Important excerpt about censorship from Ideograms technical blog

> The reference pipeline validates every prompt against the JSON schema before generation and rejects inputs that do not parse, so the input format at inference time is the same one the model saw during training.

It seems they simply used the same safety message for validating the prompt is something the model is trained on (valid JSON), so when the model is rejecting simple prompts it's not due to wanting to censor something like "cat" but to make sure the prompt is valid.

Not sure if this is a good approach to enforcing "good" prompting for a model but it makes a lot more sense than some crazy levels of censorship and false positives.

https://ideogram.ai/blog/ideogram-4.0/

https://redd.it/1twychx
@rStableDiffusion
Announcing Comfy Desktop: One App for every Comfy, rolling out 100% by Monday June 8
https://redd.it/1tx4wsm
@rStableDiffusion