Amazon Nova is Amazon’s family of foundation models for understanding text, images and video, generating images and video, and conversational speech.
Nova models
Amazon Nova is Amazon’s own model family in Amazon Bedrock. Pick by modality and size: Micro for fast text, Lite and Pro for multimodal understanding, Premier for the hardest tasks, Canvas and Reel to create images and video, and Sonic for speech.
How it works
| Model | Input | Output | Best for |
|---|---|---|---|
| Nova Micro | Text | Text | Lowest latency and cost |
| Nova Lite | Text, image, video | Text | Low-cost multimodal |
| Nova Pro | Text, image, video | Text | Balance of accuracy, speed, cost |
| Nova Premier | Text, image, video | Text | Most complex tasks |
| Nova Canvas | Text, image | Image | Image generation |
| Nova Reel | Text, image | Video | Video generation |
| Nova Sonic | Speech | Speech and text | Voice conversations |
Exam cues
- If the question says text only, fastest and cheapest, think Nova Micro.
- If the question says images or video as input at low cost, think Nova Lite.
- If the question says generate images, think Nova Canvas; generate video, think Nova Reel.
- If the question says spoken conversation, think Nova Sonic.
- If the question says one Nova model for embeddings across text, images, video and audio, think Nova Multimodal Embeddings.
Remember
- Three categories: understanding, creative, speech.
- Nova 2 adds models with up to 1 million tokens of context and reasoning.
Trap: Nova Canvas and Nova Reel create media; they aren’t the models to describe an image. For understanding images, use Lite, Pro or Premier.
From the AWS docs (11)
- Amazon Nova models come in three categories: understanding, creative and speech.
- Amazon Nova Micro is a text-only model with the lowest latency at very low cost.
- Amazon Nova Lite is a very low-cost multimodal model that processes image, video and text inputs quickly.
- Amazon Nova Pro is a multimodal model with the best combination of accuracy, speed and cost for a wide range of tasks, and Amazon Nova Premier is the most capable multimodal model for complex tasks.
- Amazon Nova Canvas generates images and Amazon Nova Reel generates videos, both from text and image inputs.
- Amazon Nova Sonic understands and generates conversational speech, taking speech in and returning speech with text transcriptions.
- In Nova version 1, Nova Micro has a 128K-token context window, Nova Lite and Nova Pro 300K, and Nova Premier 1M.
- Amazon Nova 2 adds new models with up to 1 million tokens of context and extended reasoning.
- Amazon Nova 2 includes Nova 2 Lite (text, image, video and document input with text output, for fast, cost-efficient high-volume use), Nova 2 Sonic (speech in and out) and Nova Multimodal Embeddings (embeddings from text, images, documents, video and audio).
- Nova Canvas edits images with a mask: inpainting replaces the content inside the masked region, and outpainting changes the area outside it, for example to replace a background.
- Nova Canvas adds an invisible watermark to every generated image, and watermark detection can confirm whether an image came from an Amazon Nova model.
Flashcards
Nothing to review here right now. Pick other topics or another filter.
·
All 9 cards as a list
A chatbot only handles text and needs the fastest, cheapest responses. Which Amazon Nova model fits?
Amazon Nova Micro, the text-only, lowest-latency model.
An app must answer questions about product photos and short videos at very low cost. Which Amazon Nova model fits?
Amazon Nova Lite, a low-cost multimodal model for text, image and video inputs.
A marketing team wants to generate short promotional video clips from text and image prompts. Which Amazon Nova model fits?
Amazon Nova Reel. (Nova Canvas generates images.)
Which Amazon Nova model generates images from text descriptions?
Amazon Nova Canvas.
A voice assistant must hold natural spoken conversations. Which Amazon Nova model fits?
Amazon Nova Sonic, the speech-to-speech model.
Can Amazon Nova Micro analyze an uploaded image?
No. Nova Micro is text-only; Nova Lite, Pro and Premier accept images and video.
A semantic search application must find matches across text, images, video and audio with one embedding model. Which Amazon Nova model fits?
Amazon Nova Multimodal Embeddings, which turns text, images, documents, video and audio into embeddings for semantic search and similarity matching.
A marketer wants Nova Canvas to keep a product photo but replace everything behind the product. Which editing mode fits?
Outpainting, which regenerates the area outside the masked subject. Inpainting changes only the area inside the mask.
What does Amazon Nova Canvas add to every generated image to help identify its origin?
An invisible watermark that watermark detection tools can check.