crop time

noice!

submitten and approven, waiting for the frontend to update

hm who next

,,could download pony diffusion and mass-produce images for my mass-produced horse voices

yeee great idea

i wonder if it even knows about some of them though

like how regular sd doesn't seem to know about tails despite knowing about sonic

only one way to find out

i'll download it overnight then

hmm

This is the notebook. You can choose either "object" (for training a new concept) or "style" (for training a new art style). Make sure your images are 512×512 .jpeg squares, upload them somewhere that you can get the direct image link (discord works for this), and put them into the cell! Use 3-5 images. https://colab.research.google.com/github/huggingface/notebooks/blob/main/diffusers/sd_textual_inversion_training.ipynb

Good event to go to: https://www.nvidia.com/en-us/events/speech-ai-summit/

going down an openai jukebox rabbit hole

makes me wanna relearn it so i can try doing that thing again where i gave it an old oasis demo from before they released an album and tell it to continue it in the style of oasis

hi

Is it possible/easy to add a language layer in the Pipeline training?

as in language embedding or new character set

the former sounds decently easy but it'd require ipa to work well which idk if it does

a language embedding

yeah that sounds pretty easy, just copy the code for speaker embeddings

but it would only be rarely used imo, and actually making it useful like google did requires extra code for adversarial loss

me when there are one to five versions of stable diffusion

It's not the official one by CompVis tho

Just a finetuned model

what the hay

they also have this https://huggingface.co/runwayml/stable-diffusion-inpainting