What is Amazon Nova Canvas?
Amazon Nova Canvas is an image-generation and image-editing model from Amazon. It is offered through Amazon Bedrock, AWS’s managed platform for accessing foundation models, rather than as a standalone consumer image application. Developers can use it to create new images from text, modify existing images, or guide generation with one or more reference images.
The model is intended for visual-content workflows where an application needs more control than a simple prompt-and-generate interaction. Examples include product imagery, advertising concepts, fashion visualization, marketing assets, home-design concepts, and image variations for a brand or catalog.
Nova Canvas belongs to Amazon’s Nova model family, but it is specialized for images rather than general conversation, coding, reasoning, speech, or document analysis. Its canonical Bedrock model ID is amazon.nova-canvas-v1:0.
What the model can create and edit
Nova Canvas supports both generation and editing. Its capabilities include the following:
- Text-to-image generation: Creates a new image from a natural-language description.
- Image conditioning: Uses a reference image to influence composition and layout while following a text prompt.
- Image variation: Produces new images influenced by one or more supplied images.
- Inpainting: Replaces or changes a selected region using a mask.
- Outpainting: Extends an image beyond its original boundaries or changes surrounding background areas.
- Background removal: Removes the background and returns transparent pixels.
- Color-guided generation: Uses between one and ten hexadecimal color values to influence the image palette.
- Virtual try-on: Places an object, garment, or accessory from a reference image onto a source image.
These functions make the model useful for controlled production tasks. For example, a retailer could use a product reference image to generate alternate compositions, remove its background for catalog use, or place an item into a lifestyle scene. A fashion workflow could combine a source person image with a garment reference for a virtual try-on result.
Inputs, outputs, and image limits
Nova Canvas accepts text and image inputs and produces image output. The documented input image formats are PNG and JPEG, and the documented language for prompts is English. Prompts can contain up to 1,024 characters.
For generation tasks, the total output size can be up to 4.19 million pixels. Documented examples include 2,048-by-2,048 and 2,816-by-1,536 images. The general resolution rules require each side to be between 320 and 4,096 pixels, each side to be divisible by 16, an aspect ratio between 1:4 and 4:1, and fewer than 4,194,304 total pixels.
Editing tasks can use images with a longest side of up to 4,096 pixels, subject to the total-pixel and aspect-ratio constraints. These limits matter when designing an application around large product photographs, wide advertising banners, or other fixed output formats. An application may need to resize or crop source images before sending them to the model.
Nova Canvas returns images rather than text. It is therefore not a suitable choice when an application needs a conversational answer, structured text generation, code, transcription, embeddings, or speech output.
Subject consistency and customization
Amazon documents customization through fine-tuning for workflows that need a selected subject to remain recognizable across generated images. Reference images can represent a product, pet, shoe, handbag, or another subject whose appearance should be preserved.
This is different from merely adding a product description to a prompt. Fine-tuning is intended to help the model learn the visual characteristics of a particular subject, which can be useful for repeated brand or catalog imagery. Amazon documents customization settings such as batch size, training steps, and learning rate, although the supplied research does not provide a universal quality guarantee or a fixed number of reference images for every use case.
Fine-tuning should be evaluated against the operational cost and maintenance requirements of a production workflow. If an application only needs occasional image variations, ordinary reference-image conditioning may be sufficient. If it repeatedly generates images of the same product or subject, customization may provide a more consistent basis for those assets.
How Nova Canvas is accessed
Nova Canvas is accessed through the Amazon Bedrock Runtime endpoint using the InvokeModel API. Amazon’s documentation identifies the model ID as amazon.nova-canvas-v1:0. The model does not use the Bedrock Converse API or the Bedrock Mantle endpoint.
The current documented regions supplied for this model are US East (N. Virginia), Europe (Ireland), and Asia Pacific (Tokyo). Availability can depend on the AWS account, region, service status, and applicable access requirements, so teams should verify access in the AWS console before committing to a deployment.
Nova Canvas is an image model rather than a tool-using language model. The research identifies no native function calling, web search, streaming output, JSON mode, batch API, or caching capability for this model. Applications that need orchestration, validation, storage, or post-processing must provide those surrounding services themselves.
Pricing and cost considerations
The supplied research does not expose a verified model-specific dollar price for Nova Canvas. Amazon’s Bedrock pricing page is the appropriate source for current pricing, and teams should check it directly before estimating production costs. Pricing may also depend on the selected AWS region and the exact service or customization workflow.
Because no reliable per-image or per-request amount is provided here, a precise cost comparison would be speculative. A practical evaluation should measure the number of generation and editing requests, image dimensions, retries, customization activity, storage, and any surrounding AWS services. Image generation can also create indirect costs when applications automatically request multiple variations or perform several editing passes.
Editorially, Nova Canvas appears best suited to organizations already using Bedrock and AWS infrastructure. That integration can reduce the need to introduce another platform, but it does not automatically make the model the least expensive option for every image task. Cost should be tested using the actual image sizes and workflow steps required by the application.
Main strengths and limitations
Nova Canvas’s main strength is controlled image production through one Bedrock-accessible model. It combines ordinary text-to-image generation with reference-image conditioning, masked editing, background removal, color control, virtual try-on, and subject-consistency customization. That range is particularly relevant to product, fashion, advertising, and brand-content workflows.
Another practical strength is its enterprise-oriented delivery through Amazon Bedrock. Teams can use the model within an AWS application architecture and apply their existing approach to permissions, regions, monitoring, and service integration. Amazon also describes built-in content moderation and invisible watermarking. The watermark is intended to support provenance and detection of AI-generated content, although it may no longer be detectable if image metadata or related content is removed or altered.
The limitations are significant for some projects. Prompt support is documented as English-only, image dimensions and aspect ratios are constrained, and the model does not provide text, code, audio, video, embedding, or web-search output. It also lacks the general reasoning and coding capabilities associated with language models. Nova Canvas should not be selected as a general-purpose assistant simply because it is part of the broader Nova family.
Most importantly, Amazon currently classifies the model as Legacy and lists a scheduled end-of-life date of September 30, 2026. As of September 25, 2026, the supplied research indicates that it remained listed as accessible in the documented regions, but availability and migration guidance should be confirmed directly with AWS. The end-of-life schedule makes the model a risky foundation for a new long-lived application unless the deployment has a clear migration plan.
Reasoning, coding, and modality profile
Nova Canvas is designed to interpret visual-generation instructions, not to perform open-ended reasoning or software development. No reasoning score or coding score is applicable in the supplied model data, and the model should not be evaluated as a competitor to text-focused models for planning, analysis, or code generation.
| Area | Nova Canvas support |
|---|---|
| Text input | Yes; prompts up to 1,024 characters |
| Image input | Yes; PNG and JPEG are documented |
| Image output | Yes |
| Audio or video input/output | No |
| Text output | No; image output is primary |
| Fine-tuning | Yes, for subject consistency and related customization workflows |
| Function calling and web search | Not supported in the supplied model data |
When to choose Amazon Nova Canvas
Choose Nova Canvas when the application needs an AWS-accessible image model with both generation and editing controls. It is a reasonable fit for:
- Product photography variations and catalog imagery.
- Marketing and advertising concepts generated from prompts and reference assets.
- Background removal and transparent product assets.
- Inpainting and outpainting for controlled image revisions.
- Fashion visualization and virtual try-on workflows.
- Brand or subject-consistent image generation using customization.
- Applications already built around Amazon Bedrock and its regional infrastructure.
Consider another option when the project requires a long-term model with no announced retirement date, multilingual prompt support, text or code generation, audio or video capabilities, or a verified and easily comparable image price. A general-purpose language model is more appropriate for reasoning, coding, and orchestration tasks, while a dedicated video or audio model is needed for those media types.
Nova Canvas can still be useful for an existing application with a short remaining operating horizon, but new production deployments should account for the September 30, 2026 end-of-life date from the beginning. Confirm the current Bedrock model catalog, replacement recommendations, regional access, and pricing before investing in customization or tightly coupling application logic to this model.