MI-GAN 512 inpainting (LiteRT float16) v1

LiteRT float16 conversion of the MI-GAN Places2 512×512 generator (Picsart AI Research, ICCV 2023), from andraniksargsyan/migan@406830d0 (migan.onnx, SHA-256 593eba0b…c5ae), for Lightforge's on-device object eraser.

• Asset: migan-512-fp16.tflite, 14,048,880 bytes, SHA-256 a27126ba65208314b113ca82fa70d9a9e119a74237037cb4cc63609a3adc4eeb
• Signature serving_default: input float32 [1,512,512,4] NHWC = concat(mask − 0.5, (rgb·2 − 1)·mask), mask 1 = keep → output float32 [1,512,512,3] NHWC, RGB in [−1, 1]
• Parity with the ONNX pipeline on 17 photos, inside the hole: minimum PSNR 50.6 dB
• Runs fully on LiteRT's GPU accelerator: ~40 ms on a Galaxy S24 Ultra (CPU: ~690 ms)
• License: MIT. Attribution in migan-512/ATTRIBUTION.md. Reproduce with migan-512/convert.py.

This asset is pinned by size and hash in the app; it will never be replaced.

Pet Recognition SMALL (LiteRT float16) v1

LiteRT float16 conversion of open-noodle/pet-recognition-small@9dd4c915 (recognition/model.onnx) for Lightforge's on-device pet re-identification.

• Asset: pet-recognition-small-fp16.tflite, 43,832,152 bytes, SHA-256 63f88741ce15406e90f6ce2194f4fc5018345f03cb408c97405a714ae1cc5759
• Signature serving_default: input float32 [1,224,224,3] NHWC (ImageNet-normalized RGB) → embedding float32 [1,512], L2-normalized
• Parity with the ONNX original on 64 crops / 16 identities: minimum cosine 0.99997, identical nearest neighbours
• License: Apache-2.0. Dataset attributions in ATTRIBUTION.md. Reproduce with pet-recognition-small/convert.py.

This asset is pinned by size and hash in the app; it will never be replaced.

UGallery semantic search models v1

Signed TinyCLIP LiteRT packages for UGallery's private on-device semantic search. Packages contain image/text encoders, the CLIP tokenizer, and license notice. User media and queries are never uploaded.