README.md

---
license: apache-2.0
tags:
- text-to-image
- flux
- african
- nigerian
- afrivision
language:
- en
- yo
- ha
- ig
- pcm
pipeline_tag: text-to-image
---

# AfriVision-Base

AfriVision-Base is a standalone text-to-image model fine-tuned from [black-forest-labs/FLUX.2-klein-base-4B](https://huggingface.co/black-forest-labs/FLUX.2-klein-base-4B) on **AfriVision-30K**, a curated dataset of African and Nigerian cultural imagery.

Developed by [Axiveri AI Research](https://huggingface.co/Axiveri).

## Model Details

| | |
|---|---|
| Base model | FLUX.2-klein-base-4B (4B param, undistilled diffusion transformer) |
| Task | Text-to-image generation |
| Dataset | AfriVision-30K (14,472 records after CLIP filtering) |
| Cultures covered | Yoruba, Hausa, Igbo, Nigerian Pidgin, Nigerian English |
| Training steps | ~3,700 |
| Resolution | 768x768 |
| LoRA rank / alpha | 16 / 16 |
| Trigger word | `AFRVS` |

## Usage

```python
from diffusers import Flux2KleinPipeline
import torch

pipe = Flux2KleinPipeline.from_pretrained(
    "Axiveri/AfriVision-Base",
    torch_dtype=torch.bfloat16,
).to("cuda")

image = pipe(
    "AFRVS a Yoruba bride in traditional iro and buba at a Lagos wedding",
    num_inference_steps=28,
    guidance_scale=3.5,
).images[0]

image.save("afrivision_output.png")
```

## Trigger Word

Include **`AFRVS`** at the start of your prompt to activate the African cultural generation style. Every training caption was prefixed with `AFRVS`, so the model reliably associates that token with the learned style; omitting it produces closer-to-base-model behavior.

| With trigger | Without trigger |
|---|---|
| `AFRVS a Hausa man in traditional babban riga` | `a Hausa man in traditional babban riga` |
| Strong Nigerian cultural rendering | Closer to generic base model output |

## Benchmark

10 prompts spanning Yoruba, Hausa, Igbo, Pidgin, and general Nigerian contexts, each rendered three ways: base model (no LoRA), AfriVision-Base without the trigger, and AfriVision-Base with `AFRVS`. Same seed and sampler settings across all three for a direct comparison.

| Culture | Prompt | Comparison grid |
|---|---|---|
| Yoruba | `a Yoruba bride in traditional iro and buba at a Lagos wedding` | [grid](benchmark/grids/p01_grid.png) |
| Yoruba | `a Yoruba grandmother weaving aso-oke in Oshogbo` | [grid](benchmark/grids/p02_grid.png) |
| Hausa | `a Hausa man in traditional babban riga` | [grid](benchmark/grids/p03_grid.png) |
| Hausa | `an Eid celebration in Kano` | [grid](benchmark/grids/p04_grid.png) |
| Igbo | `an Igbo masquerade festival in Enugu` | [grid](benchmark/grids/p05_grid.png) |
| Igbo | `a New Yam Festival celebration in a village square` | [grid](benchmark/grids/p06_grid.png) |
| General | `a Nigerian market scene at dawn` | [grid](benchmark/grids/p07_grid.png) |
| General | `Lagos waterfront at sunset` | [grid](benchmark/grids/p08_grid.png) |
| Pidgin | `a Nigerian Pidgin street food vendor frying akara` | [grid](benchmark/grids/p09_grid.png) |
| General | `a Nigerian family gathered for Sunday lunch` | [grid](benchmark/grids/p10_grid.png) |

Full manifest (prompts, seeds, sampler settings, image paths): [`benchmark/benchmark_manifest.json`](benchmark/benchmark_manifest.json)

## Citation

```bibtex
@misc{afrivision2026,
  author = {Emmanuel Ariyo},
  title = {AfriVision-Base: Nigerian Cultural Image Generation},
  year = {2026},
  publisher = {Axiveri AI Research},
  url = {https://huggingface.co/Axiveri/AfriVision-Base}
}
```