conradlocke

/

krea2-identity-edit

like 156

Hey. Thanks for the feedback. It would help a lot if you could share some examples of cases where v1 worked but v1.1 failed. That way I can ensure what cases need better data for future versions.

Hey. Thanks for the feedback. It would help a lot if you could share some examples of cases where v1 worked but v1.1 failed. That way I can ensure what cases need better data for future versions.

I'll provide some examples as soon as I can; for now, I can say that male faces lose a lot of consistency, while female faces still appear consistent and similar to the base version in the final result. Perhaps the examples aren't necessary; perhaps the male dataset is less balanced than the female one? Try taking a male face or a male reference card to check the consistency. The base version still works well with male faces, but I'm putting it to the test with Hollywood-style scenarios. That's all; the new version also seems to weaken the prompt consistency a bit, partially compromising the scenario, while the base version does a great job, though I expected better from the new version. Thank you so much for all your work! I hope you can improve it in the future.

Dataset is 50/50 men and women. Photos with two people are 50/50 men and women and 50/50 same gender couples, 50/50 opposite gender couples. Mixed enthicities.
I tried to make it as diverse as I could to cover a wide range of faces. However, I've on purpose not trained the model on Hollywood actors. So the Hollywood style beauty might be less represented in my dataset.

I'll make sure to capture more of the extreme ends in terms of body's and faces in future training runs. Though I plan to stay on purely synthetic data and not use real people. (Except if they are in MIT or similiarly licenses datasets)

Dataset is 50/50 men and women. Photos with two people are 50/50 men and women and 50/50 same gender couples, 50/50 opposite gender couples. Mixed enthicities.
I tried to make it as diverse as I could to cover a wide range of faces. However, I've on purpose not trained the model on Hollywood actors. So the Hollywood style beauty might be less represented in my dataset.

I'll make sure to capture more of the extreme ends in terms of body's and faces in future training runs. Though I plan to stay on purely synthetic data and not use real people. (Except if they are in MIT or similiarly licenses datasets)

I ran a test in 16:9 widescreen, but the new version is still way off with both male and female faces—in 16:9 format this time. It completely changed the setting, taking me from outer space to right above an ocean, along with the whole spaceship. This happens with the new version. I believe the base version would have been perfect if it could just preserve male faces better. Hope this helps.

in 9:16 vertical proportions get messed up for me too, faces get squished, etc.

Thanks for your great work! This is very cool. After some initial tests, I think 1.1 is definitely an improvement with likeness retention. The identity bleed is a little frustrating when trying to create images with multiple people. Do you think it will be possible to improve that? And do you expect likeness to get even better with future versions?

Yes definitely on both counts. I'm currently working on a setup to recaption my entire dataset with spatial explanations like "the man on the right wearing the green jacket, and the man on the left wearing the brown beanie" or variations with "@image1" style referencing instead of the current caption style that only mentions roughly "these two men in a bar scene". I'm currently testing implementing insightface and adaface as loss helpers during training to further improve likeness. There's also a inference side improvement coming for the nodes, that so far does improve likeness and also potentially dual input character bleed.

Amazing! Thanks for the response. Looking forward to the next version. Having so much fun with this. Do you have an ETA for the next update?