khadim-hussain/spacevision-flux2-lora

SpaceVision FLUX.2-klein LoRA

A LoRA adapter for generating stunning space imagery, nebulae, galaxies, and cosmic phenomena

Model Description

SpaceVision is a 38MB LoRA adapter trained on FLUX.2-klein-base-9B to generate photorealistic space imagery inspired by NASA photography and sci-fi art. The model excels at creating:

• Nebulae — Pillars of Creation, Crab Nebula, Helix Nebula, Horsehead Nebula
• Galaxies — Spiral galaxies, Andromeda, cosmic structures
• Planets — Jupiter, Saturn with rings, Earth from space
• Solar phenomena — Solar flares, corona, sunspots
• Deep space — Star fields, cosmic dust, binary systems

Trigger Word

Use spacevision at the beginning of your prompt to activate the LoRA style.

Sample Outputs

Comparison: Base Model vs SpaceVision LoRA

All images generated with identical settings: seed=42, steps=16, guidance_scale=4.0, resolution=1024x1024

Prompt	FLUX.2-klein Base	+ SpaceVision LoRA
Earthrise "spacevision, Earth rising over lunar horizon, blue marble with white clouds, grey Moon surface foreground"
Solar Flare "spacevision, a violent solar flare erupting from the Sun's surface, plasma loops and magnetic field lines, extreme ultraviolet view"
Pillars of Creation "spacevision, the Pillars of Creation in Eagle Nebula, towering columns of cosmic gas and dust illuminated by newborn stars, deep space photograph"
Crab Nebula "spacevision, the Crab Nebula supernova remnant, filaments of ionized gas expanding outward, pulsar at center"

Usage

Requirements

pip install diffusers transformers accelerate safetensors optimum-quanto
pip install git+https://github.com/huggingface/diffusers.git  # For Flux2KleinPipeline

Basic Inference

import torch
from diffusers import Flux2KleinPipeline
from optimum.quanto import quantize, freeze, qint8

# Load base model
pipe = Flux2KleinPipeline.from_pretrained(
    "black-forest-labs/FLUX.2-klein-base-9B",
    torch_dtype=torch.bfloat16
)

# Quantize transformer to int8 (recommended, matches training)
quantize(pipe.transformer, weights=qint8)
freeze(pipe.transformer)
pipe.to("cuda")

# Load SpaceVision LoRA
pipe.load_lora_weights("khadim-hussain/spacevision-flux2-lora")

# Generate image
image = pipe(
    prompt="spacevision, the Pillars of Creation in Eagle Nebula, towering columns of cosmic gas and dust illuminated by newborn stars, deep space photograph",
    num_inference_steps=16,
    guidance_scale=4.0,
    height=1024,
    width=1024,
    generator=torch.Generator(device="cpu").manual_seed(42)
).images[0]

image.save("pillars_of_creation.png")

Without Quantization (Requires ~48GB VRAM)

import torch
from diffusers import Flux2KleinPipeline

pipe = Flux2KleinPipeline.from_pretrained(
    "black-forest-labs/FLUX.2-klein-base-9B",
    torch_dtype=torch.bfloat16
)
pipe.to("cuda")
pipe.load_lora_weights("khadim-hussain/spacevision-flux2-lora")

image = pipe(
    prompt="spacevision, spiral galaxy with vibrant blue arms and orange core",
    num_inference_steps=16,
    guidance_scale=4.0
).images[0]

Training Details

Parameter	Value
Base Model	FLUX.2-klein-base-9B
Training Framework	SimpleTuner
LoRA Rank	16
LoRA Alpha	16
Learning Rate	1e-4
LR Scheduler	Cosine
Epochs	3
Batch Size	1
Gradient Accumulation	4
Resolution	1024x1024
Precision	bfloat16 + int8-quanto
Optimizer	AdamW
Dataset Size	1,964 images

Dataset

Trained on kiffusion-space-scifi, a curated dataset of 1,964 image-caption pairs:

Source	Count	Description
NASA	915	Real space photography (public domain)
Wallhaven	845	High-quality sci-fi & space digital art
LAION	204	Curated from LAION-2B aesthetic subset

All images are RGB, minimum 512px, auto-captioned with JoyCaption using the trigger word spacevision.

Recommended Prompts

spacevision, the Andromeda Galaxy M31, spiral arms with billions of stars, cosmic dust lanes

spacevision, Saturn's rings in stunning detail, ice particles catching sunlight, gas giant backdrop

spacevision, a binary star system with accretion disk, matter streaming between stars

spacevision, the Horsehead Nebula silhouette against glowing hydrogen gas, Orion constellation

spacevision, Jupiter's Great Red Spot, massive anticyclonic storm, swirling cloud bands

spacevision, the Helix Nebula planetary nebula, dying star shedding outer layers, eye of God

Limitations

• Optimized for space/astronomy imagery; may not generalize well to other domains
• Best results with the spacevision trigger word
• Requires int8 quantization for consumer GPUs (~24GB VRAM)
• Generated images are artistic interpretations, not scientifically accurate

License

This LoRA adapter inherits the license from FLUX.2-klein-base-9B. Please refer to the FLUX.2 Research License for usage terms.

The training dataset contains images from NASA (public domain), Wallhaven (various licenses), and LAION (CC-BY-NC-4.0).

Acknowledgments

• Black Forest Labs — For the FLUX.2-klein base model
• SimpleTuner by @bghira — Training framework
• NASA — Public domain space imagery
• Hugging Face — Model hosting and diffusers library

Citation

If you use this model in your research or projects, please cite:

@misc{spacevision-flux2-lora,
  author = {Khadim Hussain},
  title = {SpaceVision FLUX.2-klein LoRA: Space and Astronomy Image Generation},
  year = {2025},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/khadim-hussain/spacevision-flux2-lora}},
}

Related Models

• spacevision-kolors-lora — SpaceVision LoRA for Kolors model
• kiffusion-space-scifi — Training dataset

Contact

For questions, issues, or contributions, please open an issue on the Kiffusion GitHub repository.