1950S lora update! v2 fixes and broke more things!!! xd
just a remake of my 1.5 sd models!
v1 is a test i mixed all datasets of mostly images from 1950s to 2000s but expect more trained models by era (50s,60s,70s,etc)
Description
v1 some watermarks are visible next version hopefully fixes it!
FAQ
Comments (7)
V2 looking pretty nice :) Thank you for your work. The hair can be a bit crispy and such at distance, but lowering the weight helps.you just gotta update the training prompt to call out the logos and taglines when they're present. i love these pl4yb0y models, but so often they get the watermark burn in and you can't remove it :(
yeah i messed up i will update in v2!
@marcus199911 it's okay, I mean you said you were just using an older dataset, I'm just being a snob lol. I just really like these older sets, and so many of the loras are sooo close lol
@makiaeveli in the past i used datasets with watermarks but this one is funky im removing manually with photoshop tomorrow i will make a update!
@marcus199911 I've had luck just manually calling them out, but you're right the vertical nature of some of the wording might make it hard for the model to spot. I've also used Flux 2 Klein with "remove logo" type prompting with good effect. If you have that model around it might save you some time :)
@makiaeveli i have datasets for 1960s,1970s,1980s,1990s and 2000s so its gonna take a while if im gonna make models! :(
@marcus199911 well that's what I mean, you should just let F2k rip through them. If F2k can spot them with a particular prompt, the Krea2 encoder should be catching them too. You can essentially test with F2k on a few photos to find a prompt pattern that extracts the logo perfectly, and use it in your training prompts. you then just put all the photos in a folder and have comfyui open all the imgs in the folder one by one OR skip it all together and just update your prompts for the next training run :)







