amazon-nova-canvas

What it does

The skill instructs building production-ready image-generation and image-editing workflows against the Amazon Bedrock InvokeModel API using the amazon.nova-canvas-v1:0 model. It treats each inference as a paid, synchronous operation and preserves the returned image bytes unchanged.

How it works

• Start with a lifecycle gate: do not start new production dependencies on amazon.nova-canvas-v1:0 because it is Legacy in supported Regions and has an EOL date of 2026-09-30. Before continued use by eligible existing customers, perform a series of checks: query GetFoundationModel or ListFoundationModels in the region, confirm eligibility and active status, verify current billing rate, and perform a separately authorized low-cost canary before EOL or private arrangement when paid inference is allowed. Recheck lifecycle and successor information. Do not guess a replacement model ID or broaden access.
• Establish access without embedding secrets: use the AWS SDK credential-provider chain; on AWS compute attach a least-privilege role. For local development, use IAM Identity Center, aws login, or assume-role profile. Do not store keys in source, history, manifests, prompts, or logs. Use providedbearer keywords and restrict permissions to bedrock:InvokeModel for used regions.
• Choose an official task from the supported task types (TEXT_IMAGE, COLOR_GUIDED_GENERATION, IMAGE_VARIATION, INPAINTING, OUTPAINTING, BACKGROUND_REMOVAL, VIRTUAL_TRY_ON) and provide appropriate payloads with required fields. Follow the exact parameter names and structures shown (e.g., textToImageParams, inPaintingParams, etc.). Do not invent additional task types or unsupported fields.
• Apply the common payload structure precisely, including width, height, quality, cfgScale, seed, and numberOfImages as defaults unless the task omits them. Ensure input images are base64-encoded PNG or JPEG with 8-bit RGB, and that dimensions meet the dimensional and aspect-ratio constraints. For each task type, supply the appropriate fields and borders (e.g., conditionImage, controlMode, controlStrength for TEXT_IMAGE; colors and text for COLOR_GUIDED_GENERATION; images and similarityStrength for IMAGE_VARIATION).
• Build prompts as image captions: write concise captions in English describing subject, action, setting, composition, camera/viewpoint, lighting, palette, and medium. Place unwanted concepts in negativeText; keep both positive and negative fields within the character caps. Use double quotes around short scene text and ensure caption length adheres to limits.
• Use a bounded, provenance-preserving Python invocation: install boto3 and Pillow, configure credentials outside the script, obtain approval for paid inference, and perform a single text-to-image request with validation and atomic write of the resulting PNG/JPEG bytes, plus a non-secret manifest.

When to use it

Use when tasks include Nova Canvas or Bedrock image generation/editing scenarios such as canvas inpainting/outpainting/background removal, image conditioning, color guidance, image variation, virtual try-on, or Canvas fine-tuning (noting fine-tuning is restricted due to Legacy status).

What it can touch

The skill specifies using AWS Bedrock InvokeModel API with model amazon.nova-canvas-v1:0. Permissions should be restricted to bedrock:InvokeModel for the specific regional ARNs listed. It also shows examples for IAM policy statements, GetFoundationModel/ListFoundationModels calls, and typical regions us-east-1, eu-west-1, ap-northeast-1. It requires region-bound access to foundation-model resources and related model endpoints. The Python example imports boto3 and Pillow and uses Image for validation.

Caveats

• The model is Legacy in all currently supported Regions with an EOL of 2026-09-30; new production dependencies should not be started. After EOL, requests may fail unless a private arrangement exists.
• New fine-tuning jobs and new Provisioned Throughput endpoints are forbidden after Legacy; public extended access may carry higher pricing.
• Region-bound data-residency decisions apply; never move an image or prompt to a different Region without explicit control.
• Fine-tuning history indicates that existing customer ones may be restricted; otherwise evaluate an Active alternative for new tasks.
• The policy restricts sensitive data exposure: avoid embedding keys in source, history, or logs; use short-lived credentials and least-privilege roles.