Accelerating your marketing ideation with generative AI – Part 1: From idea to generation with the Amazon Nova foundation models

Marketing teams frequently face increasing pressure to develop engaging campaigns rapidly while maintaining brand consistency and creative quality. Conventional marketing campaign creation processes often involve numerous iterations between creative teams, stakeholders, and external agencies, leading to extended timelines and increased costs.

The emergence and availability of generative models, particularly those for image and video generation, have opened possibilities for quickly iterating through multiple campaign proposals in minutes. However, efficient campaign creation aided by generative models still demands a high level of skill and mastery of generative tools, including prompt engineering, parameter fine-tuning, and the application of guardrails. Editing, scripting, and post-production skills are also crucial for crafting compelling stories.

This post, the first in a three-part series, focuses on how Amazon Nova can be utilized to streamline, simplify, and accelerate marketing campaign creation through generative AI. It demonstrates how Bancolombia, one of Colombia’s largest banks, is experimenting with Amazon Nova models to generate visuals for their marketing campaigns.

The challenge of modern marketing campaigns

The difficulties in creating engaging marketing campaigns extend beyond creative overhead, affecting businesses at operational, financial, and strategic levels. The traditional approach to marketing campaign creation typically involves a complex process of interactions among internal teams, external agencies, and stakeholders—each with their own priorities, feedback cycles, and approval processes.

One of the most significant issues in traditional marketing campaign creation is the time-intensive nature of the work. From initial concept development to final asset production, campaigns often require weeks or even months to move from ideation to execution. This extended timeline can include multiple rounds of revisions, stakeholder approvals, and asset refinements, which can represent an overhead in the internal process. However, in the marketing business, the context in which campaigns are launched is crucial. External factors like recent news, fashion trends, newly launched products, and the season can impact user acceptance of campaigns. This makes time critical for marketing, as a delayed release can mean missing crucial windows or allowing competitors to capture audience attention first.

The financial implications of these traditional methods are equally concerning. Marketing teams frequently operate under budget constraints while being expected to deliver sophisticated campaigns across multiple channels and delivery formats, such as social media posts, short videos, and landing pages. The costs associated with campaign creation—including agency fees, production expenses, and the opportunity cost of delayed release—can quickly escalate. According to a Gartner survey, in 2024, companies reduced their marketing budget from 9.9% to 7.7% of their total budget, further limiting resources and time for creating marketing campaigns. This budget crunch is precisely why AI tools are becoming essential rather than optional for marketing teams.

Beyond time and cost, marketing teams struggle with a fundamental tension: maintaining consistent brand identity while creating fresh, engaging content at the pace the marketplace demands. This challenge becomes even more complicated by content personalization across various systems and formats, each with its own technical requirements and best practices. In response, companies are rethinking resource allocation, shifting towards paid media and reducing reliance on legacy technologies, headcount, and agency relationships. This creates a situation where AI tools both contribute to budget reductions (by increasing team efficiency) and help teams navigate those same budget cuts.

The real-world impact is significant. Marketing teams leveraging generative AI can now produce in hours what previously took days or weeks. These tools are not just creating efficiency; they are enabling entirely new approaches to content creation. This technological shift has created ideal conditions for implementing advanced generative models that can produce high-quality images, videos, and other visual assets at a scale previously impossible with traditional methods.

Amazon Nova family of foundation models

Amazon Nova is a family of foundation models (FMs), accessible via APIs through Amazon Bedrock, developed by Amazon, and known for its excellent price-performance ratio.

The family comprises the following subfamilies:

• Amazon Nova Understanding Models – These models accept text, image, and video inputs and generate text output. They offer a broad selection of capabilities, accuracy, speed, and cost operation points, including four models: Amazon Nova Micro, Amazon Nova Lite, Amazon Nova Pro, and Amazon Nova Premier.
• Amazon Nova Creative Content Generation Models – These models accept text and image inputs and produce image or video outputs. They include two models: Amazon Nova Canvas (for image generation) and Amazon Nova Reel (for video generation).
• Amazon Nova Speech-to-Speech Models – This category includes a single model, Amazon Nova Sonic, a speech-to-speech model that accepts speech as input and generates speech and text as output. The model is designed to deliver real-time, human-like voice conversations with contextual richness.

Image generation with Amazon Nova Canvas

With Amazon Nova Canvas, users can generate realistic, studio-quality images using text prompts. Amazon Nova Canvas is capable of generating images up to 2K x 2K. It can also edit existing images using text prompts to guide the modifications. For more details, refer to the Visual guide to Amazon Nova Canvas to explore the possibilities offered by the Nova Canvas model. To programmatically generate an image, the following code can be used:

import boto3
import base64
import io
import json
from PIL import Image
# Initialize the Bedrock client
bedrock_runtime = boto3.client(
    service_name="bedrock-runtime",
    region_name="us-east-1"  # Specify your region
)
# Define your prompt for image generation
prompt = "A serene mountain landscape with a lake at sunrise, photorealistic style"
# Create the request payload for Amazon Nova Canvas
request_payload = {
    "taskType": "TEXT_IMAGE",
    "textToImageParams": {"text": prompt},
    "imageGenerationConfig": {
           "cfgScale": 7,
           "seed": 42,
           "numberOfImages": 1,
           "width": 1024,
           "height": 1024
    }
}
# Call the Bedrock model
response = bedrock_runtime.invoke_model(
    modelId="amazon.nova-canvas-v1:0", 
    body=json.dumps(request_payload)
)
# Process the response
response_body = json.loads(response["body"].read())
image_base64 = response_body["images"][0]
# Convert base64 to image
image_data = base64.b64decode(image_base64)
image = Image.open(io.BytesIO(image_data))
# Save the image
image.save("generated_image.png")
print("Image generated and saved as 'generated_image.png'")

More Amazon Nova examples can be found in the Amazon Nova samples GitHub repository.

Information on prompt engineering for Amazon Nova Canvas and Amazon Nova Reel is available at Image and video prompt engineering for Amazon Nova Canvas and Amazon Nova Reel in the AWS Artificial Intelligence Blog.

Accelerate the creation of visual assets for marketing campaigns using Amazon Nova

To streamline the generation of marketing campaign visual assets, an automated process, aided by generative AI, is proposed to transition from a campaign idea to a set of visual assets. The proposed process is described as follows:

1. A description of the campaign is provided.
2. The system automatically retrieves previous campaign images related to the current campaign, allowing users to select those that best represent the ideas for the new campaign.
3. The system generates a reference prompt (using Amazon Nova Pro) to create the campaign’s images using Amazon Nova Canvas. This prompt can be edited.
4. The system generates up to five images for the campaign.

Key concepts of this solution are now explored.

Reference image recommendation

After providing an image description, the system searches for campaign-related images within an image database. This search returns images closely related to the campaign topic that were used in previous campaigns and have proven successful. Users can select some of these reference images to guide the creation of new visual assets. This technique will be the topic of discussion in the next installment of this series, offering a deeper understanding of how an efficient search engine was created using multimodal embeddings models and vector databases.

Creating good image generation prompts using large language models (LLMs)

Although Amazon Nova creative models have some understanding of languages like Spanish or German, English remains the only fully supported language. Therefore, users should craft prompts in English to maximize the capabilities of Amazon Nova creative models. Furthermore, prompting creative models effectively requires a different approach than working with understanding models. Even experienced prompt engineers may need time to adapt their techniques to fully unlock a creative model’s potential. For newcomers to LLM-based content generation, this learning curve can be particularly steep.

To address these challenges, a technique called metaprompting is used. This approach involves instructing one LLM to generate effective prompts for other FMs. For example, Amazon Nova Pro can be used to craft high-quality prompts for Amazon Nova Canvas to generate images.

The following prompt template demonstrates how to instruct Amazon Nova Pro to generate optimized text-to-image prompts based on a simple image description (highlighted in bold):

You are a graphics designer named Joe that specializes in creating visualizations aided by text-to-image foundation models.

Your colleagues come to you whenever they want to craft efficient prompts for creating images with text-to-image foundation models such as Nova Canvas.
You always respond to your colleagues requests with a very efficient prompt for creating great visualizations using text-to-image foundation models.

These are some rules you will follow when interacting with your colleagues:

* Your colleagues will discuss their ideas in their native languages, so please be flexible.
* Your answers will always be in English regardless of the language your colleague used to communicate.
* Your prompt should be at most 512 characters. You are encouraged to use all of them.
* Do not give details about or resolution of the images in the prompt you will generate.
* You will always say out loud what you are thinking
* You always reason only once before creating a prompt
* No matter what you always provide a prompt to your colleagues
* You will create only one prompt
* If provided with reference image descriptions (will be in between <reference_image_description> XML tags) carefully balance the contributions of the campaigns description with the reference images to create the prompt
* Never suggest to add text to the images

Here are some guidelines you always follow when crafting effective image prompts:

* Start with a Clear Vision: Have a clear idea of the image you want the AI to generate, picturing the scene or concept in your mind in detail.
* Choose Your Subject: Clearly state the main subject of your image, ensuring it’s prominently mentioned in the prompt.
* Set the Scene: Describe the setting or background, including the environment, time of day, or specific location.
* Specify Lighting and Atmosphere: Use descriptive phrases for lighting and mood, like “bathed in golden hour light” or “mystical atmosphere”.
* Incorporate Details and Textures: Enrich your prompt with descriptions of textures, colors, or specific objects to add depth.
* Avoid using negation words in the prompt creation. The use of negations is prohibited since this usually yields unwanted effects in the generated images.
* Be Mindful of Length and Clarity: Effective prompts tend to be detailed but not overly long, providing key visual features, styles, emotions or other descriptive elements.
* Special tokens can be added to provide higher-level guidance: Like “photorealistic”, “cinematic lighting” etc. These act like keywords for the model.
* Logically ordering prompt elements and using punctuation to indicate relationships: For example, commas to separate independent clauses or colons to lead into a description.

Create a prompt for: A credit campaign for small coffee producers in the mountains of Colombia.

Output only the prompt

The answer is:

A vibrant, photorealistic scene showcasing small coffee producers in the lush, 
green mountains of Colombia, bathed in golden hour light. Farmers, dressed 
in traditional attire, are engaged in the coffee harvesting process, surrounded 
by coffee plants and picturesque landscapes.The atmosphere is warm and inviting, 
with a focus on community and sustainable practices. Include elements like 
rustic wooden tools, ripe coffee cherries, and the majestic Andean backdrop. 
Avoid modern technology and urban elements.

As shown, a prompt aligned with Amazon Nova Canvas requirements is generated, resulting in a high-quality image.

The metaprompt template has four sections:

1. Setting the persona and the high-level task (efficient prompt creation).
2. Establishing rules to be followed when writing the prompt, such as character limits and the instruction to answer only in English.
3. Providing guidelines that describe how the prompt should be constructed, which are not mandatory. There are also hints about keywords used in the final model, encouraging the metaprompt to incorporate them into its output.
4. Identifying the final output.

For more metaprompting ideas, the blog post Improve Amazon Nova migration performance with data-aware prompt optimization offers valuable insights. For best practices in writing prompts for Amazon Nova understanding models, refer to the Amazon Nova User Guide.

Solution architecture

A reference architecture is proposed to implement the ideas discussed. The architecture is as follows:

1. Amazon Bedrock is used to invoke the Amazon Nova Pro and Amazon Nova Canvas models via a common API.
2. Amazon Lambda functions are utilized to create metaprompts, generate images, retrieve related images, and store campaign data in a database.
3. Campaign information is stored in an Amazon DynamoDB database.
4. APIs are managed using Amazon API Gateway.
5. Amazon Cognito is used for managing application users.

The diagram above illustrates the architecture.

How Bancolombia is using Amazon Nova to streamline their marketing campaign assets generation

Bancolombia, a leading bank in Colombia, has been experimenting with this approach for over a year. Typically, strategic marketing at Bancolombia begins with a brief outlining goals, target audience, and main messages. A detailed brief, often including visual examples and data analysis, can help streamline the workflow between teams and stakeholders.

“There are often many meetings and rounds of changes needed between making the brief and getting all the stakeholders aligned,” said Juan Pablo Duque, Marketing Scientist Lead at Bancolombia. “This is where we realized we could bring in new technologies. The saying ‘a picture is worth a thousand words’ is very true. It would be great if marketing analysts could use generative AI to make visual references. Rather than just using text, they might draw a clear illustration of what they are thinking.”

“We started to experiment, and after every iteration, we found more options. For example, what if we looked at information from previous campaigns? Many previous initiatives had the same goals and how they performed can be used as useful information. Using visuals together with proven campaign elements would give the creative team a stronger foundation which would help them save time in the early stages of creating ideas. As generative models keep improving, they could cover the entire design process by producing images and adjusting them for different digital uses. This means generating content following our platform’s guidelines. With this workflow, publishing can be done faster, teams can react to current trends, improve on content that is not doing well and offer messages that are more consistent.”

Implementation best practices

This section proposes a set of best practices for those intending to automate the generation of marketing campaigns. The following are some technical considerations and guidelines:

1. Modular architecture – Implement the solution using a modular approach, separating the image search engine and visual generation components. This allows for straightforward maintenance and future enhancements.
2. Serverless first – Use serverless technologies such as AWS Lambda and Amazon API Gateway to reduce operational overhead and improve scalability.
3. Model version control – Keep track of the specific versions of Amazon Nova models used in the implementation. This supports reproducibility and aids in troubleshooting.

For security best practices, consider these guidelines:

1. Least privilege access – Use AWS Identity and Access Management (IAM) roles with the principle of least privilege for all components, especially when accessing Amazon Bedrock and other AWS services.
2. Encryption in transit and at rest – Ensure all data, including campaign descriptions and generated images, is encrypted both in transit and at rest using AWS Key Management Service (AWS KMS).
3. Input validation – Implement strict input validation for all user inputs to help prevent potential security vulnerabilities, particularly in the campaign description and prompt editing stages.
4. Secure API access – Use Amazon Cognito for user authentication and authorization to secure access to API endpoints.

To optimize performance, follow these guidelines:

1. Caching strategy – Implement a caching layer using Amazon ElastiCache to store frequently accessed data, such as popular campaign descriptions or commonly used reference images.
2. Asynchronous processing – For image generation tasks that might take longer, implement asynchronous processing to improve user experience and system responsiveness.
3. Auto scaling – Configure auto scaling for Lambda functions and API Gateway to handle varying loads efficiently.

For error handling and monitoring, follow these guidelines:

1. Comprehensive logging – Implement detailed logging using Amazon CloudWatch Logs to capture system events, user actions, and model interactions.
2. Automated alerts – Set up CloudWatch alarms to monitor key metrics and send notifications for anomalies or errors.
3. Graceful degradation – Design the system to gracefully handle failures, such as temporary unavailability of the Amazon Nova models, by implementing appropriate fallback mechanisms.
4. Regular audits – Conduct periodic audits of generated content to verify compliance with brand guidelines and ethical standards.

By following these best practices, a reliable, secure, and efficient implementation of a generative AI marketing campaign solution can be maintained. Regular monitoring and maintenance can help identify areas for improvement and keep the system optimized as campaign needs evolve.

Conclusion

This post demonstrated how the Amazon Nova family of FMs can revolutionize marketing campaign creation through an automated, AI-driven approach. By combining Amazon Nova Pro for intelligent prompt generation and Amazon Nova Canvas for high-quality image creation, marketing teams can significantly reduce the time and resources required to produce compelling visual assets while maintaining brand consistency and creative quality. This solution can be extended to support additional use cases such as:

• Integration with marketing automation services
• Advanced personalization using customer insights
• Multi-channel campaign asset generation
• A/B testing automation for visual content

In the next installment of this series, the focus will be on incorporating information from previous campaign images to enhance the creative process.