realvis51-inpaint-int8-onnx

realvis51-inpaint-int8-onnx

int8-quantized ONNX export of Uminosachi/realisticVisionV51_v51VAE-inpainting, built to run fully on-device (CPU) on Android phones.

Category: Edit — Photoreal inpainting -- fills masked regions cleanly. The first conv is kept in fp32: the 9-channel input mixes a large-range noisy latent with a 0/1 mask, and one shared int8 scale would crush the mask.

What was changed

This is a derivative work. The original weights were:

1. exported from PyTorch to ONNX (via optimum), then
2. dynamically quantized to int8 (quantize_dynamic, QInt8 weights) for the UNet and text encoder. The VAE is left in fp32 for output quality.

Nothing else was altered — no retraining, no merging. Shrinks a ~4.1 GB fp16 pipeline to ~1.3 GB so it fits comfortably in phone RAM.

Layout

unet/model.onnx (+ .onnx.data)
text_encoder/model.onnx (+ .onnx.data)
vae_decoder/model.onnx
vae_encoder/model.onnx
tokenizer/{vocab.json, merges.txt}

License

Inherited from the source model: creativeml-openrail-m. If that is CreativeML OpenRAIL-M, the Attachment A use-restrictions apply to you as well — they travel with the weights and cannot be removed by downstream redistribution. Please read them before use.

Credit for the underlying model belongs entirely to the original author.