During my latest round of training I encountered a better than average number of the TI's that failed to recreate an accurate representation of the subjects. Normally, I would simply pick a few versions of it I like, keep those for myself to play with, and retrain the model (revise tags, find different pictures, etc.) to try to improve the accuracy. With the number of version 1 and 2 models that I have now that still aren't close, I thought it would be fun to share them to let others play.hMonk's Failed Celebrity TI's will be a collection of these models. Each new version will either be a different model or another version of a different training run of the same subject, I have a couple that I just can't get to turn out... That said, these are all designed to work with realistic checkpoints on SD1.5, not PonyXL or SDXL. I've stuck them in XL for giggles a couple of times and while interesting, as expected they don't come anywhere near to accurate.I'm setting the permissions on this as I would for an accurate TI, if someone from CivitAI wants to tell me if they should remain under the "Depicts an actual person" rules and "Celebrity" category, or can be moved to the "Character" category and I can loosen things up a bit. Please let me know.In a SD1.5 anime checkpoint they mostly work, and I've heard Dreamshaper does a good job there. I don't do anime, but please do let me know how they turn out by posting to the models.I'm working on redoing most of these for their own release, but some may never see another version. If there is one on the list you would really like to see a "fixed" version of, please be sure to let me know.<ul><li>July 20, 2024 Yael Cohen Aris (yes I know I messed up in the trigger...)<ul><li>Comes close, but didn't get her jaw line, high cheeks, or nose exactly right.</li></ul></li><li>July 20, 2024 Ariela Donovan<ul><li>Made her prettier, looks like it had issues with her high forehead and jaw. Might revisit, but what came out was very nice even if not her.</li></ul></li><li>July 21, 2024 - Blake Eden<ul><li>Change the face shape a lot, but what came out was interesting. Revisiting with an expanded data set later.</li></ul></li><li>July 24, 2024 - Caitriona Balfe<ul><li>Not sure what went wrong on this one. She looks younger and the face is off. Going to revisit the data set after I finish up what I already have in the testing phase.</li></ul></li><li>July 25, 2024 - Carlotta Bonke<ul><li>Rounded her face and did some other things. Not sure I like how it turned out and suspect it had to do with the heavy make up use in the data set.</li></ul></li><li>July 26, 2024 - Yaslen Clemente<ul><li>Jawline and cheeks are off. Lighting in the data set was all over and may have played a factor.</li></ul></li><li>July 27, 2024 - Victoria Turner<ul><li>Cheeks, jawline, and eyes are off between steps. One step got one right, next step got the another, couple got 2 of 3, none got all at the same time.</li></ul></li><li>November 29, 2024 - Lexi Belle<ul><li>I've tried a couple of times at her. It likes to round off her face. I think I need to find some pictures with her less expressive and try that.</li></ul></li><li>December 22, 2024 - Carmen Electra<ul><li>Going to try again with different settings. I tried to limit the data to the Baywatch years. Only one image because of how that affected the TI...</li></ul></li><li>December 28, 2024 - Yvonne Craig<ul><li>She played Batgirl in the 60's and an Orion on Star Trek back in the day. Version 1 shows the age of the images hard. I'm going to redo it and see if I can get it to sharpen up.</li></ul></li></ul>Please do not post content that violates CivitAI's policies. Using this resource to generate adult content may violate the law in your area and I take no responsibility or liability for what you use this for after downloading. Please be responsible and ethical in your use of AI tools

Celebrity TI Fails

Round 2 of trying to get her to turn out. The first version wasn't anywhere close to a resemblance. Round 2 used a different captioning method and tighter training settings. I'm going to revisit the data set and see if that helps.

Files with SHA256: 07cc919fbf0cb01bf61f6807d62a81fdd2e84d13f448e2e0b8e87e270eaa59f7