capable of doing that. There is a dedicated service for that AWS DeepComposer, but unfortunately it's deprecated

After careful consideration, we have made the decision to end support for AWS DeepComposer, effective 17-Sep-2025. New customer sign-ups and account upgrades are no longer available.

That's why I decided for a voice-over narration option, and there is a very good AWS service exactly for that purpose.

Amazon Polly

Amazon Polly was created precisely for that - it's AI voice generator and text-to-speech tool, which can create high-quality, natural-sounding human voices in dozens of languages.

It's also super easy to use - we just need to enter text we want to synthesize, select engine, language and voice and download it as an mp3 file.

Workflow

High level overview

Putting this all together - we will create a workflow, which looks like this:

1) We will use prompt engineering techniques to create input for Amazon Nova Micro, asking it to generate JSON metadata for the video, including category, title, slug (id), description and prompts for other foundation models and services to generate thumbnail, audio and video.

2) Output from Amazon Nova Micro will look like this:

{
  "category": "Video category",
  "title": "Video title",
  "slug": "video-title",
  "description": "Video description",
  "thumbnail": "Prompt for Amazon Nova Canvas",
  "video": "Prompt for Amazon Nova Reel",
  "audio": "Text for Amazon Polly"
}

3) We will use "thumbnail": "Prompt for Amazon Nova Canvas" from step 2 as an input for Amazon Nova Canvas to generate 3 thumbnails. One of them (arbitrary selected) will be image input for Amazon Nova Reel to generate a video

4) We will use "audio": "Text for Amazon Polly" from step 2) as an input for Amazon Polly to generate voice-over narration audio.

5) We will use "video": "Prompt for Amazon Nova Reel" from step 2 as an input for Amazon Nova Reel, together with one of the thumbnails from step 3) to generate a video.

6) We will use Apple iMovie to merge audio from step 4 with video from step 5 and produce a final video with an audio track.

Let's try it

Prerequisites

Not all foundation models in Amazon Bedrock are supported in all regions, so if you can't find Amazon Nova models in your AWS region of choice - you can switch to US East (N. Virginia) us-east-1 region.

Models have to also be enabled before you use them. Scroll down the left sidebar to the very bottom and select Bedrock configurations -> Model access -> Modify model access and enable Nova Reel, Nova Micro, Nova Canvas models.

Prompt

Prompt engineering is process of developing, designing and optimising prompts to enhance the output of FM (Foundation Model) to our needs.

Improved prompting technique consists of:

Instruction - a task for the model to do
Context - external information to guide the model
Input data - the input for which you want a response
Output indicator - the output type or format

After some experiments I ended up with prompt like this:

Instruction:
Generate an example video metadata for a given [CATEGORY]
and [TOPIC] suggestion (if provided). Output should be in
a JSON format, with the following kay-value pairs

- category: provided category in the input
- title: video title
- slug: title slug
- description: video description
- thumbnail: prompt for Amazon Nova Canvas model, which will generate video thumbnail
- video: prompt for Amazon Nova Reel, which will generate video (based on that prompt and generated thumbnail)
- audio: text for Amazon Polly, which will be synthesised into speech and will be part of the video's audio track
  Make all parts of the metadata are concise, consistent and easily understandable by a 12 year old child. Make sure that video is exactly 6 seconds long and synthesised audio is 5 seconds long. Video script should be simple but engaging animation with a moving parts and changing lights. Don't add any statistics or text overlays to the thumbnail or video.

Context:  
This video content will be used on YouTube clone, which is an educational software engineering project. The video content will be used to showcase the capabilities of the platform.

Examples:

Input:
[CATEGORY] is "Autos & Vehicles"
Output:
{
  "category": "Autos & Vehicles",
  "title": "Magic Wheels: The Future of Cars",
  "slug": "magic-wheels-the-future-of-cars",
  "description": "Discover how cars of the future will move with magic wheels! Watch the cool animation of a car zooming around with flashing lights and see how it might change our rides!",
  "thumbnail": "Create a dynamic and colorful thumbnail showcasing a futuristic car with spinning 'magic wheels' and flashing lights. The car should be zooming through a bright, neon-lit futuristic cityscape.",
  "video": "Generate a 6-second dynamic video with a futuristic car featuring magic wheels that spin and emit flashing lights. The car zooms through a vibrant, neon-lit city with animated movement and bright colors. The scene should be fast-paced and full of energy.",
  "audio": "Welcome to the future of cars! Watch this dynamic future car with magic wheels zooming through a vibrant city."
}

Input:
[CATEGORY] is "Nonprofits & Activism"

Output:

It's not perfect, but good enough for our use case. It's important to note that there are a few prompt engineering techniques e.g.

Enhanced prompt

Instruction

Generate an example video metadata for a given [CATEGORY]
and [TOPIC] suggestion (if provided)
...
Make all parts of the metadata are concise, consistent and easily understandable by a 12 year old child. Make sure that video is exactly 6 seconds long and synthesised audio is 5 seconds long. Video script should be simple but engaging animation with a moving parts and changing lights.

Output indicator

Output should be in a JSON format, with the following kay-value pairs
...

category: provided category in the input

title: video title

slug: title slug

description: video description

thumbnail: prompt for Amazon Nova Canvas model, which will generate video thumbnail

video: prompt for Amazon Nova Reel, which will generate video (based on that prompt and generated thumbnail)

audio: text for Amazon Polly, which will be synthesised into speech and will be part of the video's audio track

Context

This video content will be used on YouTube clone, which is an educational software engineering project. The video content will be used to showcase the capabilities of the platform.

Input

[CATEGORY] is "Nonprofits & Activism"

Negative prompting

A technique where you explicitly instruct the model on what not to include or do in the response.

Don't add any statistics or text overlays to the thumbnail or video.

Few-shot prompting

Provide examples of a task to the model to guide its output. I used one example (one-shot prompting). Probably I could get better results providing more examples, but it was good enough for now.

Input:
[CATEGORY] is "Autos & Vehicles"
Output:
{
"category": "Autos & Vehicles",
"title": "Magic Wheels: The Future of Cars",
"slug": "magic-wheels-the-future-of-cars",
"description": "Discover how cars of the future will move with magic wheels! Watch the cool animation of a car zooming around with flashing lights and see how it might change our rides!",
"thumbnail": "Create a dynamic and colorful thumbnail showcasing a futuristic car with spinning 'magic wheels' and flashing lights. The car should be zooming through a bright, neon-lit futuristic cityscape.",
"video": "Generate a 6-second dynamic video with a futuristic car featuring magic wheels that spin and emit flashing lights. The car zooms through a vibrant, neon-lit city with animated movement and bright colors. The scene should be fast-paced and full of energy.",
"audio": "Welcome to the future of cars! Watch this dynamic future car with magic wheels zooming through a vibrant city."
}

Prompt templates

Templates simplify and standardize the process of generating prompts.

Generate an example video metadata for a given [CATEGORY]
and [TOPIC] suggestion (if provided)
...
[CATEGORY] is "Nonprofits & Activism"

Amazon Nova Micro

We can access Amazon Nova Micro foundation model using Chat/Text playground. Choose "Single prompt" mode from the dropdown. Then click "Select model" and choose Amazon category, Nova Micro model and On-demand inference (which is pay-as-you-go model, where you are charged based on the number of input and output tokens processed for each request).

Let's also adjust some inference parameters which are located in the "Configurations" segment on the left. There are at least two interesting sections:

Randomness and diversity
Length

I increased the "Response length" parameter to 1024, because the default setting of 512 was not enough for our video metadata JSON output.

I also increased "Temperature" and "Top P" parameters to a maximum value of 1 to make model responses more random, diverse and creative.

Let's now paste our prompt and modify input category to "Sports" like this:

Input:
[CATEGORY] is "Sports"

Here is JSON video metadata output generated by the Amazon Nova Micro foundation model:

{
  "category": "Sports",
  "title": "Amazing Soccer Tricks",
  "slug": "amazing-soccer-tricks",
  "description": "Watch this cool soccer player do amazing tricks on the field! See how he spins the ball and makes it move in crazy ways!",
  "thumbnail": "Create a dynamic and colorful thumbnail showing a soccer player doing a spinning trick with the ball, surrounded by flashing lights and a vibrant soccer field.",
  "video": "Generate a 6-second dynamic video featuring a soccer player performing amazing tricks with the ball, including spins and flips. The scene should be fast-paced, full of energy, and colorful with animated movement and lights.",
  "audio": "Check out these amazing soccer tricks! The player spin and flip the ball in incredible ways!"
}

Amazon Nova Canvas

Now it's time to generate thumbnails using Amazon Nova Canvas.

We can access Amazon Nova Canvas foundation model using Image/Video playground. Click "Select model" and choose Amazon category, Nova Canvas model and On-demand inference.

Default inference parameters ("Configurations" segment on the left) are just fine. We want "Generate image" action with 3 images with 1280 x 720 size and 16:9 aspect ratio.

Let's use thumbnail prompt generated by Amazon Nova Micro:

Create a dynamic and colorful thumbnail showing a soccer player
doing a spinning trick with the ball, surrounded by flashing
lights and a vibrant soccer field.

Here are 3 thumbnails generated by the Amazon Nova Canvas foundation model:

Amazon Polly

Let's generate our audio track using Amazon Polly. Some voice engines were available only in certain AWS regions, so I switched to US East (N. Virginia) us-east-1 region.

I was trying different engines and voices, but for that one I used "Generative" engine (the most expressive and adaptive speech using Generative Al), "English, Indian" language variant and "Kajal, Female" voice.

Let's use audio input generated by Amazon Nova Micro:

Check out these amazing soccer tricks!
The player spin and flip the ball in incredible ways!

Here is an voice-over narration generated by the Amazon Polly:

Amazon Nova Reel

Now it's finally time to generate the video using Amazon Nova Reel.

We can access the Amazon Nova Reel foundation model using Image/Video playground. Click "Select model" and choose Amazon category, Nova Reel model and On-demand inference.

Let's also adjust some inference parameters which are located in the "Configurations" segment on the left.

"Generate video" action is already selected by default. One thing we have to provide is a reference image, which will be a start frame for our video. We will of course use one of the thumbnails generated by the Amazon Nova Canvas foundation model.

Let's also use video prompt generated by Amazon Nova Micro:

Generate a 6-second dynamic video featuring a soccer player
performing amazing tricks with the ball, including spins and flips.
The scene should be fast-paced, full of energy, and colorful with
animated movement and lights.

Here are the video generated by the Amazon Nova Reel foundation model:

Definitely it's not perfect, but funny for sure - with all those strange movements of the soccer player body. Indeed he does the tricks in some "incredible ways" ;)

Apple iMovie

The last step will be to merge audio generated by Amazon Polly with video generated by Amazon Nova Reel.

We will use the Apple iMovie for that purpose. We just need to drag and drop both resources, place it on the timeline and export it.

Here is the final result:

Content

Here's a list of all the 32 generated videos.

Autos & Vehicles