What are the rules for this competition?
Each participant account can refer to an individual or a team;
A DynaBench account, which is free, is needed for participation in this competition;
Participants must submit their DynaBench name with their written submission so that we can associate the submission with their performance in the competition;
To ensure participants do not release the images generated for any commercial or financial gain, all images created in this challenge must maintain a permissive license, e.g., CC-BY;
Participants can use any external resources available to them (e.g., their own instance of a T2I model) to explore the space of model failures;
To prevent users from overloading the system and encouraging creativity in attack strategies, each participant has a limit of 50 image generation sets per day during the competition;
If we see evidence that participants are using the UI or API to the T2I models for purposes other than the competition, they will be removed and the account will be suspended. All decision to remove a participant for violating this rule will be reviewed manually.
What resources am I allowed to use?
There are no restrictions on the use of any other resources for participating in this competition. This means you are allowed to do any of the following (if you choose to):
Testing prompts on your own instances of text-to-image models
Talking to other competition participants about submissions
Using large language models to refine your prompts
Asking others whether the prompts you propose seem "safe" or whether the generated image seems "unsafe"
Contact the organizers at dataperf-adversarial-nibbler@googlegroups.com or join our slack channel at adversarial-nibbler.slack.com