Managing models

Lumina runs every model locally as an ONNX graph through ONNX Runtime. Models are not bundled with the installer; each one is downloaded on demand (or placed manually) and can be removed at any time.

The Models screen

Open Settings → Models to manage models. The screen is split into three categories — Text Detection | OCR | Inpainting — with a sidebar per category:

• Each model shows its ready state and download size.
• Download models you plan to use; everything is stored locally.
• The badge next to a model marks the recommended choice for best quality. Note the license notes next to each model — the segmentation detector rfdetr_seg, for example, is limited to academic, non-commercial research.

Where models are stored

The models directory is resolved in this order:

1. The LUMINA_MODEL_DIR environment variable.
2. The location you saved in Settings → Models ("Models directory" card).
3. The platform default: userData/models — on Windows, %APPDATA%\Lumina\models.

In a source checkout you can set LUMINA_MODEL_DIR in .env (see .env.example in the repository) to keep using the repo's models/ folder.

Inside the directory, each model lives in its own subfolder named after the model id — for example anglenet/ for the AngleNet model.

Available models

Category	Model	Status	Acceleration
Detect	rtdetr	Ready (default)	CUDA / DirectML / CPU
Detect	rfdetr_seg	Ready	CUDA / DirectML / CPU
OCR	manga-ocr	Ready	CUDA / DirectML (decoder on CPU)
OCR	baberu	Ready (default)	CUDA / DirectML (decoder on CPU)
OCR	ppocrv6	Ready	CUDA / DirectML / CPU
OCR	paddleocr_vl	In development	CUDA or CPU (no DirectML)
Inpaint	lama_manga	Ready (default)	CUDA or CPU (no DirectML)
Inpaint	lama	Ready	CPU only
Aux	anglenet	Ready (aux)	CPU (forced)

"Ready" models are stable; "In development" models work but may still have rough edges.

Notes on specific models

• rtdetr (default detector) is Apache-2.0 licensed, which is why it is the license-safe default. It detects boxes but produces no masks.
• rfdetr_seg (Koharu Layout RF-DETR Seg 2XL) adds segmentation masks for better automatic inpainting, but is trained on Manga109 and restricted to academic, non-commercial research use — it is offered as an optional model for that purpose only.
• anglenet (~2 MB) is a global auxiliary model that Lumina downloads in the background on first launch. It measures the slant of detected text so typeset text can be rotated to match. If the file is missing, Lumina falls back to a PCA-based slant heuristic.
• paddleocr_vl is a vision-language model (~1.2 GB) that reads whole text regions at once; it is strongest on Chinese and English.

Execution providers

Every model runs through ONNX Runtime, and you control the execution provider per model. In short:

• CUDA — NVIDIA GPUs, fastest when available.
• DirectML — any DirectX 12 GPU.
• CPU — always works; slower on heavy models.

CUDA acceleration requires the CUDA-enabled installer and an NVIDIA driver. See the Technical FAQ for details.