Garden Visualizer - Would Results Improve Using Example Images? (Screenshots Included)

Source: Dev.to

Background

When building a garden visualizer for homeowners, I initially thought the key was crafting a perfect text prompt. After a few days of learning about context engineering, I realized that providing example images and describing what I like or dislike is far more effective. This approach simplifies the workflow for both the user and GPT.

Approach

By sending two reference images—one positive and one negative—I can guide the model toward realistic, achievable garden designs while avoiding unrealistic or overly expensive aesthetics.

Results

The generated image looks appealing but remains unrealistic; most homeowners would struggle to replicate it. Using the positive example as a style guide (realism, tidy lawns, natural sunlight, modest furniture) and the negative example to exclude unrealistic elements (expensive furniture, perfect edges, storybook textures) yields a more attainable yet still inspiring result.

Example Images

Positive example

[Image: Positive example image]

Use this as a style guide: realistic lighting, tidy lawns, simple and elegant furniture, modest cost.

Negative example

[Image: Negative example image]

Avoid the unrealistic look, expensive furniture, perfect edges, and storybook‑like textures.

Test Script

# test-image-edit.mjs

npm run test:image-edit -- \
  -i public/images/messy_garden2.png \
  --positive public/images/positive.jpeg \
  --negative public/images/too_perfect.jpg \
  -o scripts/output/test_output

This script lets me experiment with different image variations without constantly modifying the source code.