Improve text and logo integrity

Hi Shontel,

Unlike other avatar models (and the reason you can generate a video with an avatar holding a sign) Hedra employs a large, end to end model. Thus there are some things that might not be intuitive from a consistency standpoint.

That being said, one solution to ensure text is replicated well is to increase the size of the text, that way the model has a stronger reference of the text from frame to frame. Perhaps that might help?

Michael

Hi Shontel,

I think I mentioned this in the other message, but we don’t know why the model changes texts. Image to video models do not always respect text in video, our model often excels in this but does not always nail it.

It’s something we’ll work on for future models.

Hi Ryan,

Do you mind sending me the video? Diffusion models in general definitely struggle with longform text coherence.

Michael

Thanks for the feedback Felix, definitely top of mind for our next model update

Hi there,

This is a common request we’re keeping in mind for our next model. I’m merging with the other tracking issue.

Hello,

Thanks for reaching out and flagging this. We are currently working on a model that has improved preservation of text and logos but we don’t have an ETA yet.

Best,

Will

Hi,

Thank you for reaching out and flagging this. We are working on a model that has improved preservation of text and logos but we don’t quite have an ETA yet.

Thanks,

Will

Hi,

Thanks for reaching out. We are actually working on a model that has improved preservation of text and logos currently but we don't have an ETA on it just yet. That same model should handle objects better as well. For now it’s suggested to crop or remove the objects in hand to limit any further complications.

Best,

Will

This has amazing features however one of the issues i am consistently seeing are my customers logo’s are being distorted. The Lettering, phone numbers on shirts, hats, trucks consistently change once it’s completed. That is my biggest issue. Also, if there is an object in a hand, the obtain will separate in half and then re-attached multiple times. Hoping those issues can be fixed soon.