Update README.md
Browse files
README.md
CHANGED
|
@@ -1,3 +1,62 @@
|
|
| 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 |
+
- photorealistic
|
| 12 |
+
base_model: SG161222/Realistic_Vision_V6.0_B1_noVAE
|
| 13 |
+
pipeline_tag: text-to-image
|
| 14 |
+
language:
|
| 15 |
+
- en
|
| 16 |
---
|
| 17 |
+
|
| 18 |
+
# Realistic Vision V6 + Hyper-SD (4-step) — ONNX
|
| 19 |
+
|
| 20 |
+
ONNX export of [SG161222/Realistic_Vision_V6.0_B1_noVAE](https://huggingface.co/SG161222/Realistic_Vision_V6.0_B1_noVAE) paired with [stabilityai/sd-vae-ft-mse](https://huggingface.co/stabilityai/sd-vae-ft-mse) and the [ByteDance/Hyper-SD](https://huggingface.co/ByteDance/Hyper-SD) 4-step LoRA fused into the UNet. SD 1.5 architecture, 512×512 native, Euler scheduler, CFG = 1, **4 steps**.
|
| 21 |
+
|
| 22 |
+
Realistic Vision V6 is the photorealistic-portrait flagship of the SD 1.5 ecosystem. Trained on a narrow distribution (people, portraits, photography aesthetics), which is exactly *why* it's more stable across seeds than base SD 1.5 for those subjects.
|
| 23 |
+
|
| 24 |
+
> **Heads-up:** Realistic Vision is more NSFW-permissive than the other Hyper variants in this collection. Pair with content filters if that matters for your application.
|
| 25 |
+
|
| 26 |
+
Converted artifact. Training credit: SG161222 (Realistic Vision), Stability AI (sd-vae-ft-mse), ByteDance (Hyper-SD).
|
| 27 |
+
|
| 28 |
+
## What this repo contains
|
| 29 |
+
|
| 30 |
+
```
|
| 31 |
+
model_index.json
|
| 32 |
+
feature_extractor/
|
| 33 |
+
scheduler/
|
| 34 |
+
text_encoder/
|
| 35 |
+
tokenizer/
|
| 36 |
+
unet/ # RV6 UNet + Hyper-SD-15 4-step LoRA fused in
|
| 37 |
+
vae_decoder/ # sd-vae-ft-mse (RV6 ships without VAE — paired here)
|
| 38 |
+
vae_encoder/
|
| 39 |
+
```
|
| 40 |
+
|
| 41 |
+
## How it was produced
|
| 42 |
+
|
| 43 |
+
1. Load `SG161222/Realistic_Vision_V6.0_B1_noVAE` via `diffusers`.
|
| 44 |
+
2. Replace the (missing) VAE with `stabilityai/sd-vae-ft-mse` — the SD 1.5 community-standard fine-tuned VAE.
|
| 45 |
+
3. Load `ByteDance/Hyper-SD/Hyper-SD15-4steps-lora.safetensors` via `peft`, `fuse_lora()` into UNet.
|
| 46 |
+
4. `optimum-cli export onnx`.
|
| 47 |
+
|
| 48 |
+
Toolchain: `optimum 1.24.0`, `diffusers 0.31.0`, `transformers 4.45.2`, `torch 2.4.x` (CUDA 12.4). Conversion script: [`scripts/export-realistic-vision-hyper.ps1`](https://github.com/HeliosophLLC/DatumIngest/blob/main/scripts/export-realistic-vision-hyper.ps1).
|
| 49 |
+
|
| 50 |
+
## Inference notes
|
| 51 |
+
|
| 52 |
+
| Setting | Value |
|
| 53 |
+
|---|---|
|
| 54 |
+
| Scheduler | Euler |
|
| 55 |
+
| Steps | 4 |
|
| 56 |
+
| CFG / guidance scale | 1.0 |
|
| 57 |
+
| Negative prompt | Skip |
|
| 58 |
+
| Resolution | 512×512 native (best results); 768×768 OK |
|
| 59 |
+
|
| 60 |
+
## License
|
| 61 |
+
|
| 62 |
+
CreativeML OpenRAIL-M (SD 1.5 + Realistic Vision + Hyper-SD). License files included. By using this model you accept those terms.
|