Hey everyone, been lurking here for a bit while trying to get my Airflow DAGs to behave. This community has been super helpful already, so thanks! 🙏
I saw the announcement about Perplexity's new image generation feature. I tried it out by asking it to create a diagram for a simple ELT pipeline. The results were... okay? It gave me a very generic flowchart.
It got me wondering:
* Is this just a direct API call to DALL-E 3 or Stable Diffusion with a simple prompt wrapper?
* Or is there actually some deeper integration or post-processing happening?
* Has anyone done a proper technical comparison? I'm thinking about:
* Prompt adherence – does it understand complex data workflows better than going direct?
* Output consistency for technical subjects.
* Speed/cost compared to using the image models directly via their native APIs.
I'm asking because, for my work, automatically generating simple architecture diagrams from text descriptions could be a huge time-saver. But if it's just a thin wrapper, I might be better off building a small internal tool using the APIs directly.
Would love to hear if anyone has poked at this more deeply or seen any details on the implementation. The blog post was a bit high-level for my taste.
null
I ran a quick benchmark on this. It's almost certainly DALL-E 3 via API.
I gave Perplexity and the OpenAI API the same prompt: "Create a sequence diagram for a user login flow with Redis session storage."
* Output from both was nearly identical. Same style, same layout quirks.
* Perplexity added a simple border. That's the only post-processing I saw.
* The latency for Perplexity was about 200ms slower, which fits a basic wrapper adding overhead.
So for your use case, generating architecture diagrams, you're right. It's a thin wrapper. You'd get more control building your own tool, especially for iterative refinement of technical diagrams.
Benchmarks don't lie.