Image Studio browser model resources

Browser-oriented model/runtime resources for Image Studio app 1.1.2. Photos are processed locally in the user's browser; this repository does not receive photos or provide inference. These are application-specific packs, not a generic Transformers model loading layout.

Four explicit choices

  • MODNet CPU and GPU share the same unchanged FP32 model.
  • HR + ViTMatte GPU uses the fixed HR graph stages and three-phase full-key row8 Loop MultiHeadAttention ViTMatte refinement, with CPU-owned phase seams.
  • HR + ViTMatte CPU uses the recorded HR stages and separate three-phase query-chunked ViTMatte derivative. CPU/GPU stage catalogs are not interchangeable.

The five ZIP_STORED packs preserve exact bytes, stream through the existing verified cache importer and avoid whole-archive RAM expansion. Every declared entry and transport has a byte count and SHA-256. MODNet transport is shared. The CPU high-quality route is slow and memory intensive. No silent quality/model fallback is provided. Interrupted downloads may require downloading a full pack again; Hub service/rate limits and cache eviction can affect availability.

Licenses and source

Model weights, model code, ONNX Runtime and its third-party components retain their own licenses. The app's MIT does not relicense them. LICENSES.md and all original license/notice texts remain included. Source06 model-transform sources retain the Swin MIT supplement and exclude the unused deployment adapter. SOURCE_AVAILABILITY.md provides the fixed Eigen source/patches and source offer. Historical review JSON records upstream export evidence, not current release bindings or a claim of independent native/compiler rebuild.

Immutable release identity

APP_RELEASE_PAIRING.json is the release index. REPLAY.md gives the canonical web/background-release placement. MODEL_FILES.json identifies the selected five packs and all entries. App version and model graph identity are separate. The index's review sourceId identifies its input review; later app-only updates bind their own Git commit/build-info/inventory while retaining this immutable input. Never put credentials in browser code or replace fixed resource URLs with main.

Source/model/runtime revision: 3888c6cf29bacaf2ec5a1d7ce4e5ad1ab1109823. App source: https://huggingface.co/otama3/image-studio-models/resolve/3888c6cf29bacaf2ec5a1d7ce4e5ad1ab1109823/releases/app-1.1.2-01791fc98d226aef/source/app-source.tar.gz Source SHA-256: 01791fc98d226aef1fa69b64815bfdf42c05e8a7d4b33d20bee2beeed3d4074c. Input: releases/app-1.1.2-input-297572aaed26f8b9/build-input.tar.gz at the SAME immutable revision as this release index. Input review sourceId: sha256:661dc1d2122f4bbd5730124e27fd771d1379e17690668f1f9ff3d3d7903453be.

Validation scope

Local strict build and distribution inventory identify this artifact. Existing synthetic-portrait/cancellation/cache evidence is bounded, not general quality, mobile-device or physical-print certification. Saved outputs should be inspected. Anonymous Hub delivery, Pages delivery and actual-device checks are recorded separately. No complete offline startup, service warranty or automatic compliance with identity-photo rules is promised.

This repository has clean file-only history under otama3. The old otama repositories are retained separately and are not active source download dependencies.

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for otama3/image-studio-models

Base model

Xenova/modnet
Quantized
(3)
this model