| --- |
| license: openrail++ |
| base_model: |
| - oyiya/SuperMat |
| - sd2-community/stable-diffusion-2-1 |
| pipeline_tag: image-to-image |
| tags: |
| - webgpu |
| - kaminos |
| - material-decomposition |
| - pbr |
| --- |
| |
| # SuperMat WebGPU Weights |
|
|
| Browser-loadable single-image SuperMat weights for the [Kaminos WebGPU Inference Kit](https://github.com/lyonsno/kaminos/tree/main/webgpu-inference-kit). |
|
|
| The [Material Photograph demo](https://lyonsno.github.io/kaminos/inference-kit/material-photo/) combines MoGe geometry and SuperMat albedo, roughness and metallic maps in an interactive relighting comparison. Its included examples can be explored before downloading weights. |
|
|
| ## Package |
|
|
| `f16/package.json` describes 13 resources, tensor layouts and SHA-256 hashes. The chunked route downloads 1,939,155,200 weight bytes. Sixteen-MiB chunks permit verification and upload without handling a whole large bundle at once. Whole-bundle copies are also included for the adapter's bundle route; they are alternate storage of the same tensors, not another model needed by the demo. |
|
|
| Use an immutable Hub commit in the package URL: |
|
|
| ```javascript |
| const weightsUrl = 'https://huggingface.co/BasinShapers/supermat-webgpu-weights/resolve/<commit>/f16/'; |
| const adapter = await createSuperMatAdapter({route, weightsUrl, weightLoading: 'chunks'}); |
| ``` |
|
|
| The browser adapter and its device/profile setup are in [Kaminos models/supermat](https://github.com/lyonsno/kaminos/tree/main/models/supermat). This repository contains weights, not executable model code. |
|
|
| ## Sources And Terms |
|
|
| SuperMat checkpoint: [oyiya/SuperMat](https://huggingface.co/oyiya/SuperMat/tree/91ffb8edaf259a1e5bbcbcd53b702ad44d70b709), `supermat.pth`, SHA-256 `21d171726e8170e1ed566985be68ef3c5009d762800926ad85c1f288747b7b29`. |
|
|
| VAE: [sd2-community/stable-diffusion-2-1](https://huggingface.co/sd2-community/stable-diffusion-2-1/tree/bb2154823665391b4fb29b0b9cf82a198964ee05), SHA-256 `a1d993488569e928462932c8c38a0760b874d166399b14414135bd9c42df5815`. |
|
|
| The source SuperMat project is [hyj542682306/SuperMat](https://github.com/hyj542682306/SuperMat), by Yijia Hong and collaborators. Its MIT notice is included as `LICENSE-SUPERMAT`. The Stable Diffusion 2.1 components retain the CreativeML Open RAIL++-M license and its use restrictions, included as `LICENSE-MODEL`. Use and redistribution of this package are subject to that license, including Section III and Attachment A; the MIT notice does not remove those restrictions. |
|
|
| Modification notice for all `f16/` weight bundles and chunks: Kaminos converted matrix/convolution weights from float32 to binary16, retained vectors and conditioning in float32, and packed the tensors for browser GPU loading. No fine-tuning or new learned parameters were introduced. The manifests retain the source revision and checksums and identify every packed file. |
|
|
