Beynele

Text-to-image generation adapted to Kazakh culture.

Beynele adapts Lumina-Image 2.0 with curated cultural images, synthetic supervision, and recovery of useful images that would otherwise be rejected. It covers contemporary people and places alongside cultural objects, food, activities, and Kazakh writing.

Benchmark and data · Quick start · Evaluation

Contemporary portrait	Baiterek, Astana	Beshbarmak

Illustrative generated outputs from the paper. These examples are not an unbiased evaluation sample.

Quick start

Install PyTorch for your system, then install diffusers, transformers, accelerate, and sentencepiece in your Python environment. The example below follows the current release's 40-step setting. See reproducibility for the distinction from the paper's 30-step latency measurement.

import torch
from diffusers import Lumina2Pipeline

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

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

image.save("beynele_dombra.png")

Use English scene descriptions with the downloaded checkpoint. Kazakh writing can be requested inside an English description, but spelling and letter forms may still fail.

Checkpoint and prompt mediation

Component	Role
This checkpoint	Adapted Lumina-Image 2.0 image generator, retaining the original text-conditioning stack
Optional prompt mediation	A separate language-model step that translates and enriches Kazakh or underspecified prompts
Prompt safety screening	A separate part of the deployed service

Loading this repository does not install the prompt-mediation service. The paper evaluates the checkpoint and the mediated system separately. Direct Kazakh prompting of the checkpoint is substantially weaker than mediated prompting in the reported experiments.

Evaluation

The paper reports the following automatic benchmark results. Beynele-Bench is a held-out evaluation resource; it is not the released training corpus.

System	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 scores average the Qwen3-VL 32B and Gemini 2.5 Pro judgments; per-judge values and further baselines are listed on the dataset card. Other columns retain each benchmark's native scale. The manuscript specifies an automatic 0–10 scale, and the saved Qwen-judge files include zero scores.

These are the manuscript's published values, not a re-run. The benchmark has 750 rows, but the saved Qwen3-VL exports for Beynele, Lumina-Image 2.0, and Qwen-Image contain 744 scores each. The Gemini rubric and effective per-judge denominators have not yet been re-verified for this release.

Training and limitations

Training branch	Size reported in the paper
Core cultural data	196k image–text pairs; about 73k unique images
Text-rendering data	128k examples
Base-model anchor data	109k pairs

The full training corpus is not publicly distributed because of privacy and licensing constraints. The reported adaptation uses a roughly 2.6B-parameter generator; this count does not include all components of the inference pipeline.

Outputs may contain incorrect cultural details, misspelled text, distorted hands, and blurred faces. The reference benchmark covers selected instances and does not define a single authentic appearance for Kazakh people or culture. The model is intended for research and creative exploration with appropriate human review; generated images are not documentary evidence.

Reproducibility

Item	Value
Weights revision	1113f57b7463355862ac0624612a2d5f92fc59fd (all weight and config files)
Source checkpoint	EMA checkpoint consolidated_ema.00-of-01.pth, converted to the Diffusers transformer/ folder (see release_config.json)
Tested on 8 October 2026	Quick-start example above, unchanged, on one H100 80GB: Python 3.11.11, PyTorch 2.1.0, diffusers 0.33.1, transformers 4.52.4, accelerate 1.3.0, huggingface_hub 0.35.1 (about 4 minutes including loading)

Pin the revision when reporting results, for example Lumina2Pipeline.from_pretrained("issai/Beynele", revision="1113f57b7463355862ac0624612a2d5f92fc59fd", ...). Evaluation code and run metadata are not yet publicly available.

The published benchmark results come from the evaluation runs Beynele1_retrain_1, Lumina2_Original, and Qwen_Image_Original. No immutable record yet links Beynele1_retrain_1 to the weights revision above. An internal evaluation note mentions Qwen prompt enrichment, while the saved Qwen-judge exports contain identical original and generation prompts. The exact historical prompt processing is therefore unconfirmed.

The quick-start example uses 40 steps. The paper's efficiency table uses 30 steps at 1024 × 1024, batch size 1, on an H100; its 15.5 GB figure should not be presented as a universal minimum VRAM requirement for the offloaded example.

Licence and citation

The model repository declares Apache-2.0 for the adapted generator. Bundled components keep their upstream terms:

Folder	Origin	Terms
transformer/	Fine-tuned from Alpha-VLLM/Lumina-Image-2.0	Apache-2.0
text_encoder/, tokenizer/	Unmodified google/gemma-2-2b, as packaged by Lumina-Image 2.0	Gemma Terms of Use
vae/, scheduler/	Unmodified from the Lumina-Image 2.0 Diffusers release	As declared by that release (Apache-2.0)

The repository does not include separate licence files; consult the upstream repositories for full terms.

@unpublished{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 = {Manuscript in preparation}
}