Carousel Fit / 차곡 — LaMa FP32
Public, byte-identical mirror of Carve Photos' LaMa FP32 ONNX model for local photo background extension. This repository contains only the model and notices; no user photos, app source, or generated user images are published here.
Original LaMa: Copyright [2021] Samsung Research, Apache License 2.0.
Original project: https://github.com/advimman/lama
ONNX conversion: https://github.com/Carve-Photos/lama
Upstream model: https://huggingface.co/Carve/LaMa-ONNX
Upstream file: lama_fp32.onnx, file commit a3ee2fca54baebec351b8fa7786154ffa7555aa6.
The upstream model card declares Apache-2.0; LICENSE-LAMA.txt and LAMA-NOTICE.txt
retain the license and attribution. No endorsement is implied.
Integrity
- File:
lama-fp32-1faef5301d78.onnx - Size: 208,044,816 bytes
- SHA-256:
1faef5301d78db7dda502fe59966957ec4b79dd64e16f03ed96913c7a4eb68d6 - Weights/graph unmodified; only the local filename is content-addressed.
GitHub release mirror: https://github.com/neureps/carousel-fit-models/releases/tag/lama-v1 Hugging Face mirror: https://huggingface.co/neureps/carousel-fit-models
Runtime contract and limitations
ONNX Runtime 1.24.2 CPU. Input image: float32 RGB [1,3,512,512] in 0–1;
mask: float32 [1,1,512,512], 1 generates and 0 preserves. Output RGB values
are in 0–255. The app symmetrically pads a ratio-preserving working canvas,
crops the result back, and restores original photo pixels. Post-processing is
separate from the model and does not modify these weights.
Model download requires explicit consent, with an additional non-Wi-Fi data confirmation. After downloading and validating SHA-256, processing stays offline. The model is not bundled with the app. High memory usage can cause failure or OS termination on some phones; availability is not a guarantee of device performance. Generated backgrounds can contain artifacts and should be reviewed before saving.