Skip to content

Releases

GLiNER2.5 Decide — Android sample APK 0.1 (debug)

Choose a tag to compare

@john-rocky john-rocky released this 02 Oct 14:52

Pre-built APK of the gliner25decide sample, built from 2a160f8 (the same code as android/ in litert-community/GLiNER2.5-Decide-LiteRT).

  • gliner25decide-0.1-debug.apk: 19,291,588 bytes, Android 8.0 or newer (API 26), arm64-v8a. It is a debug build because the model installer copies the files into the app with run-as.
  • SHA-256: 7cb637d9366a4bb0bdcb8dd3cc8c332c378421754d4fadc78def2e2dc5593cef (also in gliner25decide-0.1-debug.apk.sha256).

The model files (2.45 GB) are not in the APK. Download them and run the installer as in Download, build and install, with adb install gliner25decide-0.1-debug.apk in place of the Gradle build.

v3: MoGe-2 Monocular Geometry

Choose a tag to compare

@john-rocky john-rocky released this 08 Apr 04:21

MoGe-2 ViT-S (35M params)

Monocular 3D geometry estimation — affine point map, surface normals, depth mask, and metric scale from a single image.

File Size Input Output
moge.tflite 136 MB Float32 [1, 3, 448, 448] NCHW Points [1,448,448,3] + Normal [1,448,448,3] + Mask [1,448,448,1] + Scale [1,1,1,1]
  • CompiledModel GPU: 835 ops, all GPU-native, ~522 ms on Pixel 8a
  • Preprocessing: RGB / 255 (range [0, 1]), NCHW layout. DINOv2 applies ImageNet stats internally.
  • Original: microsoft/MoGe (CVPR'25 Oral) | MIT License

Sample app: moge/

v2: CLIP + DSINE + Whisper + SmolVLM

Choose a tag to compare

@john-rocky john-rocky released this 05 Apr 03:12

New Models

CLIP ViT-B/32 — Zero-Shot Classification

  • clip_image_encoder.tflite (352 MB) — image encoder, CompiledModel GPU
  • text_embeddings.bin (192 KB) + labels.txt — 96 pre-computed label embeddings

DSINE — Surface Normal Estimation (CVPR 2024)

  • dsine.tflite (282 MB) — encoder + decoder initial prediction, CompiledModel GPU

Whisper-tiny — Speech Recognition

  • whisper_encoder.tflite (33 MB) — mel spectrogram encoder, CompiledModel GPU
  • whisper_decoder.onnx (199 MB) — autoregressive text decoder, ONNX Runtime CPU
  • vocab.json + mel_filters.bin + config.json — tokenizer and preprocessing

SmolVLM-256M — Vision-Language Model

  • smolvlm_vision.tflite (357 MB) — SigLIP vision encoder + connector, CompiledModel GPU
  • smolvlm_decoder.onnx (515 MB) — SmolLM2 language model, ONNX Runtime CPU
  • embed_tokens.bin (108 MB) — token embedding matrix

Setup (large models)

Models are too large for APK assets. Push via adb:

adb push <model> /data/local/tmp/
adb shell "run-as <package> cp /data/local/tmp/<model> /data/data/<package>/files/"

v1 - YOLO11n + Depth Anything V2

Choose a tag to compare

@john-rocky john-rocky released this 03 Apr 13:15

Models

YOLO11n (Object Detection)

  • yolo11n.tflite (10 MB) — 18+ FPS on Pixel 8a GPU
  • Input: Float32 [1, 384, 384, 3] NHWC, normalized 0-1
  • Output: Float32 [1, 84, 3024] — bbox + 80 COCO classes
  • Converted from SavedModel (GPU-compatible, no PACK/SPLIT ops)

Depth Anything V2 (Monocular Depth)

  • depth_anything_v2_fp16w.tflite (47 MB) — FP16 weights
  • Input: Float32 [1, 392, 518, 3] NHWC, ImageNet normalized
  • Output: Float32 [1, 392, 518, 1] — relative inverse depth
  • Native Keras conversion (correlation 0.9995 with PyTorch)

Both models run on LiteRT CompiledModel GPU (ML Drift).

DepthAnything V2 Small — Native Keras TFLite (v2.0)

Choose a tag to compare

@john-rocky john-rocky released this 02 Apr 05:53

DepthAnything V2 Small — Native Keras NHWC Models

Native Keras conversion with near-perfect PyTorch quality. Replaces v1.0 onnx2tf models.

Model Input Speed (Pixel 8a) Quality (vs PyTorch) Size
depth_anything_v2_keras.tflite 518x518 9ms corr=0.9995 99 MB
depth_anything_v2_keras_392x518.tflite 392x518 ~7ms corr=0.9999 98 MB

What's New (vs v1.0)

  • Quality: corr=0.845 → 0.9995 (native Keras, no onnx2tf conversion artifacts)
  • Speed: 575ms → 9ms (CompiledModel ML Drift GPU, was old Interpreter API)
  • Compatibility: Works on Pixel 6+ with LiteRT 2.1.3 CompiledModel API

Usage

implementation("com.google.ai.edge.litert:litert:2.1.3")

val options = CompiledModel.Options(Accelerator.GPU)
options.gpuOptions = CompiledModel.GpuOptions(
    precision = CompiledModel.GpuOptions.Precision.FP32
)
val model = CompiledModel.create(context.assets, "depth_anything_v2_keras.tflite", options, null)

Place model in sample-app/app/src/main/assets/ and see sample-app/ for full example.

DepthAnything V2 Small - Model Files

Choose a tag to compare

@john-rocky john-rocky released this 01 Apr 09:39

Model Files

File Format Size Description
depth_anything_v2_nhwc.tflite TFLite FP32 NHWC 94 MB Recommended - GPU delegate compatible, 579ms on Pixel 10 Pro
depth_anything_v2.onnx ONNX FP32 94 MB ONNX Runtime CPU inference, 1437ms
depth_anything_v2_qualcomm.tflite TFLite FP32 NHWC 94 MB Qualcomm AI Hub pre-converted model, 571ms

Usage

Place model files in app/src/main/assets/ and build with Android Studio.

Conversion Pipeline (for NHWC TFLite)

  1. Patch PyTorch model (4D CLS concat, GELU tanh approximation)
  2. Export to ONNX (opset 18)
  3. Fix ConvTranspose kernel_shape, replace Erf→Tanh
  4. Simplify with onnxsim
  5. Convert with onnx2tf (-osd -coion)

See scripts/ for conversion code.