Omartificial-Intelligence-Space/Khatt-Vision-Arabic-Calligraphy-OCR
KhattVision OCR
An interactive research demo for Arabic calligraphy understanding using a Muse Glimmer 30B vision-language model fine-tuned with LoRA on DuwatBench.
The demo supports:
- full-image Arabic OCR;
- structured JSON extraction;
- six-way calligraphic style recognition;
- nine-way theme classification; and
- normalized text-region localization.
Model configuration
Create a Space variable named ADAPTER_REPO containing the Hugging Face model ID of the pushed LoRA adapter, for example:
Omartificial-Intelligence-Space/KhattVision-Muse-Glimmer-30B-LoRAIf the adapter or base model is gated/private, also create a private Space secret named HF_TOKEN with read access.
Hardware
The model is loaded in 4-bit. 1x L40S (48 GB) is the recommended reliable public-demo configuration. 1x L4 (24 GB) may work with embedding offload but has much less memory headroom.
Held-out results
Structured evaluation: 100% valid JSON, 76% exact style accuracy, 50% exact theme accuracy, 0.707 mean matched IoU, and 71.0% box recall at IoU 0.5.
Limitation
This is a research demo. Dense, overlapping, highly decorative calligraphy can cause plausible but incorrect Arabic text generation. Verify transcriptions before using them in research, archival, religious, or legal contexts.
