SDXL-Turbo-LiteRT

New LiteRT conversion of stabilityai/sdxl-turbo. Powered by Stability AI.

Source: stabilityai/sdxl-turbo, revision 71153311d3dbb46851df1931d3ca6e939de83304. Output: 512 × 512, 4 sampling step(s).

Contents

Float32 graphs for CLIP text encoding, UNet denoising, the Euler ancestral sampler and VAE decoding. Graphs are streamed one at a time. The scheduler and default step count are fixed by the manifest. The bundle is portable and is not Google Tensor TPU AOT-compiled. Integer tensor I/O is preserved.

Weights retain the Stability AI Community License in LICENSE.md, including its use and commercial-use conditions. They are not relicensed under Apache 2.0. The newly added companion scripts use CODE_LICENSE.

manifest.json enumerates the tensor execution plan and exact graph/tokenizer/constant hashes. Large source Diffusers checkpoints, download caches, credentials and compiler working files are not included.

Run on CPU

Use Python 3.12 in a virtual environment. Download the repository with hf download frankdu/SDXL-Turbo-LiteRT --local-dir SDXL-Turbo-LiteRT, then:

cd SDXL-Turbo-LiteRT
python3.12 -m venv .venv
. .venv/bin/activate
pip install -r requirements.txt
python run_portable.py --model-dir . --prompt "a red apple on a wooden table, studio lighting" --seed 42 --out output.png

These are diffusion graph pipelines, not language-model .litertlm checkpoints. A standalone graph call with random inputs does not generate a meaningful image; use the full pipeline. Android integrations must run off the main thread, release tensor buffers/models between stages and use a shared LiteRT Environment. Source/model graph layout is preserved for reuse by the Research Lab Android application.

Validation and limitations

The actual model graphs completed a full desktop LiteRT 2.3.0 CPU generation check on Linux x86_64. See validation.json. The sample was visually checked as a recognizable apple. This is a smoke check, not a comparative quality or speed benchmark. SD/SDXL source-parity checks use export-time PyTorch noise; normal inference uses the runner's documented seeded stream. Different runtimes can generate different samples with the same integer seed.

Android CPU/GPU execution, phone memory requirements and Google Tensor TPU execution have not been validated for this repository. Successful desktop execution does not establish phone performance. Graph loads, compilation and host tensor copies can be expensive. Use the upstream model cards for intended use, limitations and applicable restrictions.

Attribution and license

The original model and conversion authors retain their copyrights. This distribution retains the source license and notices; conversion/mirroring does not change model license terms. See the included LICENSE/NOTICE files and SOURCE_MODEL_CARD.md. Added helper code is identified separately where applicable.