Nova Canvas image consistency

To generate consistent characters using Amazon Nova Canvas, the recommended approach is to use the "subject consistency" feature. This feature allows you to fine-tune the model with reference images to preserve a chosen subject (such as a specific character) across multiple generated images.

When you want to create consistent characters, you can provide one or more reference images of the character along with your text prompt. This helps the model understand and maintain the character's appearance, style, and key features across different generations.

Regarding your observation about characters appearing more realistic when not explicitly asking for realism in the prompt, this is an interesting point. It's important to note that the model's output can be influenced by many factors in the prompt, and sometimes less specific descriptions can allow the model more creative freedom. If you're finding that your characters look more realistic without explicitly requesting realism, you might want to experiment with different phrasings and levels of detail in your prompts.

To refine your results, consider these tips:

1. Use a consistent seed value when making small changes to your prompt. This allows you to better understand how specific wording affects the output.
2. Once you've refined your prompt, generate multiple variations using the same prompt but different seed values to find the perfect output image.
3. Remember to phrase your prompts as image captions rather than commands or conversations for best results.

If you're aiming for a specific style or level of realism, try describing the visual style or medium in your prompt (e.g., "photorealistic", "oil painting", "cartoon style") without necessarily using the word "realistic". This can help guide the model towards the desired aesthetic while maintaining the character's consistency.
Sources
Generating images with Amazon Nova - Amazon Nova
Amazon Nova Canvas prompting best practices - Amazon Nova

One of the things you can use is 'CANNY_EDGE' mode for consistent character generation.

The CANNY_EDGE control mode in Amazon Nova Canvas is like providing a detailed outline or blueprint for your character. It's like giving the model a precise sketch to colour in and elaborate upon.

For example:

1. Setting Up CANNY_EDGE Control

inference_params = {
    "taskType": "TEXT_IMAGE",
    "textToImageParams": {
        "text": "3d animated film style, a character with [your details]",
        "conditionImage": your_reference_image,
        "controlMode": "CANNY_EDGE",
        "controlStrength": 0.7  # This can be adjusted between 0-1
    }
}

2. How to Use Control Strength:

• 0.7-0.8: Maintains strong character consistency while allowing some artistic interpretation
• 0.8-1.0: Very strict adherence to the original outlines and shapes
• 0.3-0.6: More creative freedom while keeping basic character features
• Below 0.3: Looser interpretation of the reference

3. Best Practices:

• Start with a clean, well-defined reference image
• CANNY_EDGE works best when your reference image has:
  • Clear outlines
  • Good contrast
  • Distinct facial features
  • Well-defined clothing details
• The model will follow edge details precisely, including:
  • Facial structure
  • Hair style
  • Clothing folds
  • Accessories
  • Pose and positioning

4. Let's say you're creating a cartoon character:

"textToImageParams": {
    "text": "3d animated film style, a young woman with a confident pose, 
             wearing a stylish outfit, bright expression",
    "conditionImage": character_reference,
    "controlMode": "CANNY_EDGE",
    "controlStrength": 0.75  # Good balance of consistency and creativity
}

CANNY_EDGE vs SEGMENTATION:

• Use CANNY_EDGE when:
  • You need exact pose replication
  • Character proportions must stay consistent
  • Specific details are crucial (like facial features)
• Use SEGMENTATION when:
  • You want more flexible pose interpretation
  • General layout is more important than exact details
  • You're okay with creative reinterpretation

Tips for Better Results:

• Combine with clear text prompts that match your reference image
• Use consistent lighting descriptions
• Include style keywords that match your intended output
• Test different controlStrength values to find the sweet spot for your character

Remember, CANNY_EDGE is particularly powerful for maintaining character consistency across multiple generations, especially when you need to maintain specific design elements or brand characteristics.