I would really love to see a 'pose mimic' feature in version 1.2. It would be fantastic if it could accurately replicate human movements.
I would really love to see a 'pose mimic' feature in version 1.2. It would be fantastic if it could accurately replicate human movements. Specifically, it would be great if it could handle not just a single person's pose, but also those of a couple. While many people use ControlNet for this, it’s quite cumbersome, and it’s uncertain if or when a stable version will be available. Besides, there is no ControlNet for Krea 2 yet. I saw so much potential in version 1.0, so I’m really hoping you’ll consider this. I’ve been enjoying your work so much, but I was a little disappointed that this wasn't possible in version 1.1. I truly hope you can add this feature in the next update. Thank you for your hard work!
Thank you! :)
That's a good idea. But I simply don't have the dataset for it currently. Do you mean: Input1 picture of a person, Input2 person striking a pose, output person from image1 striking the pose from image2?
Two people could be potentially tricky, and also only really doable if both images include two people. The model was only trained with up to two images, three just takes too much compute for me.
If you have a dataset with examples of such generations, I would greatly appreciate you uploading them to your huggingface and notifying me, that way I could include them in the next training run. :)
Hey, I noticed you set your comment as resolved. If you found a way to do this with the current model I'd love to know how and the community would benefit as well. :)
I hesitated to ask about this because you mentioned it was difficult to implement and that there wasn't enough data yet; I didn't want to trouble you with a difficult request. However, if you happen to add this feature later on, or if you ever create a separate 'Pose LoRA' for it, it would be a huge help to so many users. Thank you so much for all the work you do—it’s truly appreciated.