This is a dumping ground for models I've trained that I'm not fully happy with, but I also don't think I'll keep burning buzz to try and retrain. Posting them in hopes someone finds some use for them.
I don't guarantee that these'll be very good but, well ,try them out and see!
Description
Trained on about 222 hand-picked frames from vintage (the early 2000s are vintage now right?) media. Mean to have that that early user-generated Internet media feel to it. Digital cameras with low MP and that.
FAQ
Comments (6)
dude yes!!!! I need this so bad for a project!
Enjoy! Would love to see what you make with it
what strength do you use for these results?
For these examples it was between 1.4 and 2.2
amazing LoRA!!! I train with AI-TOOLKIT on krea raw... I can't get these kind of results by training. Is there a training procedure you use? It would be awesome if you could upload the dataset as well in the optional materials... or just an image preview of a few of the samples! How many images? LR? Steps? Thank you sooo much in advance!! :)
Hey there, glad you like it!
If I recall both of these were trained on AI-TOOLKIT via runpod, just on the default krea raw preset. Web 2000 ended up being a merge of a few of the last epochs (I think 10 and maybe 5 & 7?). I unfortunately don't have the training sets anymore I don't think but I'll check.
For the Web 2000 it was 222 images, for the VHS one I don't remember how many but it would've been more than that. The images they were trained on were kind of opposite methods.
Web 2000 it was mostly images from the same production studio, roughly the same size image, similar color grading, clear subjects in view, yadda yada.
With VHS was just a mishmash of sources of whatever I could find that fit the aesthetic from Archive and YouTube and stuff. The images were of irregular size with some of them being really low-res and small dimensions, some of them pretty blurry, and some I cropped pretty aggressively so the data set wouldn't have too many images with those burnt-in timestamps.
For all screencaps, I make sure the final image sizes are divisible by 64 (which I just run some vibecoded script to do it for me) and they were captioned using JoyCaption on ComfyUI with a manual QC pass from me afterward.
I think the results were pretty ok but as you'll notice, you have to up the strength on the LoRas a lot when you use them. I can't tell if it's my captions, the data set I used, or if its just the nature of a LoRa that wants to do low quality fighting uphill against checkpoints trained on high quality.