this post was submitted on 08 Aug 2024
15 points (94.1% liked)

Artificial Intelligence

1338 readers
30 users here now

Welcome to the AI Community!

Let's explore AI passionately, foster innovation, and learn together. Follow these guidelines for a vibrant and respectful community:

You can access the AI Wiki at the following link: AI Wiki

Let's create a thriving AI community together!

founded 1 year ago
 

You heard me, I'm curious, I know there's all those dumbass deepnude programs but has anyone actually tried to make a model that takes images of nude humans and puts clothing on them? I guess they don't have to be nude but that does remove a lot of variables in the generation.

I think it would be an interesting little tool to try out new looks you never would really mess with before

all 4 comments
sorted by: hot top controversial new old
[–] [email protected] 3 points 3 months ago

Aren't those already available in some (online) stores?

[–] j4k3 2 points 3 months ago

Yeah there are people messing with this. There are models for product photography and such too. You don't really need to use nudes.

You first need to train a LoRA on the subject, such as yourself. You need a bunch of angles in various lighting conditions. Hundreds of images are best but like 20-50 will do. You need to be wearing similar clothes and as much of a variety of outfits as possible. The hard part is that you need very detailed captions that are unique to each image. You are training on the things that are the same, so if you are the only consistency in the data set, it will train to know you. If you wear the same black shirt in every image, the black shirt is a feature of you, and the model does not differentiate. There is no real logic in generation. Most of the logic is in training captions.

Now do the exact same thing with each piece of clothing; all lighting conditions, worn with many other clothes, etc.

Now you can stack the LoRA layers and see yourself anywhere while wearing said garment. That is how it works.

Now I can set up a complex toolchain to do image to image, but saving the face to make it the same is a pain in the ass, and takes a lot of tuning for each instance. I still need a trained fine tuning LoRA to get a specific near replica of a product. I can easily make a dong like an ankle leash or boobs drag the ground. Those are like universal products. Hell, I can even gen a woman lying in grass with SD3.

The real question is easily accessible datasets and the motivation to caption. There are auto-caption tools, but they suck for the level of detail desired here.

[–] [email protected] 2 points 3 months ago

there was a big thing with software that would put you in a suite for video meetings while in pjs or I guess nude but you sure has hell better have faith it works 100% of the time then.