ModelListedExternal

Lama Npu 8elite

Precompiled QNN-ONNX build of LaMa (lamadilated, Apache-2.0) for Snapdragon 8 Elite (Hexagon v79), used by Latent Studio for on-device object removal. All 332 layers run on the Hexagon NPU (40 ms/image) instead of...

Sourcehuggingface.co/latentdivergence/lama-npu-8elite/tree/165f5cf3b10eeb261697ae169882197d82f8918b

KindOtherVersionLama dilated qairt contextBgrade

Model source

Kind: Other
File size: 459 B–93.1 MB
Version: vLama dilated qairt context
Updated: Oct 9, 2026
Source: Hugging Face

Source description

LaMa object-removal eraser — Snapdragon 8 Elite (Hexagon v79), Hexagon NPU

Precompiled QNN-ONNX build of LaMa (lama_dilated, Apache-2.0) for Snapdragon 8 Elite (Hexagon v79), used by Latent Studio for on-device object removal. All 332 layers run on the Hexagon NPU (~40 ms/image) instead of seconds on CPU.

Files:

lama_dilated.onnx — EPContext wrapper
lama_dilated_qairt_context.bin — the Hexagon context binary (chip-specific)

Inputs: image (1,3,512,512) RGB 0..1, mask (1,1,512,512). Output: repainted image. Compiled via Qualcomm AI Hub at float precision (no quantization).

Sources

1 source

Hugging FacePrimaryVersion 165f5cf3b10eeb261697ae169882197d82f8918b

huggingface.co/latentdivergence/lama-npu-8elite/tree/165f5cf3b10eeb261697ae169882197d82f8918b

Verified Oct 9

Download files

Model artifacts

2 artifacts

lama_dilated_qairt_context.bin

bin · 93.1 MB · SHA-256 5827365f304c…7ed6 · Hugging Face

Download file

lama_dilated.onnx

onnx · 459 B · SHA-256 04d544d62b53…92f6 · Hugging Face

Download file

Source excerpts

2 excerpts

ReadmeJul 31huggingface.co/latentdivergence/lama-npu-8elite

--- license: apache-2.0 tags: [lama, inpainting, object-removal, qnn, hexagon, npu, on-device] --- # LaMa object-removal eraser — Snapdragon 8 Elite (Hexagon v79), Hexagon NPU Precompiled QNN-ONNX build of **LaMa** (`lama_dilated`, Apache-2.0) for **Snapdragon 8 Elite (Hexagon v79)**, used by Latent Studio for on-device object removal. All 332 layers run on the Hexagon NPU (~40 ms/image) instead of seconds on CPU. Files: - `lama_dilated.onnx` — EPContext wrapper - `lama_dilated_qairt_context.bin` — the Hexagon context binary (chip-specific) Inputs: image `(1,3,512,512)` RGB 0..1, mask `(1,1,512,512)`. Output: repainted image. Compiled via Qualcomm AI Hub at float precision (no quantization).

NoteJul 31huggingface.co/latentdivergence/lama-npu-8elite

Source context: 0 downloads · 0 likes · Repo latentdivergence/lama-npu-8elite