Local AI models

Oraphim keeps several focused AI workflows local through ONNX Runtime. The models are optional packages installed on demand; they are not required for ordinary editing, compositing, color, or export.

EfficientTAM — Magic Mask and smart selection

The segmentation package uses EfficientTAM-S 512×512 ONNX encoder/decoder graphs. Oraphim uses it for Magic Mask, smart selection, tracking-assisted masks, and object-removal mask generation.

For best results, start from a frame where the subject is clearly visible, place positive/negative prompts carefully, inspect the generated matte, then track or refine through difficult frames. Hair, motion blur, occlusion, similar foreground/background colors, and frame exits deserve manual inspection.

LaMa — Content Aware Fill

The fill package uses a LaMa FP32 ONNX model. It requires a source frame and an active mask. Use the smallest clean mask that fully covers the unwanted object plus enough boundary context for the fill to infer surrounding image structure.

Check texture continuity and temporal stability after filling. A good still-frame fill can still flicker across a shot, so review the complete affected range.

Whisper — speech to text

The speech package uses quantized Whisper tiny.en ONNX encoder/decoder models plus its vocabulary. Use it as a transcription/caption starting point, then proofread names, technical terms, punctuation, timing, and line breaks before delivery.

Runtime and privacy

These packages execute through Oraphim's native ONNX path and can use DirectML where supported. Model files are downloaded only when required. Once installed, the focused inference operation does not need to send project frames to a third-party chat model.

Verification

After the first model install, run a short representative test. Confirm the model reports ready, the generated mask/fill/transcript appears in the intended project context, undo or revert a test mutation where applicable, and save/reopen before using the workflow at scale.