Stellar Diffusion

Stellar Diffusion v0.2 vs Base Stable Diffusion v1.5

prompt = A hubble photograph of a galaxy
seed = 42
size = 512x512

Version: 0.2 (Nebula) (Dreambooth / .ckpt formats)

Stable Diffusion 1.5 finetuned on high quality processed space imagery.

Python Usage

from diffusers import StableDiffusionPipeline
import torch

model_id = "rexwang8/stellar-diffusion"
pipe = StableDiffusionPipeline.from_pretrained(model_id, torch_dtype=torch.float16)
pipe = pipe.to("cuda")

prompt = "A hubble photograph of a galaxy"
image = pipe(prompt).images[0]  
    
image.save("hubble_galaxy.png")

Example Results

Suggested parameters

512x512

Place subjects and styles at the very front, roll a few times if you don't get the results you want.

Great - Nebulas, Galaxies

Good - Black holes, Pulsars, Comets, Jupiter

Decent - All other solar system planets

Reconized Tags

All reconized tags can be found in the tags.txt file. They are generated from the annotated descriptions of the photograph.

Partial support for scientific celestial body tags as follows:

NGC - New General Catalogue of Nebulae and Clusters of Stars

M / Messier - A set of 110 astronomical objects catalogued by the French astronomer Charles Messier

UGC – (catalog) Uppsala General Catalogue, a catalog of galaxies

Partial support for the following classification methods as follows:

By recording instrument/spacecraft (ex. Voyager, Hubble)

By Color

By Celestial Body type

Dataset and Credits

Model

Rex Wang (me!)

RunwayML for their SD 1.5

Compute

Coreweave - 2x A40s (~3 A40 hours)

Dataset

91 of the 100 images from https://esahubble.org/ Top 100 Hubble Images ESA/Hubble

~100 additional images from ESA/Hubble and ~10 images from ESA/Webb

~50 additional images from images.nasa.gov

Version History

V0.2 (Codename: Galaxy) - 264 image dataset

V0.1 - 91 image dataset

Contact me

All feedback, criticisms, complaints, etc to discord preferably.

wang5009@purdue.edu

Discord: bob#1236

rexwang8
/

stellar-diffusion