Heliosoph/sd-turbo-onnx

SD Turbo — ONNX

ONNX export of stabilityai/sd-turbo — Stability AI's Adversarial Diffusion Distillation (ADD) of Stable Diffusion 2.1. 512×512 native, designed to run at CFG = 1 in 1-4 inference steps (1 = design point for max speed, 4 = quality sweet spot).

This is a converted artifact, not a new model. All training credit belongs to Stability AI (Sauer, Lorenz, Blattmann, Rombach — ADD method, 2023).

What this repo contains

A standard ONNX diffusers pipeline layout:

model_index.json
feature_extractor/
scheduler/
text_encoder/           # OpenCLIP ViT-H/14 (1024-dim embeddings — NOT 768 like SD 1.5)
tokenizer/
unet/                   # SD 2.1 UNet, encoder_hidden_states dim 1024
vae_decoder/
vae_encoder/
LICENSE.md              # Stability AI Community License

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

How it was produced

optimum-cli export onnx --model stabilityai/sd-turbo <output> against a pinned toolchain. No LoRA fuse step — SD Turbo is the distilled base, not a fine-tune.

Toolchain: optimum 1.24.0, diffusers 0.31.0, transformers 4.45.2, torch 2.4.x (CUDA 12.4). Full conversion script: scripts/export-sd-turbo.ps1 in the Heliosoph repo.

Inference notes

Setting	Value
Architecture	Stable Diffusion 2.1 (single CLIP-H text encoder, 1024-dim)
Scheduler	Euler (or EulerAncestral) — ADD-distilled for short schedules
Steps	1 (design point) to 4 (quality sweet spot); beyond 4 returns diminishing gains
CFG / guidance scale	1.0 (no classifier-free guidance — Turbo was distilled without it)
Negative prompt	Skip — CFG = 1 ignores it
Resolution	512×512 native (other resolutions degrade fast)
VAE scale	0.18215 (SD 2.x family)
Latent shape	[1, 4, 64, 64]

SD Turbo vs SDXL Turbo vs the SD 1.5 Hyper family

Need	Pick
Fastest 512×512 from a clean Stability baseline	SD Turbo (this)
Better prompt adherence + composition at 512×512	SDXL Turbo (heavier — ~5× the disk and ~2× the VRAM)
512×512 with a specific photoreal / fantasy / illustrative fine-tune	SD 1.5 Hyper family (AbsoluteReality / DreamShaper / epiCRealism / RealisticVision / etc.)

SD Turbo is the right "small baseline" pick when you want Stability's canonical fast model rather than a community fine-tune, and when 512×512 is enough.

License

Stability AI Community License — LICENSE.md included in this repo and travels with redistribution.

⚠️ Commercial revenue threshold: This license is free for research, individuals, and commercial use below $1M annual revenue. Above the threshold, commercial use requires a separate Stability AI Enterprise License. By downloading you agree to these terms and to Stability's Acceptable Use Policy, which prohibits CSAM, non-consensual deepfakes, harassment, malware generation, and similar misuse. The AUP propagates with the model — your derivatives and downstream redistributions must impose the same policy.

Citation

@article{sauer2023adversarial,
  title   = {Adversarial Diffusion Distillation},
  author  = {Sauer, Axel and Lorenz, Dominik and Blattmann, Andreas and Rombach, Robin},
  journal = {arXiv preprint arXiv:2311.17042},
  year    = {2023}
}