ONNX export of emilianJR/epiCRealism (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 sharp, prompt-faithful output epiCRealism is known for — at a higher per-image cost (the UNet runs twice per step, over ~25 steps).

epiCRealism is a photoreal SD 1.5 fine-tune with broad subject coverage — strongest on environments, landscapes, architecture, interiors, and natural lighting.

Converted artifact. Training credit: emilianJR (epiCRealism).

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

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-epicrealism-cfg.ps1.

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.

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