---
license: openrail++
library_name: onnx
pipeline_tag: image-to-image
tags:
- stable-diffusion-xl
- onnx
- unity
- unity-inference-engine
- sentis
- localimagegen
- image-to-image
base_model: stabilityai/stable-diffusion-xl-refiner-1.0
---

# Stable Diffusion XL Refiner 1.0 — corrected ONNX export

This repository contains a standard-operator, external-data ONNX export of the
official [Stable Diffusion XL Refiner 1.0](https://huggingface.co/stabilityai/stable-diffusion-xl-refiner-1.0)
UNet. It is the full-precision source used to produce the optimized Unity
Inference Engine model for
[LocalImageGen](https://huggingface.co/dtaddis/localimagegen-sdxl-refiner-fp16).

> **Correction notice (26 July 2026):** the UNet in the repository's earlier
> revisions produced incorrectly low-amplitude noise predictions and must not
> be used. The current `main` branch replaces it with a fresh export from the
> official Stability AI checkpoint. Existing clones should pull the current
> revision and replace both files in `unet/`.

## Which repository should I download?

- **LocalImageGen customer:** use LocalImageGen's Model Manager. It downloads
  the tested FP16 `.sentis` artifact from
  [`dtaddis/localimagegen-sdxl-refiner-fp16`](https://huggingface.co/dtaddis/localimagegen-sdxl-refiner-fp16).
- **ONNX conversion or engine integration:** use this repository's
  `unet/model.onnx` and `unet/model.onnx_data`.

The Refiner is an optional second stage. It consumes low-noise latents from
SDXL Base; it is not a standalone text-to-image model.

## Export details

- Upstream: `stabilityai/stable-diffusion-xl-refiner-1.0`
- Pinned upstream revision:
  `5d4cfe854c9a9a87939ff3653551c2b3c99a4356`
- Precision: FP32 ONNX source weights
- ONNX opset: 14
- Operators: standard ONNX operators; no ONNX Runtime custom attention fusion
- External data filename: `model.onnx_data`
- Dynamic axes: batch, latent height/width, and text sequence length
- Inputs: `sample`, `timestep`, `encoder_hidden_states`, `text_embeds`,
  `time_ids`
- Output: `out_sample`

The included `export_sdxl_refiner_unet.py` documents the reproducible export
wrapper and command-line interface. The tokenizer, text encoder, scheduler,
and VAE files retained in this repository are shared SDXL pipeline components;
LocalImageGen's Refiner profile reuses those components from its SDXL Base
profile and downloads only the optimized Refiner UNet.

Keep `model.onnx` and `model.onnx_data` together in the same directory without
renaming either file.

## Integrity

| File | Bytes | SHA-256 |
|---|---:|---|
| `unet/model.onnx` | 2,858,274 | `47957faab78710349a73de650b1c6590b98f950e4c384903a8c7264fc20c09ef` |
| `unet/model.onnx_data` | 9,038,106,640 | `b60900c7457ccf354cd72e56a34c0dad25911a900a00cad2e84be7962875f28e` |

## Validation

The corrected export was converted to FP16 with Unity 6000.5.5f1 and Unity
Inference Engine 2.6.1. LocalImageGen's SDXL Base + Refiner pipeline was tested
under D3D11 at 512×512 and 768×768 with a clean four/five-step low-noise
handoff and no mosaic artifacts.

## License

CreativeML OpenRAIL++-M. See `LICENSE.md`. This is a transformed inference
artifact, not a newly trained model; all model credit belongs to Stability AI
and the upstream contributors. Downstream users are responsible for reviewing
and complying with the license and applicable laws.