README.md

---
license: apache-2.0
language:
- en
- kk
library_name: diffusers
pipeline_tag: text-to-image
tags:
- text-to-image
- diffusers
- safetensors
- Lumina2Pipeline
- lumina
- kazakh
- central-asia
- cultural-alignment
- beynele
base_model:
- Alpha-VLLM/Lumina-Image-2.0
- google/gemma-2-2b
datasets:
- issai/BeyneleBench
---

# Beynele

Beynele is a Lumina-Image 2.0 based text-to-image model adapted for Kazakh
cultural image generation. It is trained with a data-centric pipeline that
combines curated cultural data, synthetic supervision, revive-before-reject
curation, a base-model anchor dataset, and reference-based evaluation with
Beynele-Bench.

![Beynele qualitative comparison](assets/beynele_teaser.jpg)

## Use With Diffusers

```python
import torch
from diffusers import Lumina2Pipeline

pipe = Lumina2Pipeline.from_pretrained(
    "issai/Beynele",
    torch_dtype=torch.bfloat16,
)
pipe.enable_model_cpu_offload()

prompt = "A Kazakh dombra resting on a patterned felt carpet."
image = pipe(
    prompt,
    height=1024,
    width=1024,
    guidance_scale=4.0,
    num_inference_steps=40,
    cfg_trunc_ratio=0.25,
    cfg_normalization=True,
    generator=torch.Generator("cpu").manual_seed(42),
).images[0]
image.save("beynele_dombra.png")
```

## Model Details

| Field | Value |
| --- | --- |
| Architecture | Lumina-Image 2.0 / flow-based diffusion transformer |
| Base pipeline | `Alpha-VLLM/Lumina-Image-2.0` |
| Text encoder | `google/gemma-2-2b` |
| Diffusers class | `Lumina2Pipeline` |
| Resolution | 1024 x 1024 |
| Recommended dtype | `torch.bfloat16` |
| Recommended steps | 40 |
| Recommended guidance | 4.0 |

Only the diffusion transformer is adapted. The tokenizer, text encoder,
scheduler, and VAE are carried over from the Lumina-Image 2.0 Diffusers release
to provide a direct `from_pretrained` loading path.

## Training Data Summary

The final training pool contains three branches:

| Branch | Examples |
| --- | ---: |
| Core cultural dataset | 196k image-text pairs, about 73k unique images |
| Text-image dataset | 128k examples |
| Base-model anchor dataset | 109k examples |

The cultural dataset covers Kazakh people, material culture, buildings,
landmarks, food, national symbols, natural scenes, activities, and text-bearing
images. The full fine-tuning corpus is not released because of privacy,
licensing, and cultural-data governance constraints.

## Evaluation

| Model | Beynele-Bench | GenEval | WISE | UniGenBench++ |
| --- | ---: | ---: | ---: | ---: |
| Lumina-Image 2.0 | 4.85 | 0.73 | 0.54 | 64.98 |
| Qwen-Image | 6.51 | 0.87 | 0.62 | 78.36 |
| Beynele | 7.29 | 0.74 | 0.51 | 65.53 |
| Beynele + prompt mediation | 7.01 | 0.78 | 0.73 | 68.89 |

Beynele-Bench uses 750 prompt-reference pairs and reports the arithmetic mean of
Qwen3-VL 32B and Gemini 2.5 Pro similarity scores on a 1-10 scale.

## Intended Use

Beynele is intended for research on cultural text-to-image generation,
low-resource visual adaptation, Kazakh cultural representation, benchmarked T2I
evaluation, and data-centric model adaptation.

## Limitations and Safety

The model may hallucinate cultural details, produce imperfect Kazakh text, blur
faces under difficult compositions, or shift prompt details under strong cultural
specialization. It should not be used for identity verification, historical
authentication, or high-stakes cultural documentation. Human review and local
cultural expertise remain important for sensitive uses.

## Provenance

The Hub package contains the converted Diffusers `transformer/` safetensors used
by `Lumina2Pipeline.from_pretrained`. The source EMA checkpoint is retained in
the local release backup and cache for internal traceability.

## Licensing

Beynele is released under the Apache License 2.0. The model is adapted from
`Alpha-VLLM/Lumina-Image-2.0`; users should also follow the licenses and terms
of any bundled or upstream components used by the Diffusers pipeline.

## Citation

```bibtex
@article{aikyn2026beynele,
  title = {A Data-Centric Framework for Adapting Text-to-Image Models to Low-Resource Cultural Domains},
  author = {Aikyn, Nartay and Aryngazin, Anuar and Maxutov, Akylbek and Varol, Huseyin Atakan},
  year = {2026},
  note = {Pre-release manuscript}
}
```