Modelpublic

khadim-hussain/spacevision-kolors-lora

sourceHugging Faceapache-2.0updated 8mo agoView on Hugging Face

0likes14downloads

Model Card

Spacevision LoRA - Space & Sci-Fi Image Generation for Kolors

Spacevision is a LoRA (Low-Rank Adaptation) fine-tune of Kwai-Kolors/Kolors-diffusers trained on ~1,964 high-quality space, NASA, and sci-fi images. It generates photorealistic space imagery including satellites, nebulae, galaxies, astronauts, rocket launches, planetary surfaces, and more.

Trigger word: spacevision

Sample Outputs

Example 1: International Space Station

Prompt: "spacevision, Photograph of the International Space Station floating in space with large solar panels extended, Earth's curvature visible below showing blue ocean and white cloud formations against the blackness of space"

Base Kolors (no LoRA)	Spacevision LoRA
[image]	[image]

Example 2: Satellite in Orbit

Prompt: "spacevision, High-resolution photograph of a satellite in space against a pitch-black background, featuring a cylindrical main body with solar panels extending horizontally showing a golden-brown hue with grid-like pattern of solar cells"

Base Kolors (no LoRA)	Spacevision LoRA
[image]	[image]

Quick Start

Install dependencies:

bash

pip install diffusers transformers torch accelerate

Using the LoRA adapter (recommended)

python

import torch
from diffusers import KolorsPipeline

# Load the base Kolors model
pipe = KolorsPipeline.from_pretrained(
    "Kwai-Kolors/Kolors-diffusers",
    variant="fp16",
    torch_dtype=torch.float16,
)

# Load the Spacevision LoRA weights (one line!)
pipe.load_lora_weights("khadim-hussain/spacevision-kolors-lora")

# Move to GPU
pipe.to("cuda")

# Generate an image
image = pipe(
    prompt="spacevision, Photograph of the Pillars of Creation nebula captured by the Hubble Space Telescope, towering columns of gas and dust glowing in shades of gold and green against a dark starfield",
    num_inference_steps=50,
    guidance_scale=3.4,
    height=1024,
    width=1024,
    generator=torch.Generator(device="cuda").manual_seed(42),
).images[0]

image.save("spacevision_output.png")

Adjusting LoRA strength

You can control how strongly the LoRA influences the output using cross_attention_kwargs:

python

# Full LoRA strength (default)
image = pipe(prompt="spacevision, ...", cross_attention_kwargs={"scale": 1.0}).images[0]

# Half LoRA strength (more like base model)
image = pipe(prompt="spacevision, ...", cross_attention_kwargs={"scale": 0.5}).images[0]

Recommended Settings

Parameter	Value
Inference steps	50
Guidance scale (CFG)	3.4
Resolution	1024 x 1024
Trigger word	`spacevision`
Precision	fp16

Training Details

Parameter	Value
Base model	Kwai-Kolors/Kolors-diffusers (2.6B UNet + ChatGLM-6B text encoder)
LoRA rank	16
LoRA alpha	16
LoRA type	Standard (PEFT)
Training epochs	3
Batch size	2
Gradient accumulation	4 (effective batch size: 8)
Learning rate	4e-5
LR scheduler	Cosine
Warmup steps	50
Optimizer	AdamW (bf16)
Mixed precision	bf16
Noise offset	0.02
Resolution	1024
Caption dropout	10%
Training framework	SimpleTuner
GPU	NVIDIA RTX 5090 (Blackwell, 32GB)
File size	~46 MB

Dataset

This model was trained on the khadim-hussain/kiffusion-space-scifi dataset, which contains ~1,964 captioned images across two subsets:

—NASA subset (~915 images): Real space photography from NASA archives including ISS, Hubble telescope imagery, satellite photos, astronaut EVAs, rocket launches, planetary surfaces, nebulae, and galaxy observations
—LAION Sci-Fi subset (~1,049 images): Science fiction artwork and concept art featuring spaceships, futuristic technology, alien worlds, and space exploration themes

All images are auto-captioned with the spacevision trigger word prefix.

The dataset is publicly available on HuggingFace: khadim-hussain/kiffusion-space-scifi

Prompt Tips

—Always include the trigger word spacevision at the beginning of your prompt for best results
—Be descriptive: The model responds well to detailed, NASA-style descriptions mentioning specific objects, colors, and compositions
—Works great with: Nebulae, galaxies, satellites, space stations, astronauts, rocket launches, planetary surfaces, Hubble/JWST imagery, Earth from orbit

Example Prompts

spacevision, NASA photograph of an astronaut performing a spacewalk outside the International Space Station, white spacesuit illuminated by sunlight against the deep black of space with Earth's blue atmosphere glowing on the horizon

spacevision, Hubble Space Telescope deep field image showing thousands of distant galaxies of various shapes and colors, spiral and elliptical galaxies scattered across a dark background revealing the vastness of the observable universe

spacevision, Detailed photograph of the surface of Mars taken by the Curiosity rover showing reddish-brown rocky terrain with layered sedimentary formations, distant mountains on the horizon under a dusty pinkish-tan sky

spacevision, NASA photograph of a SpaceX Falcon 9 rocket launching at night from Cape Canaveral, brilliant orange flame and exhaust plume illuminating the launch pad and surrounding water with streaks of light against a dark sky

License

This LoRA adapter is released under the Apache 2.0 License. The base Kolors model has its own license terms - please refer to Kwai-Kolors/Kolors-diffusers for details.

Acknowledgments

—Kolors by Kwai Kolors Team - The base diffusion model powering this LoRA
—SimpleTuner by bghira - The training framework used for fine-tuning
—LoRA by Edward J. Hu et al. - Low-Rank Adaptation technique
—Diffusers by HuggingFace - The inference library

Citations

If you use this model in your work, please cite:

Kolors (base model)

bibtex

@article{kolors2024,
  title={Kolors: Effective Training of Diffusion Model for Photorealistic Text-to-Image Synthesis},
  author={Kolors Team},
  year={2024},
  publisher={Kwai},
  url={https://huggingface.co/Kwai-Kolors/Kolors-diffusers}
}

LoRA

bibtex

@article{hu2021lora,
  title={LoRA: Low-Rank Adaptation of Large Language Models},
  author={Hu, Edward J and Shen, Yelong and Wallis, Phillip and Allen-Zhu, Zeyuan and Li, Yuanzhi and Wang, Shean and Wang, Lu and Chen, Weizhu},
  journal={arXiv preprint arXiv:2106.09685},
  year={2021}
}

Kiffusion Space Sci-Fi Dataset

bibtex

@dataset{kiffusion_space_scifi,
  title={Kiffusion Space Sci-Fi Dataset},
  author={Khadim Hussain},
  year={2025},
  publisher={HuggingFace},
  url={https://huggingface.co/datasets/khadim-hussain/kiffusion-space-scifi}
}