Tips for better results
Abi Raja edited this page Apr 15, 2024
·
7 revisions
Not getting code that resembles your screenshot enough? Try these tips:
- Try a different model - We support 3 best in-class models (GPT-4 Turbo (Apr 2024), GPT-4 Vision (Nov 2023), Claude 3 Sonnet). Depending on the exact screenshot and design, one model can perform significantly better than another model.
- Regenerate multiple times - Use the Regenerate button to easily re-try a generation. AI can be a bit random and one generation could be a lot better than the others.
- In a follow-up, tell the AI what it missed or got wrong
- For longer pages, one technique that works well is chopping the page up into smaller pieces, converting them individually to code and piecing them together.
GPT-4 Turbo ranks slightly higher on our evaluations than GPT-4 Vision and Claude 3 Sonnet. But one model might perform significantly better than another model on certain screenshots. Here's a helpful comparison of the three:
GPT-4 Turbo
- Good at matching colors and fonts
GPT-4 Vision
- Good at matching colors and fonts
Claude 3 Sonnet
- Faster
- Better at not being lazy with larger/longer screenshots
- Problem: The generated code is missing many elements. Solution: Try Claude 3 Sonnet or write a follow-up prompt asking the AI: "Please fill in the missing pieces and make it look exactly like the reference screenshot"
- Problem: The font doesn't match. Solution: Instruct the AI with the exact font
We'd love to hear from you so we can improve our model. Email us at support@picoapps.xyz with your screenshot.