Following the popularity of avatar creators of 2023, Tencent has just released a repository of photo generation tools that claims to preserve the identity of input users.
This method uses SDXL and is supposed to be more consistent than previous known methods according to the paper.
>Did you folks experiment with building an Agent-like interface that asks more questions before the LLM finally answers?
We built something similar to query DB. Created two versions, one of which was agent based that had a maker-checker style of generation. Basically one generates and the other checks its correctness and if objective has been acheived.
The accuracy improves in the agent driven framework, at the cost of latency.
great tool and as someone mentioned there are quite a few AI products/libraries that accomplish this.
The main challenge for production settings
1) soon becomes optimizing the latency for each step and
2) dealing with complex queries at a reasonable cost (GPT-4 is bit expensive at scale)
We evaluated multiple options and settled on GPT-3Turbo in the short term. However we are not happy with the latency, especially if you create an agentic solution.