nexus-sd15-inpaint-mnn-int8

Nexus Realistic Vision v5.1 Inpaint MNN INT8

Complete MNN Stable Diffusion 1.5 pack for offline Android text-to-image, image-to-image, and true 9-channel inpainting. The pack is intended for the Nexus Android app and its MNN 3.4.1 diffusion runtime.

Contents and provenance

• unet.mnn and unet.mnn.weight are an MNN INT8 conversion of the UNet from SG161222/Realistic_Vision_V5.1_noVAE revision 1e9f017a7b1eaefb63a1900ea6c5953d2739fd21.
• inpaint_unet.mnn and inpaint_unet.mnn.weight are an MNN INT8 conversion of the trained 9-channel UNet from stablediffusionapi/realistic-vision-v51-inpa revision 842b8e36533db07f4f7d421001d423f1788f4a0e.
• The tokenizer, text encoder, VAE decoder, scheduler data, and VAE encoder are the SD1.5-compatible MNN components used by the Nexus mobile runtime.

The dedicated inpaint UNet uses the Nexus MNN diffusion contract:

• sample: [2, 9, 64, 64] float32
• timestep: [1] int32
• encoder_hidden_states: [2, 77, 768] float32
• out_sample: [2, 4, 64, 64] float32

The 9 input channels are the noised latent, masked-image latent, and downsampled mask. Both UNets use 8-bit asymmetric weight quantization.

Pixel 8 Pro validation

The exact four UNet artifacts in this revision were validated through the production Nexus Studio UI on a Pixel 8 Pro at 512x512 and seed 424242:

• Text-to-image: 30/30 MNN denoise calls in 197.1 seconds produced a detailed, prompt-aligned blue bird on a branch.
• Inpaint: 30/30 dedicated 9-channel calls in 200.8 seconds produced a recognizable blue bird inside the white mask while preserving unmasked pixels.
• Both jobs completed on backend MNN without an app-process restart.

The inpaint MNN output has cosine similarity 0.99999428 against the source PyTorch UNet for a fixed test tensor (MAE=0.002806, RMSE=0.003535).

Integrity

SHA256SUMS contains hashes for every runtime file. The four model hashes that distinguish this revision are:

License

This model inherits CreativeML OpenRAIL-M restrictions from its upstream Stable Diffusion derivatives. Review the upstream model cards before redistribution or deployment:

• https://huggingface.co/SG161222/Realistic_Vision_V5.1_noVAE
• https://huggingface.co/stablediffusionapi/realistic-vision-v51-inpa

The pack is optimized for mobile correctness and quality, not minimal storage. Peak memory and speed vary by Android device and MNN backend.