Cadryl — your images, your perspective; approved English brand artwork

Cadryl · LiteRT models by task

An Android studio for importing images, reviewing model proposals, correcting annotations and exporting datasets. Historical task catalogue: 48 model groups · 99 LiteRT exports · 93 image / 6 audio. Additional conversions are listed separately below.

Quick card

Field Reference
Use Choose a specialized inference model for one documented task.
Expected input An image and its exact export configuration; audio uses a separate native frontend.
Example Import a model/config pair, try a representative image, review and preserve human annotations.
Limits Historical inference collection, plus the separately documented trainable-head exports below; benchmark scores do not establish accuracy on your photos or complete Release/device qualification.
Contract Per-export catalogue · Import guide

Choose a variant · Resource navigation · Update log

15 task comparisons · All variants · JSON index

Task group Available uses
Classification Waste, materials, PCB1 anomalies, sharpness, clothing, fruit freshness, gestures, signed alphabet, indoor/outdoor scenes and Kuzushiji.
Detection Helmets/heads/people and document text blocks.
Segmentation Pothole foreground masks.
Points Four document corners.
Audio Spoken digits, with the exact audio frontend.

Added 14 November 2026 · FireViewer trainable-head LiteRT exports

Two existing FP32 conversions, originally published on 12 November, are now grouped here with their unchanged model bytes, Android settings, provenance and cloud execution receipts.

Model Input Graph Matching Cadryl settings Scope
FireViewer YOLO11M · fire / smoke · 928 px 928 × 928 RGB model.tflite model.tflite.model_config.json Trainable final projections; frozen feature extractor
FireViewer ConvNeXt-Tiny · six alert categories · 920 px 920 × 920 RGB model.tflite model.tflite.model_config.json Trainable final projections; frozen feature extractor

Both graphs retain real train, infer, save and restore signatures. Their artifact manifests record the original parity, parameter-update and checkpoint/reopen tests. These are cloud technical checks; native Cadryl/Android qualification and post-adaptation accuracy remain unmeasured. Use the exact settings shipped next to each graph.

The YOLO export comes from fireviewer/yolo11m-fire-smoke at ecf4ca351b5dbd4ab84a92c0402e6f45f8001de9. It is not a conversion of the YOLO11 A100 resume stopped on 14 November. The ConvNeXt export comes from fireviewer/convnext-tiny-alert-categories-v2 at 82d7af7e8f1e0126309f26b19c285b26909a5acb.

The source model cards retain the applicable licences and original label names. No training, fresh conversion, new accuracy evaluation or paid compute was run for this relocation.

Added 27 October 2026 · FireViewer YOLO26m checkpoint 54

New three-class FP32 inference export for Cadryl Play, kept separate from the historical task catalogue and its qualification counts.

RGB NCHW [1,3,768,768], centered letterbox, /255; output [1,300,6] contains XYXY in input pixels, confidence and class ID. Classes: probable_fire_unspecified, visible_flame, visible_smoke. PyTorch parity passed on 14 validation images; LiteRT 2.2.0 CPU execution also passed. Inference only; no physical Android qualification is claimed. Original epoch54 weights, all existing variants and their evidence are preserved.

Use a matching pair

  1. Retrieve the chosen .tflite and its exact .tflite.model_config.json at one immutable revision.
  2. Import both into Cadryl; preserve labels, preprocessing, tensor outputs and thresholds.
  3. Try a representative image, review proposals and preserve corrected human annotations.

Application · Downloads · Getting started · About Unicorn Who Dev

Evidence and remaining limits

The existing execution receipt reports 93 cloud image executions and 66/93 image variants executed on ARM64 Debug, with synthetic device inputs. Complete physical-device coverage and Release APK qualification remain open. These are execution checks, not representative photo accuracy or phone timing benchmarks.

The comparisons retain full-test scores, recorded intervals and exact source hashes. Historical scores without a public common-test identity are not ranked. Fashion-MNIST at 460 px remains deferred. No training, new evaluation or paid compute is started by this update.

General-purpose conversions · Audio frontend · Hashes · Attributions. Original corpus/checkpoint restrictions remain applicable; no new blanket weight licence is assigned.

Likes are always welcome! ❤️ If these resources help you, a little heart makes our day and helps others discover the work. Thank you!

Downloads last month
2,318
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Collection including unicornwhodev/lite-rt_models_cadrylpro