Heliosoph/absolute-reality-hyper-onnx

AbsoluteReality + Hyper-SD (4-step) — ONNX

ONNX export of Lykon/AbsoluteReality with the ByteDance/Hyper-SD 4-step LoRA fused into the UNet. SD 1.5 architecture, 512×512 native, designed to run with the Euler scheduler at CFG = 1 in 4 inference steps.

This is a converted artifact, not a new model. All training credit belongs to Lykon (AbsoluteReality) and ByteDance (Hyper-SD).

What this repo contains

A standard ONNX diffusers pipeline layout:

model_index.json
feature_extractor/
scheduler/
text_encoder/
tokenizer/
unet/                   # AbsoluteReality UNet + Hyper-SD-15 4-step LoRA fused in
vae_decoder/
vae_encoder/

unet/model.onnx is paired with unet/model.onnx_data (external-weights file). Both must be downloaded.

How it was produced

1. Load Lykon/AbsoluteReality via diffusers (uses its bundled VAE — no separate VAE pairing needed).
2. Load ByteDance/Hyper-SD/Hyper-SD15-4steps-lora.safetensors via peft and call fuse_lora() on the UNet.
3. Save the fused pipeline to a temp directory.
4. optimum-cli export onnx --model <temp> <output>.

Toolchain: optimum 1.24.0, diffusers 0.31.0, transformers 4.45.2, torch 2.4.x (CUDA 12.4), peft latest. Full conversion script: scripts/export-absolute-reality-hyper.ps1 in the DatumIngest repo.

Inference notes

Setting	Value
Scheduler	Euler (sample/x0 prediction is not required — 4-step Hyper is epsilon)
Steps	4
CFG / guidance scale	1.0 (no classifier-free guidance)
Negative prompt	Skip — CFG = 1 ignores it
Resolution	512×512 native (768×768 works, smaller resolutions degrade fast)

License

This export inherits CreativeML OpenRAIL-M from the base SD 1.5 lineage and AbsoluteReality. The Hyper-SD LoRA also ships under OpenRAIL-M (ByteDance). Both LICENSE-*.md files are included in this repo and travel with redistribution. By using this model you accept those terms — see the included license files for acceptable-use clauses.

Citation

If you use this in academic work, cite both the base model and the distillation method:

@misc{lykon-absolutereality,
  author = {Lykon},
  title  = {AbsoluteReality},
  howpublished = {\url{https://huggingface.co/Lykon/AbsoluteReality}}
}
@article{ren2024hypersd,
  title   = {Hyper-SD: Trajectory Segmented Consistency Model for Efficient Image Synthesis},
  author  = {Ren, Yuxi and others},
  journal = {arXiv preprint arXiv:2404.13686},
  year    = {2024}
}