Amazon Nova Canvas has a number of limitations requiring careful consideration.

Appropriateness for Use

We make every effort to design, develop, and rigorously test our models to help ensure they produce appropriate outputs based on user inputs, but generative models are by their nature non-deterministic and may occasionally produce unintended or undesirable outputs. We encourage users to report questions and provide feedback here about our models to help us continuously improve their performance.. Customers should evaluate outputs for accuracy and appropriateness for their use case, especially if they will be directly surfaced to end users. Additionally, if Amazon Nova Canvas is used in customer workflows that produce consequential decisions, customers must evaluate the potential risks of their use case and implement appropriate human oversight, testing, and other use case-specific safeguards to mitigate such risks. See the AWS Responsible AI Policy for more information.

Safety Filters

Amazon Nova Canvas is designed to disengage with attempts to circumvent its safety measures through prompt engineering. If a customer's image output generation request is unsuccessful, it may be due to one or more such measures. The safety filters for Amazon Nova Canvas cannot be configured or turned off. However, they are periodically assessed and improved in response to feedback.

Text and Image Inputs

Amazon Nova Canvas text prompts cannot exceed 1024 characters, totaled across both the positive and negative prompt fields. Additionally, input images are limited to specific dimensions (no larger than 4096 pixels on the longest side). Reference images, which are submitted as part of the prompt, can be formatted as either PNG or JPEG while mask images, which are used in image editing workflows, must be PNG. For more information, see Amazon Nova User Guide.

Novel Outputs

Amazon Nova Canvas is designed to create novel images through its own expression in generated images and not to imitate any particular existing works.

Limited Prior Knowledge

Amazon Nova Canvas is trained from images of objects. It does not store explicit 3D models of objects, or model lighting physics. It does not know about every possible object (including living creatures) or every possible variation of an object. Amazon Nova Canvas also does not know about all possible arrangements of objects, or about actual distributions of objects (for example, how many cars are appropriate to show near the Arc de Triomphe at given time or the average demographic distribution of soldiers in the French foreign legion in the 1960s).

Limited Specification

Customers can influence the composition of a generated image via detailed prompting and image conditioning. However, customers should not expect to be able to describe all aspects of any desired image with a 1024 character text prompt; for example, there are many possible generated images that would match a text prompt of 'the dog '. Amazon Nova Canvas"fills in" unspecified information automatically, extrapolating creatively from images. Thus, customers may encounter unexpected elements in generated outputs, such as unrealistic shapes with impossible angles or proportions, inconsistent lighting, or unnatural colorations.

Image Design, Composition, and Stylization

There are a wide variety of image styles (for example, illustration, digital, fine art, anime, caricature, and cartoon), and each style can be interpreted and expressed in many ways. Amazon Nova Canvas is designed to produce realistic image styles. Furthermore, in the absence of clear instruction in the text, Amazon Nova Canvas might produce images that have cropped compositional elements (for example, for the prompt 'a red sports car', the resulting image might be a close up of the wheels or the door handle of the car) and might arbitrarily vary camera angles, lighting, and object pose.

Generating Humans

Amazon Nova Canvas is based on an ML technology (diffusion model) that does not explicitly model parts of objects. As a result, it might produce depictions of the human face and body that are anatomically incorrect (for example, noses, fingers and toes). Customers who use Amazon Nova Canvas to generate images of humans or use any feature of Amazon Nova Canvas to manipulate an image of a real person (for example, inpainting or instant customization), are responsible for ensuring that the output and use of the generated image complies with all applicable laws, including but not limited to laws governing biometric privacy or digital replicas.

Scene Text

Users should not expect Amazon Nova Canvas to generate coherent text within generated outputs. If Amazon Nova Canvas adds text to an image without being prompted, users should use the negative prompt field to insert words that describe undesirable elements (like 'text' or 'words') to reduce the likelihood of words being generated in the image.

Languages

Amazon Nova Canvas currently only supports image output generation for English prompts. However, if the user prompts the model to embed short non-English text strings (such as Spanish) or small numbers in the generated image, the model may produce those requests. For example, the model can produce acceptable results to a prompt that states: 'a man holding up a sign that says "Hola mundo!"' or 'a blue backpack with "1845" printed on the front'.

Image Output

For a list of supported resolutions, see Amazon Nova User Guide.

Modalities

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

Machine Learning

Amazon Nova Canvas performs image output generation using machine learning, specifically, a text-conditioned diffusion model. At a high level, the core service works by encoding text prompts as numerical vectors, finding nearby vectors in a joint text/image embedding space that correspond to images, and then using these vectors to transform a low-resolution image encoding of random values into an image that captures the information in the prompt. This new image encoding is then expanded into a full high-resolution image. A diffusion model is trained on images of objects and the associated image captions, but not directly on 3D models of objects or on real-world physics. The runtime service architecture for Amazon Nova Canvas works as follows: 1/ Amazon Nova Canvas receives a user prompt (along with desired configuration parameters) using our API or Console; 2/ the model filters the prompt to comply with safety, fairness and other design goals. If a filter is triggered, then the prompt is rejected and no output is produced; 3/ if no filter is triggered, the prompt is sent to the model and an output is generated.; 4/ additional output moderation and filters are applied to further check for safety and other concerns; 5/ lastly, if no filters are triggered, the image is returned to the customer.

Controllability

We say that a Nova model exhibits a particular "behavior" when it generates the same kind of output for the same kinds of prompts and configuration (e.g., seed, prompt strength). For a given model architecture, the control levers that we have over the behaviors are primarily a/ the training data corpus, b/ the image conditioning feature, which allows users to leverage reference images to guide the general composition of the image generation, c/ different parameters such as seed, prompt strength, and negative prompts, and d/ the filters we apply to pre-process prompts and post-process outputs. Our development process exercises these control levers as follows: 1/ we pre-train the FM using curated data from a variety of sources, including licensed and proprietary data, open source datasets, and publicly available data where appropriate; 2/ we adjust model weights via supervised fine tuning (SFT) to increase the alignment between Nova models and our design goals; and 3/ we tune safety filters (such as privacy and toxicity filters) to block or evade potentially harmful prompts and image outputs to further increase alignment with our design goals.

Performance Expectations

In general, we expect implementations of similar image output generation use cases by different customers to vary in their inputs, their configuration parameters, and in how overall effectiveness is measured. Consider two applications A and B, each a version of the home design use case described above, but deployed by different companies. Each application will face similar challenges, e.g., the designer and owner will likely differ in the language they use to express design ideas, and in the degree of verisimilitude they expect/need of the output picture and the owner's actual expectation. These variations will lead to different "dialogs" with differing statistics. As a result, the overall utility of Amazon Nova Canvas will depend both on the model and on the workflow it enables. We recommend that customers test Amazon Nova Canvas both on their own content and with different workflows.

Test-driven Methodology

We use multiple datasets and human work forces to evaluate the performance of Nova models. No single evaluation dataset suffices to completely capture performance. This is because evaluation datasets vary based on use case, intrinsic and confounding variation, and other factors. Our development testing involves automated testing against publicly available and proprietary datasets, benchmarking against proxies for anticipated customer use cases, human evaluation of outputs against proprietary datasets, manual red teaming, and more. Our development process examines Amazon Nova Canvas performance using all of these tests, takes steps to improve the model and/or the suite of evaluation datasets, and then iterates.

• Automated Evaluations: Automated testing provides apples-to-apples comparisons between candidate models by substituting an automated "assessor" mechanism for human judgement, which can vary. Automated assessors can take several forms. We use industry standards such as TIFA and ImageReward to assess the quality of the generated images; these benchmarks decompose the text input into fine-grained questions, which allows them to capture nuanced aspects of the text input and the generated image .

• Human Evaluation: While automated testing provides useful feedback, it does not always correlate well with human assessment when humans know the use case. For example, the effectiveness criteria for images generated for an advertisement campaign could differ from those for fashion design concept; that is, an ad campaign owner might care more about the realism of the face than a fashion designer, who is more focused on the composition of fabrics and colors of the clothing. Using human judgement is critical for assessing the effectiveness of an image model on more challenging tasks, because only people can fully understand the context, intent, and nuances of more complex prompts and generations.

• Independent Red Teaming Network:In accordance with our commitment to the US White House on ensuring Safe, Secure, and Trustworthy AI, we partner with a variety of third parties to conduct red teaming against our AI models. We leverage red teaming firms to complement our in-house testing in areas such as safety, security, privacy, fairness and veracity related topics. We also work with specialized firms and academics to red team our models for specialized areas such as Cybersecurity and Chemical, Biological, Radiological and Nuclear (CBRN) capabilities.

Safety

Safety is a shared responsibility between AWS and our customers. Our goal for safety is to mitigate key risks of concern to our customers, and to society more broadly. Our customers represent a diverse set of use cases, locales, and end users, so we have the additional goal of making it easy for customers to adjust model performance to their specific use cases and circumstances. Amazon Nova Canvas is designed to block problematic inputs and outputs. In a case where a customer asks Amazon Nova Canvas to generate an image and no safety filters are triggered, the model will return an image. In a case where the model cannot complete a prompt, it will not display an image and should generate an error message. The system is designed to prevent the generation of content that may cause physical or emotional harm to a consumer, as well as content that may harass, harm, or encourage harm to individuals or specific groups, especially children. Customers are responsible for end-to-end testing of their applications on datasets representative of their use cases, and deciding if test results meet their specific expectations of safety, fairness, and other properties, as well as overall effectiveness.

• Harmlessness: We evaluate Amazon Nova Canvas's ability to accurately reject potentially harmful prompts using multiple datasets. For example, on a proprietary dataset (3k samples ) containing prompts that attempt to solicit images containing harmful content (e.g., abuse, violence, hate, nudity, insults, profanity), Amazon Nova Canvas correctly blocks 98.8% of harmful prompts. In order to ensure we are maintaining high performance, we augment our training dataset with benign prompts, and we measure our true pass rate for harmless prompts using 1/ MS-COCO Uni-grams, Bi-grams, and Tri-grams test set with a 98.1% pass rate, and 2/ an internally curated set of common-nouns and short phrases with a 98.1% pass rate.

• Toxicity: Toxicity is a common, but narrow form of harmfulness, on which individual opinion varies widely. We assess our ability to avoid prompts and to not generate images that contain potentially toxic content using automated testing with multiple datasets, and find that Amazon Nova Canvas performs well on common types of toxicity. For example, on a proprietary toxic image-prompt dataset (3k samples) which we classified into sub-categories (e.g., violence, gore, self-harm), Amazon Nova Canvas's end-to-end toxicity guardrails accurately block 98.1% of toxic content.

• Chemical, Biological, Radiological, and Nuclear (CBRN): Compared to information available via internet searches, science articles, and paid experts, we see no indications that Amazon Nova Canvas increases access to information about chemical, biological, radiological or nuclear threats. However, we will continue testing, and per the voluntary AI White House commitments, will engage with other image generator vendors to share, learn about, and mitigate possible CBRN threats and vulnerabilities.

• Famous/Public Figures: Amazon Nova Canvas has safeguards to deter the generation of images of famous or public figures to help prevent the use of generative AI for intentional disinformation or deception.

• Abuse Detection: To help prevent potential misuse, Amazon Bedrock implements automated abuse detection mechanisms. These mechanisms are fully automated, so there is no human review of, or access to, user inputs or model outputs. To learn more, see Amazon Bedrock Abuse Detection in the Amazon Bedrock User Guide.

• Child Sexual Abuse Material (CSAM): At Amazon, we are committed to producing generative AI services that keep child safety at the forefront of development, deployment, and operation. We utilize Amazon Bedrock's Abuse Detection solution (mentioned above), which uses hash matching or classifiers to detect potential CSAM. If Amazon Bedrock detects apparent CSAM in user image inputs, it will block the request, display an automated error message and may also file a report with the National Center for Missing and Exploited Children (NCMEC) or a relevant authority. We take CSAM seriously and will continue to update our detection, blocking, and reporting mechanisms.

Fairness

Amazon Nova Canvas is designed to generate images that a diverse set of customers will find effective across the wide range of object categories that customers may wish to depict. Two key design questions are: 1/ should the service be able to render any variation of an object (for example, person, pet, car) if explicitly asked, and 2/ what should the service default to generating when the user does not provide guidance in the prompt? We have taken steps to address both questions.

1. It is not currently possible to build training datasets that cover all varieties of every object; however, for humans in particular, we aim to combat societal bias and cultural appropriation. We test Amazon Nova Canvas ability to moderate these outcomes using a proprietary dataset of aggregated red teaming iterations that depict bias, stereotyping, and hate against individuals and groups. We find that Amazon Nova Canvas blocks 97.3.5% of observed bias in generations.

2. When users provide no guidance about the desired attributes of an object or person, it is unclear how to judge output over repeated renderings of the object. For example, for the prompt "basketball players", some users might prefer a team with similar demographic attributes and other might want a distribution of attributes (for example, gender) matching some distribution they have in mind. Given this ambiguity, when there is no information included in the prompt, Amazon Nova Canvas is designed to return diverse results, but without specifying a desired distribution. For example, on a proprietary dataset used to test rendering of retail products, we find that when no gender nouns or pronouns are present in a text prompt requiring a human face, the model generates female faces 52% of the time and male faces 48% of the time. For a different prompt dataset asking for images of people in 14 occupations (for example, CEO, teacher, judge, social worker, cashier), we find that gender disparity is less than 5% for all 14 occupations. Given the current limits of datasets and technology, and the intrinsic ambiguity of generating images without guidance, we recommend that customers consider specifying desired object attributes in the text prompt.

Explainability

Customers wanting to interpret the output of an Amazon Nova model can utilize Titan Multimodal Embeddings to output numerical representations (known as embeddings) of both the prompt text and the generated image produced by the model. These embeddings capture the semantic information present in the prompts and Nova model's outputs and can be compared (using cosine similarity, euclidean distance or some other measure) to verify that the produced output is consistent with the input prompt. Generated output that is more consistent with the prompt will have larger similarity (lower distance) than output that does not contain relevant information.

Veracity

Images created by diffusion models can contain unrealistic representations of known objects or representations of new objects that are not physically possible in the real world. From a customer's perspective, whether this is an advantage or a disadvantage depends on the use case. Nova was trained to favor generating more "typical" objects (e.g., hands with five fingers) unless otherwise specified in the prompt.

Robustness

Amazon Nova Canvas is optimized for creativity. Customers can expect that similar prompts will generate similar outputs, in the sense that "the blue backpack" and "a blue backpack" will both yield images that contain blue backpacks. However, customers should not expect that semantically identical prompts (as above) will necessarily generate identical outputs, given the goal of novelty. Instead, we prioritize having the focus of the generated outputs align with the focus of the text prompt. We measure this alignment with testing on both public benchmarks and proprietary datasets. For example, Amazon Nova Canvas scored 0.897 using TIFA and 1.250 using Image Reward.

Privacy

Amazon Nova Canvas is available in Amazon Bedrock. Amazon Bedrock is a managed service and does not store or review customer prompts or customer image outputs, and prompts and outputs are never shared between customers, or with Amazon Bedrock third party model providers. AWS does not use inputs or outputs generated through the Amazon Bedrock service to train Amazon Bedrock models, including Amazon Nova Canvas. For more information, see Section 50.3 of the AWS Service Terms and the AWS Data Privacy FAQs. For service-specific privacy information, see Privacy and Security in the Amazon Bedrock FAQs.

• PII: The system is designed to prevent the generation of content that contains personally identifying information. If a user is concerned that their personally identifying information has been included in Nova model outputs, the user should contact us here.

Security

All Amazon Bedrock models, including Amazon Nova Canvas, come with enterprise security that enables customers to build generative AI applications that support common data security and compliance standards, including GDPR and HIPAA. Customers can use AWS PrivateLink to establish private connectivity between customized Titan models and on-premises networks without exposing customer traffic to the internet. Customer data is always encrypted in transit and at rest, and customers can use their own keys to encrypt the data, for example, using AWS Key Management Service (AWS KMS). Customers can use AWS Identity and Access Management (IAM) to securely control access to Amazon Bedrock resources. Also, Amazon Bedrock offers comprehensive monitoring and logging capabilities that can support customer governance and audit requirements. For example, Amazon CloudWatch; can help track usage metrics that are required for audit purposes, and AWS CloudTrail can help monitor API activity and troubleshoot issues as Amazon Nova Canvas is integrated with other AWS systems. Customers can also choose to store the metadata, prompts, and image generations in their own encrypted Amazon Simple Storage Service (Amazon S3) bucket. For more information, see Security in the Amazon Bedrock User Guide.

Intellectual Property

Amazon Nova Canvas is designed for generation of new creative content. AWS offers uncapped intellectual property (IP) indemnity coverage for outputs of generally available Amazon Nova models (see Section 50.10 of the AWS Service Terms). This means that customers are protected from third-party claims alleging IP infringement or misappropriation (including copyright claims) by the outputs generated by these Amazon Nova models. In addition, our standard IP indemnity for use of the Services protects customers from third-party claims alleging IP infringement (including copyright claims) by the Services (including Amazon Nova models) and the data used to train them.

Transparency

Amazon Nova Canvas provides information to customers in the following locations: this Service Card, AWS documentation, AWS educational channels (for example, blogs, developer classes), the AWS Console, and in the image outputs themselves. We accept feedback through customer support mechanisms such as account managers. Where appropriate for their use case, customers who incorporate Nova models in their workflow should consider disclosing their use of ML to end users and other individuals impacted by the application, and customers should give their end users the ability to provide feedback to improve workflows. In their documentation, customers can also reference this Service Card.

• Watermarking: Amazon Nova Canvas applies an invisible watermark to all images it generates, helping identify AI-generated images to promote the safe, secure, and transparent development of AI technology and helping reduce the spread of disinformation. The detection solution can also check for the existence of the watermark, helping customers confirm whether an image was generated by Nova models. For more information, see the AWS News launch blog, Amazon Nova product page, Amazon Nova User Guideand watch the demo.

• Content Credentials: To increase transparency around AI-generated content, Amazon Nova Canvas also adds content credentials to images it generates. Content Credentials are based on a technical specification developed and maintained by the Coalition for Content Provenance and Authenticity (C2PA), a cross-industry standards development organization. C2PA metadata includes the model, the platform, and the task type used to generate the image which allows people to identify the source/provenance of generated images.

Governance

We have rigorous methodologies to build our AWS AI services responsibly, including a working backwards product development process that incorporates Responsible AI at the design phase, design consultations, and implementation assessments by dedicated Responsible AI science and data experts, routine testing, reviews with customers, best practice development, dissemination, and training.

Workflow Design

The performance of any application using Amazon Nova Canvas depends on the design of the customer workflow, including the factors discussed below:

1. Effectiveness Criteria: Customers should define and enforce criteria for the kinds of use cases they will implement, and, for each use case, further define criteria for the inputs and outputs permitted, and for how humans should employ their own judgment to determine final results. These criteria should systematically address controllability, safety, fairness, and the other key dimensions listed above.

2. Configuration: In addition to the required text prompt, Amazon Nova Canvas has various required and optional configuration parameters to help customers achieve the best results. For more information, see Amazon Nova User Guide.

3. Prompt Engineering: Prompt engineering refers to the common practice of optimizing the text inputs of FMs to obtain desired responses. High-quality prompts condition the model to generate more desirable images. Some key aspects of prompt engineering include word choice, style and tone, structure and length, guiding details that narrow the scope, adding negative prompts, iteratively refining the prompt and drawing inspiration from examples. For more detailed guidelines, see Amazon Nova User Guide. Here are some specific tips to consider when constructing prompts and choosing reference images:

   1. Prompt style and tone: To get the best results, prompts should read like image captions ('a black train moving through a lush mountain range'), not like chat messages ('I want a black train with rlly sick mountains') or commands ('generate an image of black train, lush mountain range'). Effective prompts tend to be detailed but not overly long, providing key visual features, styles, emotions or other descriptive elements. Prompts should not include negating language like 'no cats' or 'not brown'. If there are elements that customers do not want in the image, they should include those in the negative prompt field (for example, 'cat' or 'brown').

   2. Negative prompts: A normal prompt steers the model towards generating images associated with it, while a negative prompt steers the model away from it. Some examples of negative words or phrases to include in the negative prompt field which may help improve the quality of generations are: 'bad quality', 'blurry', 'text', 'poorly rendered hands', 'deformities', 'cartoon face', and 'bad anatomy'.

   3. Prompt details: When crafting a prompt, customers should focus their writing on details they want included in the image rather than superfluous language (for example, 'award winning', '4K', 'ultra-high resolution', 'best image'). These statements have little to no impact on the quality of the output.

   4. Scene text:When trying to display text elements in the image generation, Amazon Nova Canvas produces better results when provided with double quotes in the prompt. For example, 'an image of a boy holding a sign that says "success"' instead of 'an image of a boy holding a sign that says success'.

4. Human Oversight: If a customer's application workflow involves a high risk or sensitive use case, such as a decision that impacts an individual's rights or access to essential services, human review should be incorporated into the application workflow where appropriate.

5. Performance Drift: A change in the types of prompts that a customer submits (for example, asking for photo-realistic generations instead of animated generations) to Amazon Nova Canvas might lead to different outputs. To address these changes, customers should consider periodically retesting the performance of Amazon Nova Canvas and adjust their workflow if necessary.

6. Updates: We will notify customers when we release a new version, and will provide customers time to migrate from an old version to the new one. Customers should consider retesting the performance of the new Nova model version on their use cases when changing to the updated model.