Text-to-Image Fundamentals with Amazon Nova Canvas | AI

More Info

AI-driven image creation has quickly become one of the most groundbreaking technologies, transforming how we produce and engage with visual content. Amazon Nova Canvas stands out as a generative model within the Amazon Nova creative suite, enabling users to produce realistic and imaginative images from simple text descriptions.

This article serves as an introductory guide for utilizing Amazon Nova Canvas. We’ll start with the necessary steps to set up on Amazon Bedrock, a fully managed service that supports leading foundation models (FMs) for diverse applications like text, code, and image generation; summarization; question answering; and custom use cases involving fine-tuning and Retrieval Augmented Generation (RAG). Here, we will specifically focus on the Amazon Nova image generation models accessible in AWS Regions across the US, particularly the Amazon Nova Canvas model. Next, we will outline the image generation process (diffusion) and delve into the input parameters for text-to-image generation with Amazon Nova Canvas.

Getting Started with Image Generation on Amazon Bedrock

Follow these steps to gain access to Amazon Nova Canvas and the image playground:

1. Create an AWS account if you haven’t done so already.
2. Access the Amazon Bedrock console as an AWS Identity and Access Management (IAM) administrator or an appropriate IAM user.
3. Select one of the Regions where the Amazon Nova Canvas model is available (e.g., US East (N. Virginia)).
4. In the navigation pane, click on Model access under Bedrock configurations.
5. Under “What is Model access,” select Modify model access or Enable specific models (if not yet activated).
6. Choose Nova Canvas, then click Next.
7. On the Review and submit page, click Submit.
8. Refresh the Base models. If you see the Amazon Nova Canvas model with Access Granted status, you’re set to proceed.
9. In the navigation pane, choose Image / Video under Playgrounds.
10. Click Select model, then pick Amazon and Nova Canvas. Finally, click Apply.

You are now ready to begin generating images with Amazon Nova Canvas on Amazon Bedrock. The following screenshot illustrates an example from our playground.

Understanding the Generation Process

Amazon Nova Canvas employs diffusion-based techniques for image creation:

• Starting Point: The process kicks off with random noise (a completely static image).
• Iterative Denoising: The model progressively eliminates noise in stages, guided by your prompts. The degree of noise removal at each step is learned during training. For example, if the model needs to generate an image of a cat, it is trained on numerous cat images and gradually adds noise until it is pure noise. When learning to adjust the noise level at each step, the model effectively learns the reverse process, starting from a noisy image and incrementally subtracting noise until a cat image emerges.
• Text Conditioning: The text prompt acts as the guiding factor for the image generation. The prompt is encoded as a numerical vector, compared against similar vectors in a text-image embedding space corresponding to images. Using these vectors, a noisy image is transformed into one that reflects the input prompt.
• Image Conditioning: In addition to text prompts, Amazon Nova Canvas can utilize images as inputs.
• Safety and Fairness: To meet safety and fairness objectives, both the prompt and the generated image pass through filters. If no filter is triggered, the final image is produced.

Prompting Fundamentals

The art of image generation starts with effective prompting—crafting text descriptions that steer the model toward your desired output. Well-constructed prompts should encompass specific details about the subject, style, lighting, perspective, mood, and composition. They tend to be more effective when presented as captions rather than commands or conversational phrases. For example, instead of saying “create an image of a mountain,” a more effective prompt would be “a majestic snow-capped mountain peak at sunset with dramatic lighting and wispy clouds, photorealistic style.” For further insights on prompting, check out this blog post that keeps the reader engaged.

Let’s explore the following prompt elements and their effects on the final output image:

• Subject Descriptions: For instance, using the prompt “a cat sitting on a chair.”
• Style References: For example, prompts like “A cat sitting on a chair, oil painting style” or “A cat sitting on a chair, anime style.”
• Compositional Elements and Technical Specifications: Examples include prompts such as “A cat sitting on a chair, mountains in the background,” and “A cat sitting on a chair, sunlight from the right low angle shot.”

Positive and Negative Prompts

Positive prompts inform the model what to include, specifying the elements, styles, and characteristics desired in the final image. It’s best to avoid negation words like “no,” “not,” or “without” in your prompt. Since Amazon Nova Canvas has been trained on image-caption pairs, captions rarely describe what isn’t present in an image. Therefore, the model doesn’t grasp the concept of negation. Instead, negative prompts indicate elements to exclude from the output.

Negative prompts define what to avoid. Common examples include “blurry,” “distorted,” “low quality,” “poor anatomy,” “bad proportions,” “disfigured hands,” or “extra limbs,” which help models steer clear of typical generation artifacts. For instance, we might start with the prompt “An aerial view of an archipelago,” and then refine it as “An aerial view of an archipelago. Negative Prompt: Beaches.”

The equilibrium between positive and negative prompting creates a well-defined creative space for the model, often leading to more predictable and desirable results.

Image Dimensions and Aspect Ratios

Amazon Nova Canvas is trained with 1:1, portrait, and landscape resolutions, with generation tasks capped at a maximum output resolution of 4.19 million pixels (such as 2048×2048, 2816×1536). For editing tasks, the image should measure 4,096 pixels on its longest side, have an aspect ratio between 1:4 and 4:1, and maintain a total pixel count of 4.19 million or smaller. Being aware of these dimensional constraints helps prevent stretched or distorted results, especially for specialized compositions.

Classifier-Free Guidance Scale

The classifier-free guidance (CFG) scale dictates how closely the model adheres to your prompt:

• Low Values (1.1–3): Granting more creative freedom to the AI, which may yield more aesthetic results but with lower contrast and less adherence to the prompt.
• Medium Values (4–7): A balanced approach, typically recommended for most generations.
• High Values (8–10): Ensuring strict adherence to the prompt, which can produce detailed results but may sacrifice natural aesthetics and lead to heightened color saturation.

For instance, using the prompt “Cherry blossoms, bonsai, Japanese style landscape, high resolution, 8k, lush greens in the background,” the first image with CFG 2 captures some aspects of cherry blossoms and bonsai, while the second image with CFG 8 adheres closely to the details, creating a more serious tone.

For more authoritative insights on this topic, please visit this resource. Additionally, if you’re looking for community feedback and discussions, check out this excellent resource from Reddit.