SDXL VAE encoder — fp16 fix, compiled for Core ML

VAEEncoderSDXLFix.mlmodelc — an fp16-safe VAE encoder for Stable Diffusion XL, compiled for Core ML (Apple Silicon).

Why this exists

Apple's coreml-stable-diffusion-xl-base ships a VAE encoder that produces NaN for any input (apple/ml-stable-diffusion#297). The decoder is fine, so text-to-image works — but image-to-image is impossible, because encoding the source image is the first step.

Swapping this encoder into the pipeline folder fixes that.

Use

Replace VAEEncoder.mlmodelc inside a compiled SDXL pipeline folder:

rm -rf <pipeline>/VAEEncoder.mlmodelc
cp -R VAEEncoderSDXLFix.mlmodelc <pipeline>/VAEEncoder.mlmodelc

Provenance and license

Converted from madebyollin/sdxl-vae-fp16-fix (MIT) — an SDXL VAE retuned to run in fp16 without overflowing. This repository redistributes it as a Core ML format conversion, under the same MIT license. All credit for the fix itself goes to madebyollin.

Converted for AML — Apple ML operators for TouchDesigner, which fetches it on demand for SDXL image-to-image.