Instructions to use litert-community/Idol with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- LiteRT
How to use litert-community/Idol with LiteRT:
# No code snippets available yet for this library. # To use this model, check the repository files and the library's documentation. # Want to help? PRs adding snippets are welcome at: # https://github.com/huggingface/huggingface.js
- Notebooks
- Google Colab
- Kaggle
Commit History
Today
-
Card: date the first sentence's two timings (2026-11-09 and 2026-11-10 JST)
Verified signaturee8f6b03 -
Card: first_weight_input counts inputs in numeric order of their names; match inputs by name
Verified signaturemanifest.json lists each signature's inputs in the order the .tflite file stores them (args_0, args_1, args_10, ...), while LiteRT 2.3.0 CompiledModel lists them in numeric order. Card text only; no model file changes.54317d5 -
Card and runner: the Python runner also runs the 262k/ form
Verified signaturepython/run_idol.py --models 262k ran on an M4 Max Mac (Metal) with the reference's mode and seed (--mode synthetic --seed 7): 70 s, and its splat of the armchair is byte for byte the first release's Mac output (4-view render PSNR against the author's default 35.51 dB, threshold 25 dB). The card and the runner's docstring now say how to run that form.4ce7762 -
Faster default form in the root; the first release's form moves to 262k/
Verified signatureTwo forms. The repo's root now holds a faster default form: a photo makes a splat of 65,536 Gaussians (2,048 anchors), every graph but BiRefNet runs on the GPU at FP16, the denoiser's joint blocks 1-22 share one graph (den_joint_mid.tflite, weights in den_joint_w.f32) and the Gaussian decoder's graphs 1-6 share another (gs_mid.tflite, weights in gs_w.f32); the app copies a block's weights in before each call, so each is compiled once per photo. The first release's form (262,144 Gaussians, a graph per block, GPU FP16 with fp32 accumulation) moves to 262k/ with its own manifest.json; its files are the same bytes as the root's at commit e7ee1226. On a Galaxy S26 the first photo after start on a cool phone took 158 s with the default form and 267 s with the 262k/ form. Four-view renders against the author's default on the phone (armchair / axolotl / succulent; threshold 25 dB): 34.41 / 32.26 / 34.26 dB PSNR with the default form, 46.59 / 42.34 / 41.43 dB with the 262k/ form.74728e5 -
Card: the denoiser did not compile for the Galaxy S26 NPU on the phone
Verified signatureCompiled on the Galaxy S26 for its HTP (LiteRT 2.3.0, QAIRT 2.50), one denoiser block did not finish in 15 minutes at the default optimization level and ran out of memory after 28 s at optimization_level=1. Graph files unchanged.e7ee122 -
Memory check counts the background removal call peak
Verified signatureThe app now counts BiRefNet's measured call peak (manifest run_peak_bytes) before compiling it, closes graphs or stops with "Not enough memory for background removal (bir)" instead of starting a call that does not fit. Card line updated. Graph files unchanged.0695ada -
Card: close other apps before a run (one photo peaks at about 5.4 GiB)
Verified signature3870440 -
Android sample: the viewer reads the first number of a repeated Content-Length and grows its buffer (it showed RangeError instead of the splat)
Verified signaturee7e7df2 -
Card, hero image, and the card's Kotlin block with its buffers closed
Verified signature842ba6a -
Idol on LiteRT: graphs, manifest, Python runner and Android sample
Verified signatureed66dc1 -
initial commit
Verified signaturea8917cc