xay2001/sd-turbo-android

SD-Turbo + TAESD for Android

Static ONNX export for local, prompt-guided image-to-image generation in Relight Lab. This is a converted research package, not a new trained model.

Contents

Download every file in sd-turbo-v1/, preserving filenames and keeping the ONNX graphs beside their external .data files. Weights are shared between the 384×384 and 512×512 graphs. The package manifest records exact sizes, SHA-256 hashes, model revisions and Euler noise parameters.

• SD-Turbo CLIP text encoder: FP16 weights, int64 tokens, FP32 output, sequence length 77 and embedding dimension 1024.
• SD-Turbo UNet: static 384/512 image resolutions (48/64 latent resolutions), FP16 weights and FP32 floating-point inputs/outputs.
• TAESD encoder and decoder: static 384/512, FP16 weights and FP32 inputs/outputs.
• Shared external weight files: text_encoder.data, unet.data, encoder.data, decoder.data.

Tokenizer files are bundled with the companion app. This is the SD2 tokenizer with PAD=0, not the original IC-Light SD1 tokenizer.

Pipeline

Center-crop an RGB input image, normalize it to [-1, 1], encode it with TAESD, add Gaussian noise at t499 (standard) or t699 (stronger edit), perform exactly one Euler denoising step with guidance scale 0, and decode with TAESD. TAESD latents already use the Stable Diffusion scaling convention; do not multiply them by the original VAE's 0.18215 factor.

Android runtime

The companion app uses CPU text encoding. The image encoder, UNet and decoder can use CPU or strict QNN HTP on a compatible Snapdragon device. HTP failures are reported rather than silently falling back to CPU. A successful export or a registered provider is not a device-execution benchmark.

The Android package requires device validation for NPU operator support, memory consumption and image quality. No sub-second Redmi inference claim is made by this model card. Separate local PyTorch/MPS experiments on an Apple M1 Pro measured approximately 0.55 seconds for a preloaded, 512×512 TAESD pipeline; that is not Android/NPU performance.

Limitations

This is general image-to-image generation, not instruction-perfect editing. Identities, objects, colors and details can change. TAESD is lossy, and image quality must be checked for the intended use. This package does not include DPR, does not replace the existing IC-Light package, and does not include a safety classifier; applications exposed to end users must implement appropriate safeguards.

Sources and terms

• SD-Turbo: https://huggingface.co/stabilityai/sd-turbo — source revision b261bac6fd2cf515557d5d0707481eafa0485ec2. Model use remains subject to Stability AI's applicable license and usage terms; consult the upstream card for commercial-use requirements.
• TAESD: https://huggingface.co/madebyollin/taesd — MIT, source revision 614f76814bbe30edbe2e627ace1c2234c81a2c0e.

Conversion does not grant additional rights to the upstream weights.