Quoted By:
I will make a write-up about Noise Injection with wdv later tonight. I find it really interesting how just a tiny bit of noise can bring back backgrounds even on Pony models. (picrel is not pony but there will be more samples in the write-up using autism)
Unless somebody has the code to train a new SD model from scratch and can train it for 1 entire step (or just fill everything with random values) then saves the model as a noise outputting 7GB paperweight, wdv is kind of a requirement to get a noise generator model so I hope the wd guys will not get mad at me for using their test model this way. I am not smart enough to create my own noise generating function and save it as a safetensor checkpoint to allow merging.
There might be a way to achieve this by injecting noise into a specific layer at inference but I'm not 100% sure.
