Training FLUX.2 and Kolors LoRAs on 1,964 Space Images: The Dataset Is More In Demand Than the Models

A Space-Themed Image Generation Experiment

Khadim Hussain has released a space-themed image generation project called Kiffusion. The core workflow is straightforward: first collect space and sci-fi images from NASA, Wallhaven, and LAION-Aesthetics, use JoyCaption to automatically generate detailed image descriptions, and then train LoRA adapters based on this data.

The project consists of two parts—the Kiffusion dataset and the SpaceVision LoRA adapters. The dataset has now been downloaded over 3,300 times on Hugging Face, while the two LoRA adapters together have fewer than 300 downloads. That gap itself is a telling signal.

LoRA (Low-Rank Adaptation) is a parameter-efficient fine-tuning technique proposed by Microsoft Research in 2021. Its core idea is: while keeping all original weights of a pretrained model frozen, inject a pair of low-rank matrices A and B into the attention layers of a Transformer, and use their product to approximate the weight update. Since the rank r typically ranges from 4 to 64, the number of trainable parameters can be reduced to roughly one-thousandth of full fine-tuning. The practical significance of LoRA is especially pronounced in the diffusion model space: a style adapter is often only a few dozen MB, and can be stacked, swapped, or combined on the same base model like a plugin—without needing to duplicate tens of GB of full weights for each style. The 38 MB and 46.6 MB adapters mentioned here are a direct manifestation of this property.

The Dataset: The Story Behind 1,964 Image-Text Pairs

The Kiffusion dataset contains 1,964 image-description pairs, all centered on space and sci-fi themes. There are three sources: NASA provides real aerospace imagery, Wallhaven contributes wallpaper-grade sci-fi visual works, and LAION-Aesthetics supplements with aesthetically filtered internet images.

The key lies in the annotation method. Every image is automatically captioned with JoyCaption—not simple alt text, but detailed descriptions of composition and lighting that are needed for model training. This choice directly determines the dataset's downstream usability: diffusion models need structured visual language during training, not one-sentence summaries.

JoyCaption is an open-source vision-language model for image captioning, released by the fancyfeast team in 2024 and continuously iterated since. Unlike captioners such as BLIP and CLIP that lean toward short sentences, JoyCaption is positioned to generate "training-grade" long descriptions. It can output detailed portrayals of composition, camera, lighting, materials, and atmosphere according to user-specified style instructions, and supports customized natural-language questioning. This makes it especially valuable when preparing training data for diffusion models: the text encoder needs descriptions with sufficient information density, not a single summary-style alt text. Similar tools include GPT-4V-assisted annotation, Florence-2, and CogVLM, but JoyCaption has been widely adopted in the open-source community because it can run locally and has a permissive license.

SpaceVision LoRA: Two Models, Two Sizes

SpaceVision is a collection of LoRA adapters trained on the Kiffusion dataset, using SimpleTuner and Diffusers as the training framework. Two versions were produced:

• An adapter for FLUX.2-klein (9B), 38 MB
• An adapter for Kolors, 46.6 MB

The LoRA approach is to train only a small set of additional weights rather than fine-tune the entire model. This brings a very practical benefit: the adapter is small enough that downloading and switching can be done within seconds. For creators who need to quickly switch between different styles, this is far more flexible than complete models that are often tens of GB.

SimpleTuner and Diffusers are two components often used together in the open-source diffusion model training stack. Diffusers is the inference and training library maintained by Hugging Face, providing modular implementations of UNet, Transformer, schedulers, VAE, text encoders, and more. It is the underlying dependency for the vast majority of LoRA scripts. SimpleTuner is built on top of Diffusers and focuses on "getting training to work": built-in dataset caching, multi-resolution buckets, VRAM optimization, EMA, unified configuration for LoRA and full fine-tuning, and support for multiple architectures such as SDXL, FLUX, Kolors, and Sana. Compared with writing your own training loop, SimpleTuner lets you reproduce the entire process with a single JSON/TOML configuration—one of the key reasons this project's workflow can be transferred to other vertical domains.

What the Download Data Reveals About Real Demand

According to Hugging Face's all-time statistics as of October 2026, the download counts for the three artifacts show clear stratification:

Artifact	Type	Downloads
kiffusion-space-scifi	Dataset	3,306
spacevision-flux2-lora	FLUX.2-klein LoRA	204
spacevision-kolors-lora	Kolors LoRA	66

The dataset is used far more than the adapters themselves. The project author's own interpretation is: a well-annotated vertical-domain dataset may be a more reusable artifact than a trained product. There is truth to this judgment—adapters are tied to specific models, and once the base model iterates, the adapter's value is discounted; a dataset, however, does not depend on any model, and anyone can take it to train their own version.

From the community's behavior, a large number of developers and researchers need high-quality, domain-specific training material, not someone else's trained style weights for a particular style. Kiffusion's download data precisely confirms this.

FLUX.2-klein and Kolors are two base models with clearly different architectures. The FLUX.2 series was launched by Black Forest Labs, and the klein version is the smaller open-weights variant (about 9B) in that series, using rectified flow and MMDiT architecture, with standout performance in instruction following and text rendering, but relatively high VRAM usage. Kolors was developed by the Kuaishou team, based on a U-Net architecture and optimized for the Chinese context, with its text encoder switched to ChatGLM3, giving it noticeably better support for Chinese prompts than most English models. Training two sets of LoRAs from the same dataset is effectively validating the data's model-agnosticism: the same batch of image-text pairs can produce usable style adapters for different architectures, which echoes the argument in this article that "the dataset does not depend on any model."

Implications for Creators and Developers

The value of this project goes beyond space imagery itself. It provides a reproducible workflow: collect domain-specific images → generate detailed descriptions with JoyCaption → train LoRA with SimpleTuner and Diffusers. This workflow can be transferred to any vertical domain, such as architecture, illustration, or product photography.

Another point worth noting is the choice between datasets and adapters. If you have a batch of high-quality domain images, organizing them into an annotated dataset and releasing it publicly may have a greater community impact than training a style LoRA. A dataset is raw material; an adapter is a prepared dish—raw material naturally has a broader audience.

Related Resources

• Kiffusion dataset (Hugging Face)
• SpaceVision FLUX.2-klein LoRA
• SpaceVision Kolors LoRA