ostris / sd15-big-g-alpha

SD 1.5 Big G (alpha)

This is a Stable Diffusion 1.5 model, but it uses the CLIP Big G text encoder instead of the original CLIP-L text encoder. This is just a knowledge transfer pre-train with the goal of preserving the current knowledge of the model. It was only trained using student/teacher training from my SD 1.5 fine tune, Objective Reality v2 . To fully realize the full potential of the much larger text encoder, it would need to be further fine tuned on a large dataset.

Examples

Coming soon

Usage

For diffusers, you can use it like any other stable diffusion model.

from diffusers import StableDiffusionPipeline
import torch

model_id = "ostris/sd15-big-g-alpha"
pipe = StableDiffusionPipeline.from_pretrained(model_id, torch_dtype=torch.float16)
pipe = pipe.to("cuda")

prompt = "a photo of an astronaut riding a horse on mars"
image = pipe(prompt).images[0]  
    
image.save("astronaut_rides_horse.png")

It will not work out of the box with Comfy UI or Auto1111. There would need to be special code to load it. If there is any interest in this model, I may work on compatibility. Overall, it won't be hard to add. The only architecture change is the text encoder the and cross attention weights.

Alpha

This is just a pretrained alpha. There are some concepts that did not seem to transfer. It really needs proper training on a large dataset. Anyone is welcome to take this task on. I do not plan to at the time.

Why make this?

In the words of George Mallory, "Because it's there"

Training Method

As mentioned above, it was trained using student/teacher only. This was an iterative process over the corse of a few months, and I did not keep track of all of the exact numbers. The following are best estimates.

The cross attention layers were trained for 1-2 million steps with a batch size of 8 on a single 4090 GPU. Then the full unet was trained for around 100k steps with the same settings.