# Stylistic Conditioning Dataset with Diffusion Mixture Architecture

**URL:** <https://discuss.huggingface.co/t/stylistic-conditioning-dataset-with-diffusion-mixture-architecture/172593>\
**Category:** Research\
**Created:** [January 19, 2026, 12:33am UTC](https://discuss.huggingface.co/t/stylistic-conditioning-dataset-with-diffusion-mixture-architecture/172593 "2026-01-19T00:33:11Z")\
**Posts on this page:** 2\
**Page:** 1

<div class="post-metadata">

**Author:** ![Sabith64](https://avatars.discourse-cdn.com/v4/letter/s/6de8d8/32.png) [@Sabith64](https://discuss.huggingface.co/u/Sabith64)\
**Post date:** [January 19, 2026, 12:33am UTC](https://discuss.huggingface.co/t/stylistic-conditioning-dataset-with-diffusion-mixture-architecture/172593/1 "2026-01-19T00:33:11Z")

</div>

Hi everyone, sharing a new research paper + dataset (Apache 2.0) that may be interesting for research on **stylistic conditioning and mixture-based diffusion architecture**.

**Paper (arXiv):**  
[https://arxiv.org/pdf/2601.07941](https://arxiv.org/pdf/2601.07941)

### **Core idea**

The paper introduces the **Moonworks Lunara Aesthetic Dataset** , designed explicitly to study **style conditioning and aesthetics** , rather than treating style as an emergent byproduct of large-scale pretraining. The dataset covers regional arts (e.g., Nordic, East Asian, South Asian, Middle Eastern), as well as general art styles (e.g., oil, sketch).

### **Lunara Aesthetic Dataset**

The dataset focuses on **high-quality, human-curated image–prompt pairs** with explicit stylistic grounding.

Highlights:

- ~2k image–prompt pairs with **consistently high aesthetic scores**

- Prompts emphasize **stylistic intent** (art movement, regional style, medium, mood)

- Suitable for:

### **Diffusion mixture perspective**

The dataset is created with a sub-10B parameter at inference diffusion mixture architecture.

This aligns well with recent interest in:

- Modular fine-tuning (LoRA / adapters per style)

- Better controllability without scaling model size indiscriminately

### **Colab: quick experimentation**

Here’s a Colab notebook for loading the dataset and running basic visualizations:

[https://colab.research.google.com/drive/1beodSkLWIyiaGfJIo4kkQzDPjS8lJb0S?usp=sharing](https://colab.research.google.com/drive/1beodSkLWIyiaGfJIo4kkQzDPjS8lJb0S?usp=sharing)

### **Discussion**

I’m curious how others here are thinking about Mixture vs. monolithic diffusion models for style.

Would love to hear thoughts, critiques, or related work people are exploring.

---

<div class="post-metadata">

**Author:** ![YamSandwich](https://avatars.discourse-cdn.com/v4/letter/y/d78d45/32.png) [@YamSandwich](https://discuss.huggingface.co/u/YamSandwich)\
**Post date:** [February 6, 2026, 3:06am UTC](https://discuss.huggingface.co/t/stylistic-conditioning-dataset-with-diffusion-mixture-architecture/172593/2 "2026-02-06T03:06:24Z")

</div>

Feeling deeply grateful today ❤  
When we released Part 1, we honestly didn’t know how it would land. Seeing the community engage with it, download it, and build with it meant a great deal to me, and it’s what made Part 2 possible. This dataset carries years of memories and artwork, shaped with intention and care, including the contextual variations that quietly but fundamentally influence how Lunara learns.  
I’m deeply thankful to my coauthors who believed in doing this the right way, and to everyone supporting ethical, high-aesthetic work in this space. We’re just getting started. More to come.

📄 Paper: **[https://arxiv.org/pdf/2601.07941](https://arxiv.org/pdf/2601.07941)**  
📔 Colab notebook: [Google Colab](https://colab.research.google.com/drive/1xrtJNS4rljgVa_6UKCuanyS2syJ0QZ7b?usp=sharing)