Update README.md
Browse files
README.md
CHANGED
|
@@ -1,3 +1,77 @@
|
|
| 1 |
---
|
| 2 |
license: creativeml-openrail-m
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 3 |
---
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
---
|
| 2 |
license: creativeml-openrail-m
|
| 3 |
+
library_name: diffusers
|
| 4 |
+
tags:
|
| 5 |
+
- stable-diffusion
|
| 6 |
+
- sd-1.5
|
| 7 |
+
- hyper-sd
|
| 8 |
+
- onnx
|
| 9 |
+
- text-to-image
|
| 10 |
+
- 4-step
|
| 11 |
+
base_model: Lykon/DreamShaper
|
| 12 |
+
pipeline_tag: text-to-image
|
| 13 |
+
language:
|
| 14 |
+
- en
|
| 15 |
---
|
| 16 |
+
|
| 17 |
+
# DreamShaper + Hyper-SD (4-step) — ONNX
|
| 18 |
+
|
| 19 |
+
ONNX export of [Lykon/DreamShaper](https://huggingface.co/Lykon/DreamShaper) with the [ByteDance/Hyper-SD](https://huggingface.co/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**.
|
| 20 |
+
|
| 21 |
+
DreamShaper is Lykon's stylized SFW fine-tune — leans more illustrative / fantasy than AbsoluteReality, which is more photorealistic. Pick this one for D&D-style art, character portraits with painterly textures, and concept-art-leaning prompts.
|
| 22 |
+
|
| 23 |
+
This is a converted artifact, not a new model. All training credit belongs to Lykon (DreamShaper) and ByteDance (Hyper-SD).
|
| 24 |
+
|
| 25 |
+
## What this repo contains
|
| 26 |
+
|
| 27 |
+
```
|
| 28 |
+
model_index.json
|
| 29 |
+
feature_extractor/
|
| 30 |
+
scheduler/
|
| 31 |
+
text_encoder/
|
| 32 |
+
tokenizer/
|
| 33 |
+
unet/ # DreamShaper UNet + Hyper-SD-15 4-step LoRA fused in
|
| 34 |
+
vae_decoder/
|
| 35 |
+
vae_encoder/
|
| 36 |
+
```
|
| 37 |
+
|
| 38 |
+
`unet/model.onnx` is paired with `unet/model.onnx_data` (external-weights file).
|
| 39 |
+
|
| 40 |
+
## How it was produced
|
| 41 |
+
|
| 42 |
+
1. Load `Lykon/DreamShaper` via `diffusers` (bundled VAE).
|
| 43 |
+
2. Load `ByteDance/Hyper-SD/Hyper-SD15-4steps-lora.safetensors` via `peft`, `fuse_lora()` it into the UNet.
|
| 44 |
+
3. Save the fused pipeline to a temp directory.
|
| 45 |
+
4. `optimum-cli export onnx --model <temp> <output>`.
|
| 46 |
+
|
| 47 |
+
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-dreamshaper-hyper.ps1`](https://github.com/HeliosophLLC/DatumIngest/blob/main/scripts/export-dreamshaper-hyper.ps1).
|
| 48 |
+
|
| 49 |
+
## Inference notes
|
| 50 |
+
|
| 51 |
+
| Setting | Value |
|
| 52 |
+
|---|---|
|
| 53 |
+
| Scheduler | Euler |
|
| 54 |
+
| Steps | 4 |
|
| 55 |
+
| CFG / guidance scale | 1.0 |
|
| 56 |
+
| Negative prompt | Skip |
|
| 57 |
+
| Resolution | 512×512 native |
|
| 58 |
+
|
| 59 |
+
## License
|
| 60 |
+
|
| 61 |
+
CreativeML OpenRAIL-M (inherited from SD 1.5 + DreamShaper) + the Hyper-SD LoRA's OpenRAIL-M. Both license files are included in this repo. By using this model you accept those terms.
|
| 62 |
+
|
| 63 |
+
## Citation
|
| 64 |
+
|
| 65 |
+
```bibtex
|
| 66 |
+
@misc{lykon-dreamshaper,
|
| 67 |
+
author = {Lykon},
|
| 68 |
+
title = {DreamShaper},
|
| 69 |
+
howpublished = {\url{https://huggingface.co/Lykon/DreamShaper}}
|
| 70 |
+
}
|
| 71 |
+
@article{ren2024hypersd,
|
| 72 |
+
title = {Hyper-SD: Trajectory Segmented Consistency Model for Efficient Image Synthesis},
|
| 73 |
+
author = {Ren, Yuxi and others},
|
| 74 |
+
journal = {arXiv preprint arXiv:2404.13686},
|
| 75 |
+
year = {2024}
|
| 76 |
+
}
|
| 77 |
+
```
|