Enter a prompt to generate an image using ReNO. The method enhances text-to-image generation by optimizing the initial noise using reward models as detailed in the paper. The demo uses a lower learning rate (2.5) compared to the paper's default (5.0) for smoother trajectories - if you are looking for more drastic changes, you can increase this value. You can also adjust the reward weights to e.g. prioritize either prompt following (increase ImageReward) or aesthetic quality (increase HPS/PickScore) based on your preferences.
The first time you load this demo, it will take a bit to download and initialize the required model. Once loaded, each optimization run takes about 25-60 seconds.
