A lightweight internal tool for comparing YouTube thumbnail candidates through audience voting and multimodal AI feedback.
ThumbVote was created to support thumbnail discussions during the video production process.
Instead of relying only on subjective opinions, the tool allows users to compare multiple thumbnail candidates, collect votes, and review AI-generated feedback on their visual strengths and weaknesses.
This project was built as an internal workflow experiment rather than a general-purpose A/B testing platform.
The main goal was to explore whether human preference and multimodal AI analysis could be used together to support editorial decisions.
The tool was designed around three questions:
- Which thumbnail do people prefer?
- What visual elements may influence that preference?
- Where do audience reactions and AI analysis agree or differ?
- Uploads multiple thumbnail candidates
- Creates a voting page
- Collects and stores audience votes
- Aggregates voting results
- Uses a multimodal LLM to evaluate uploaded images and their text settings
- Presents human voting results and AI feedback together
Upload thumbnail candidates
↓
Share the voting page
↓
Collect audience votes
↓
Generate multimodal AI feedback
↓
Review both signals during the editorial discussion
- Next.js
- TypeScript
- Supabase
- OpenAI API
- Vercel
The AI analysis is not intended to select the final thumbnail automatically.
Audience voting provides a direct preference signal, while the model can help identify possible reasons behind that preference, such as readability, clickability, image-text synergy, or distinctiveness.
The final decision remains with the content team.
- Voting results may be influenced by the size and composition of the participant group.
- AI feedback can vary depending on the model and prompt.
- The tool does not predict actual YouTube click-through rates.
- It was developed primarily to support internal comparison and discussion.