madebyollin
commited on
Commit
•
8ab04db
1
Parent(s):
2fa4db2
Update README.md
Browse files
README.md
CHANGED
@@ -7,5 +7,26 @@ inference: false
|
|
7 |
---
|
8 |
# SDXL-VAE-FP16-Fix
|
9 |
|
10 |
-
SDXL-VAE-FP16-Fix is
|
11 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
7 |
---
|
8 |
# SDXL-VAE-FP16-Fix
|
9 |
|
10 |
+
SDXL-VAE-FP16-Fix is the [SDXL VAE](https://huggingface.co/stabilityai/sdxl-vae), but modified to run in fp16 precision without generating NaNs.
|
11 |
+
|
12 |
+
```python
|
13 |
+
from diffusers import DiffusionPipeline, AutoencoderKL
|
14 |
+
import torch
|
15 |
+
|
16 |
+
pipe = DiffusionPipeline.from_pretrained("stabilityai/stable-diffusion-xl-base-0.9", torch_dtype=torch.float16, variant="fp16").to("cuda")
|
17 |
+
fixed_vae = AutoencoderKL.from_pretrained("madebyollin/sdxl-vae-fp16-fix").half().to("cuda")
|
18 |
+
|
19 |
+
prompt = "An astronaut riding a green horse"
|
20 |
+
latents = pipe(prompt=prompt, output_type="latent").images
|
21 |
+
|
22 |
+
for vae in (pipe.vae, fixed_vae):
|
23 |
+
for dtype in (torch.float32, torch.float16):
|
24 |
+
with torch.no_grad(), torch.cuda.amp.autocast(dtype=torch.float16, enabled=(dtype==torch.float16)):
|
25 |
+
print(dtype, "sdxl-vae" if vae == pipe.vae else "sdxl-vae-fp16-fix")
|
26 |
+
display(pipe.image_processor.postprocess(vae.decode(latents / vae.config.scaling_factor).sample)[0])
|
27 |
+
```
|
28 |
+
|
29 |
+
| VAE | Decoding in `float32` precision | Decoding in `float16` precision |
|
30 |
+
| --------------------- | ------------------------------- | ------------------------------- |
|
31 |
+
| SDXL-VAE | ✅ ![](./images/orig-fp32.png) | ⚠️ ![](./images/orig-fp16.png) |
|
32 |
+
| SDXL-VAE-FP16-Fix | ✅ ![](./images/fix-fp32.png) | ✅ ![](./images/fix-fp16.png) |
|