| --- |
| license: openrail++ |
| library_name: diffusers |
| tags: |
| - stable-diffusion-xl |
| - sdxl |
| - lightning |
| - onnx |
| - text-to-image |
| - 4-step |
| - photorealistic |
| - fantasy |
| base_model: RunDiffusion/Juggernaut-XL-Lightning |
| pipeline_tag: text-to-image |
| language: |
| - en |
| --- |
| |
| # JuggernautXL Lightning — ONNX (fp32) |
|
|
| ONNX export of [RunDiffusion/Juggernaut-XL-Lightning](https://huggingface.co/RunDiffusion/Juggernaut-XL-Lightning) — an SDXL fine-tune + Lightning distillation. SDXL architecture (1024×1024 native, dual text encoders), runs in **4 inference steps** at CFG = 1. |
|
|
| JuggernautXL is one of the most-downloaded SDXL fine-tunes for photorealistic and fantasy-character art. Lightning distillation brings it down to 4 steps (vs SDXL's native 30+) at minimal quality cost. Pick this when SDXL-Turbo's 1-step output isn't quality-stable enough but you still want fast inference. |
|
|
| Converted artifact. Training credit: RunDiffusion (Juggernaut-XL-Lightning). |
|
|
| ## What this repo contains |
|
|
| A standard SDXL ONNX diffusers pipeline layout: |
|
|
| ``` |
| model_index.json |
| scheduler/ |
| text_encoder/ # CLIP-L (OpenAI) |
| text_encoder_2/ # OpenCLIP-G |
| tokenizer/ |
| tokenizer_2/ |
| unet/ # JuggernautXL-Lightning UNet (~2.6B params) |
| vae_decoder/ |
| vae_encoder/ |
| ``` |
|
|
| `unet/model.onnx` has external weights in `unet/model.onnx_data`. Same for `text_encoder_2/` if present. |
|
|
| ## How it was produced |
|
|
| `optimum-cli export onnx --model RunDiffusion/Juggernaut-XL-Lightning --task text-to-image --library diffusers --device cuda --no-post-process <output>` |
|
|
| via [`scripts/export-batch-onnx.ps1 -Models juggernaut-xl-lightning`](https://github.com/HeliosophLLC/DatumIngest/blob/main/scripts/export-batch-onnx.ps1). |
|
|
| Toolchain: `optimum 1.24.0`, `diffusers 0.31.0`, `transformers 4.45.2`, `torch 2.4.x` (CUDA 12.4). |
|
|
| ## Why fp32, not fp16 |
|
|
| The `optimum 1.24` + `torch 2.4` + opset-14 fp16 export path produces a numerically broken UNet for SDXL-class models (all-NaN noise predictions from valid conditioning). The fp32 export works correctly. Revisit fp16 if/when a confirmed working toolchain ships. |
|
|
| Practical impact: ~13 GB on disk vs ~6.5 GB at fp16. VRAM budget at inference: ~7–8 GB with the standard scheduler. |
|
|
| ## Inference notes |
|
|
| | Setting | Value | |
| |---|---| |
| | Scheduler | Euler (epsilon prediction — Lightning 4-step is **not** sample-prediction; the 2-step variant is, but this is the 4-step) | |
| | Steps | 4 | |
| | CFG / guidance scale | 1.0 | |
| | Negative prompt | Skip | |
| | Resolution | 1024×1024 native | |
|
|
| ## License |
|
|
| **CreativeML OpenRAIL++-M** inherited from `stabilityai/stable-diffusion-xl-base-1.0` + JuggernautXL's terms. The included `LICENSE-creativeml-openrail-pp-m.md` travels with redistribution. By using this model you accept those terms — review the use-based restrictions in section II of the license. |
