Overview
A FLUX-based LoRA fine-tune trained on nearly 15,000 captioned images for image generation reflecting African people, places, and contexts.
AfriVision-Base is a FLUX.2 klein-base-4B LoRA fine-tune, trained under Axiveri on the Axiveri/AfriVision-30K dataset (14,792 images) via a three-notebook Kaggle pipeline: Florence-2 captioning, LoRA training, and CLIP-based cleanup. The trained LoRA was fused into the base model's attention weights and published to HuggingFace.
Building the pipeline meant resolving a transformers version incompatibility, CLIP pooled-embedding extraction failures, and flash_attn compile hangs on Kaggle — plus a subtler bug where calling device_map='balanced' twice in the same process left stale multi-GPU allocations, fixed with explicit per-device cleanup after each pipeline deletion.