Really excited for our first dataset release. We're co-founders at Mendit.ai, a small startup in the PNW, we focus on customized and explainable AI models. One thing we noticed as we were building our first product was the need for a curated dataset with diverse resolutions, lighting conditions and scenarios resembling mundane, everyday depictions of characters and objects.

We tried a number of agentic ai frameworks and landed on our own lightweight implementation that first identifies the content properties that are relevant for the use case defined in the prompt and then randomly generates content based on those properties.

An interesting takeaway from our experiments is that the latest LLMs that include reasoning in their response tend to generate noisier outputs than instruction tuned LLMs that generate the output without an in depth plan and explanation.

The dataset can be found on huggingface at: https://huggingface.co/datasets/ninamoss/sleeetview_agentic_...

Keep in mind this is only a first sample, we intend to add new content to the dataset based on community feedback on a regular basis.

Looking forward to any feedback, pointers or insights!