When using XL Models, the onsite generator only aknowledges the prompt up to a certain length.

Screenshots showcase results using same seed but different prompts.

Adding a bunch of tokens to the end of the prompt resulted in the exact same images being generated. Meanwhile, adding them to the front caused the effect of the original prompt to be completely eliminated.

Conclusion: Only a fixed number of tokens are used for the generation request and the rest are ignored completely. This only happens when using XL models, non-XL models behave as expected.