openjourney-cfg-onnx

Openjourney v4 (CFG, full quality) — ONNX

ONNX export of prompthero/openjourney-v4 (which ships its own VAE). No distillation LoRA — this is the full, non-distilled UNet driven with classifier-free guidance. SD 1.5 architecture, 512×512 native, Euler scheduler, CFG ≈ 7.5, ~25 steps.

This is the quality counterpart to the 4-step Hyper export. The Hyper variant is distilled for fast, CFG-free, 1–4 step sampling; it's great for previews and batch work but caps fidelity and prompt adherence. This non-distilled export, run with classifier-free guidance, a negative prompt, and a normal step budget, recovers the cinematic, prompt-faithful output Openjourney is known for — at a higher per-image cost (the UNet runs twice per step, over ~25 steps).

Openjourney v4 is a Midjourney-v4-style SD 1.5 fine-tune. No activator phrase — unlike v1/v3 it dropped the mdjrny-v4 style trigger; prompt it like any modern fine-tune.

Converted artifact. Training credit: PromptHero (Openjourney).

What this repo contains

model_index.json
feature_extractor/
scheduler/
text_encoder/
tokenizer/
unet/                   # Openjourney v4 UNet, non-distilled (no LoRA)
vae_decoder/            # Openjourney v4 bundled VAE
vae_encoder/

How it was produced

1. Load prompthero/openjourney-v4 via diffusers (uses its bundled VAE).
2. optimum-cli export onnx (no LoRA fusion step).

Exported at FP32 — the SD 1.5 VAE is fp16-fragile (it overflows and posterizes), so the quality export stays full precision.

Toolchain: optimum 1.24.0, diffusers 0.31.0, transformers 4.45.2, torch 2.4.x (CUDA 12.4). Conversion script: scripts/export-openjourney-cfg.ps1.

Inference notes

Classifier-free guidance runs the UNet twice per step (conditional + unconditional) and combines them as uncond + guidance · (cond − uncond). The negative prompt only takes effect when guidance > 1.

License

CreativeML OpenRAIL-M (SD 1.5 + Openjourney). License files included. By using this model you accept those terms.