Comment by aaraujo002

The model should perform better with JSON prompts than with natural language prompts. The safety filter may also be triggered incorrectly by natural language prompts, we are aware of this and will make a future checkpoint update to improve it.

See how to prompt the model here: https://github.com/ideogram-oss/ideogram4/blob/main/docs/pro...

Disclaimer: I work at Ideogram.