Tried prompting to get a top-down view of a log split into 4 quarters and it would generate images with tree rings going the wrong direction, with a huge hole in the middle. Negative prompts could not remove the hole.
Seems to be based on a small quick LLM that might be best swapped out for a larger one. Though that might not fit the model.
Is it running the LLM clientside and that limits the size?