Modelpublic

thetechgeekko/latent-android-models

sourceHugging Faceotherupdated 21d agoView on Hugging Face

0likes149downloads

Model Card

LATENT Android models

This repository hosts model artifacts generated specifically for the LATENT Android camera app. It intentionally does not contain LATENT's film LUTs or Qualcomm QNN context binaries.

Each artifact below has its own immutable byte count, SHA-256, origin, and license. A repository-level license must not be inferred for a different file.

DA3 Small — LiteRT FP32, 672 × 896

da3_small_672x896_fp32.tflite is a LATENT-generated conversion of Depth Anything 3 Small for a fixed 3:4 Android input.

Property	Value
Size	105,877,404 bytes
SHA-256	`427e0cdb2ac60e3f5d1c1b1d2073ecff7469189859011f36017f1814773c1ba1`
Precision	FP32
Input	`args_0`, float32 NCHW `[1, 3, 896, 672]`
Output	`output_0`, float32 `[1, 1, 896, 672]`
Host status	Admitted on LiteRT CPU
Device status	Production promotion blocked pending NPU, quality, memory, latency, and thermal gates

The conversion retains upstream align_corners=true, keeps the head positional embedding, and replaces positional construction with an exactly checked separable x/y constant bake. It is not the community repository's published binary and must not be described as FP16.

Download the exact artifact with:

bash

hf download thetechgeekko/latent-android-models \
  da3_small_672x896_fp32.tflite

LATENT pins a repository commit, byte count, and SHA-256 before staging this artifact into an APK. See provenance/ for the guarded conversion, immutable input revisions, validation reports, and reproduction scripts.

Upstream and license

The base model is depth-anything/DA3-SMALL; source is ByteDance-Seed/depth-anything-3. Both identify Apache-2.0 as the license. The pinned community converter is litert-community/Depth-Anything-3-Small. Full revisions and hashes are in provenance/ADMISSION.json and provenance/MANIFEST.json.

See NOTICE.md for the modification and attribution record and LICENSE for the Apache License 2.0. This model card records engineering provenance; it is not a production-quality claim or legal opinion.

U2Netp d0 — LATENT TFLite FP16 conversion, 320 x 320

u2netp_d0_320_fp16.tflite is Clearstate's format conversion of the plain U2Netp salient-object checkpoint. It is not the separately trained u2net_portrait model.

Property	Value
Size	2,378,688 bytes
SHA-256	`357bd5214d6725efd88dc1578de5308f8a80c6df63ac3f35c2021e955f4a4bc5`
Precision	FP16 weights, float32 I/O
Input	float32 NHWC `[1, 320, 320, 3]`
Output	d0 saliency, float32 NHWC `[1, 320, 320, 1]`
Upstream architecture	`xuebinqin/U-2-Net`
Source checkpoint	plain `u2netp` obtained through `danielgatis/rembg`
Modification	pruned to d0 and converted ONNX to TFLite by Clearstate on 2026-07-19
License	Apache-2.0; see `LICENSE`

MiDaS v2.1 Small — legacy Intel TFLite FP32

midas_v2_1_small_256_fp32.tflite is the unmodified legacy app fallback. It is retained for exact clean-checkout reproduction while LATENT evaluates the smaller revision-pinned LiteRT Community conversion.

Property	Value
Size	66,338,288 bytes
SHA-256	`93d871071edff1218973ce25ee27ce95ccd20450c70a55e1b89efa3f5a772cdd`
Precision	FP32
Input	float32 NHWC `[1, 256, 256, 3]`
Output	relative inverse depth `[1, 256, 256]`
Upstream	Intel MiDaS v2.1 Small TFLite
Modification	none by Clearstate
License	MIT; see `MIDAS_MIT.txt`

Stock Google MediaPipe models

LATENT does not mirror Google's stock model files here. Clean checkouts fetch the exact versioned Google objects and verify these identities before staging:

App artifact	Official versioned object	Size	SHA-256	License evidence
`gesture_recognizer.task`	`gesture_recognizer/gesture_recognizer/float16/1`	8,373,440	`97952348cf6a6a4915c2ea1496b4b37ebabc50cbbf80571435643c455f2b0482`	Google MediaPipe model bundle; see the linked upstream model card in LATENT's lock file
`magic_touch.tflite`	`interactive_segmenter/magic_touch/float32/1`	6,227,884	`e24338a717c1b7ad8d159666677ef400babb7f33b8ad60c4d96db4ecf694cd25`	Apache-2.0 per Google's MagicTouch model card

The Qualcomm-generated QNN context binaries are device-specific build inputs. They remain local pending explicit authorization for a private external destination and are not licensed or presented here as generally reusable models.