Text-to-image inference script

[Suheon-Jeong]

Thanks for introducing such excellent work following Infinity!

Is the text-to-image model currently running using weights from a publicly released video model?

Could you share a demo script for text-to-image inference?

Also, I’m interested in the parameter settings used for T2I evaluation.

[JeyesHan]

Thank you for your attention. InfinityStar is initialized from a text-to-image (T2I) pretrained model and then trained exclusively on video data. As a result, compared with the pretrained T2I model, InfinityStar demonstrates a noticeable decline in image generation capability, which is caused by the domain gap between text-to-image (T2I) and text-to-video (T2V) data. Therefore, if you are solely focused on image generation, we recommend using the Infinity-8B model instead.

[Suheon-Jeong]

Thank you for your reply!
I’m currently trying to generate images using Infinity_Star.
Could you share a demo script for image generation, as well as the parameters used for the T2I evaluation in the paper?