Asset repository for the Obsidian plugin Local Image Generator (https://github.com/johannes-kaindl/local-image-generator). The plugin's built-in engine downloads these files only after the user clicks "Download", verifies each file against a SHA-256 pinned in the plugin release, and stores them outside the vault.

Two models are hosted here, chosen from a dropdown in the plugin: SD-Turbo (≈ 2.5 GB, 512×512 only) and SDXL-Turbo (≈ 7.0 GB, up to 1024×1024). Both are the plugin's own ONNX conversion of the respective official Stability AI weights — no third-party conversion is used.

Path	What it is
`sd-turbo/text_encoder/model.onnx`	CLIP text encoder, fp16 weights, fp32 inputs/outputs
`sd-turbo/unet/model.onnx`	UNet, fp16 weights, fp32 inputs/outputs (`timestep` is a 0-d float32 scalar)
`sd-turbo/vae_decoder/model.onnx`	VAE decoder, fp16 weights, fp32 inputs/outputs
`sd-turbo/tokenizer/vocab.json`, `merges.txt`	CLIP BPE tokenizer data
`sd-turbo/LICENSE.md`, `sd-turbo/NOTICE.md`	Stability AI Community License and attribution (SD-Turbo)
`sdxl-turbo/text_encoder/model.onnx`	CLIP-L text encoder (primary), fp16 weights, fp32 inputs/outputs
`sdxl-turbo/text_encoder_2/model.onnx`	OpenCLIP-bigG text encoder (secondary), fp16 weights, fp32 inputs/outputs
`sdxl-turbo/unet/model.onnx` + `unet/unet_*.onnx_data`	UNet (≈ 5.1 GB), fp16 weights, split across 13 external-data buckets — the model exceeds ONNX's and the browser's single-buffer limits
`sdxl-turbo/vae_decoder/model.onnx`	VAE decoder, fp32 weights and inputs/outputs (kept fp32 — SDXL's activations exceed fp16's range under WebGPU, which produced a silent pure-black image; ≈ 198 MB vs. ≈ 99 MB for the fp16 build)
`sdxl-turbo/tokenizer/vocab.json`, `merges.txt`	Primary (CLIP-L) BPE tokenizer data
`sdxl-turbo/tokenizer_2/vocab.json`, `merges.txt`	Secondary (OpenCLIP-bigG) BPE tokenizer data
`sdxl-turbo/LICENSE.md`, `sdxl-turbo/NOTICE.md`	Stability AI Community License and attribution (SDXL-Turbo)
`runtime/ort-<version>/ort-wasm-simd-threaded.asyncify.wasm`	ONNX Runtime Web (MIT), the exact build the plugin version links against

Provenance

The ONNX files are the plugin's own conversion of the official stabilityai/sd-turbo and stabilityai/sdxl-turbo weights: optimum ONNX export, then onnxruntime's float16 conversion with keep_io_types=True (SDXL-Turbo's UNet additionally split across external-data buckets, see above). The script is tools/convert/convert_model.py in the plugin repository; no third-party conversion is used for either model. Powered by Stability AI.

License

The model files are redistributed under the Stability AI Community License — see SD-Turbo's license and SDXL-Turbo's license (research, non-commercial and limited commercial use free of charge — see the respective license for the revenue threshold; each model's own LICENSE.md/NOTICE.md under its subdirectory above is the one that actually applies to those files). The ONNX Runtime WASM is MIT. The plugin itself is AGPL-3.0-or-later.

johannes-kaindl/local-image-generator-models

Local Image Generator — model assets

Contents

Provenance

License

johannes-kaindl/
local-image-generator-models