Doppelganger

• 

  Invite your famous figure in

• 

  Have a live conversation with them

• 

  Based on real problems of interacting with 1 million+ followers

• 

  AI generated selfie with Harrison Ford

• 

  Users can input prompts to generate pictures with famous figures

• 

  AI generated photo of Lady Gaga handshaking

Inspiration

Imagine you attend a Ed Sheeran concert. He takes a few selfies with lucky fans – you're not one of them. You're too far away from him, he's too busy, it's too expensive to get better seating.

What would be cool – to give everyone a selfie of their style with Ed Sheeran. What would be cooler - to let everyone talk to Ed Sheeran.

Ethan, our teammate has 1.4 million followers, and 70% of his income comes from brand deals. Agents and social media manager take 40% of his income. Having this middleman takes a lot of effort

What it does

• Democratise brand deals through AI
• Multiplication of presence
• Technology that augments the talent, allowing their likeness and brand to scale
• Voice to voice communication with an AI personality.
• AI generation of selfies and photos of you with a famous figure in the location, pose and style you want.

How we built it

• Beam: For model training and webhooks
• Python
• Vana: text-to-image, for facial stable diffusion
• OpenAI: GPT-3
• Uberduck.ai: text-to-speech
• Whisper: speech-to-text

Challenges we ran into

• Speed and fidelity of vocal output
• Stability of stable diffusion
• Latency of multiple services
• Orchestration of API calls

Accomplishments that we're proud of

• End to end system
• Live product

What we learned

• Beam hosting, training, development flow
• Audio engineering
• Prompt engineering

What's next for Doppelganger

• Further training and tuning of audio and visual models.

Built With

• beam
• openai
• python
• uberduck.ai
• vana
• whisper