File size: 1,596 Bytes
e82aa0e
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
---
license: openrail++
tags:
- text-to-image
- PixArt-Σ
---
THIS IS A REDISTRIBUTION OF PIXART-Σ-XL-256x256

### 🧨 Diffusers 
> [!IMPORTANT]  
> Make sure to upgrade diffusers to >= 0.28.0:
> ```bash
> pip install -U diffusers --upgrade
> ```
> In addition make sure to install `transformers`, `safetensors`, `sentencepiece`, and `accelerate`:
> ```
> pip install transformers accelerate safetensors sentencepiece
> ```
> For `diffusers<0.28.0`, check this [script](https://github.com/PixArt-alpha/PixArt-sigma#2-integration-in-diffusers) for help.

To just use the base model, you can run:

```python
import torch
from diffusers import Transformer2DModel, PixArtSigmaPipeline

device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
weight_dtype = torch.float16

pipe = PixArtSigmaPipeline.from_pretrained(
    "dattrong/pixart-sigma-512", 
    torch_dtype=weight_dtype,
    use_safetensors=True,
)
pipe.to(device)

# Enable memory optimizations.
# pipe.enable_model_cpu_offload()

prompt = "A small cactus with a happy face in the Sahara desert."
image = pipe(prompt).images[0]
image.save("./catcus.png")
```

When using `torch >= 2.0`, you can improve the inference speed by 20-30% with torch.compile. Simple wrap the unet with torch compile before running the pipeline:
```py
pipe.transformer = torch.compile(pipe.transformer, mode="reduce-overhead", fullgraph=True)
```

If you are limited by GPU VRAM, you can enable *cpu offloading* by calling `pipe.enable_model_cpu_offload`
instead of `.to("cuda")`:

```diff
- pipe.to("cuda")
+ pipe.enable_model_cpu_offload()
```