justinpinkney commited on
Commit
383b0d8
1 Parent(s): 9c0622a

initial test

Browse files
Files changed (3) hide show
  1. app.py +91 -4
  2. requirements.txt +1 -0
  3. unsafe.png +0 -0
app.py CHANGED
@@ -1,7 +1,94 @@
1
  import gradio as gr
 
 
2
 
3
- def greet(name):
4
- return "Hello " + name + "!!"
5
 
6
- iface = gr.Interface(fn=greet, inputs="text", outputs="text")
7
- iface.launch()
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
  import gradio as gr
2
+ import torch
3
+ from PIL import Image
4
 
5
+ from lambda_diffusers import StableDiffusionImageEmbedPipeline
 
6
 
7
+ def main(
8
+ input_im,
9
+ scale=3.0,
10
+ n_samples=4,
11
+ seed=0,
12
+ steps=25,
13
+ ):
14
+ generator = torch.Generator(device=device).manual_seed(seed)
15
+
16
+ images_list = pipe(
17
+ n_samples*[input_im],
18
+ guidance_scale=scale,
19
+ num_inference_steps=steps,
20
+ generator=generator,
21
+ )
22
+
23
+ images = []
24
+ safe_image = Image.open(r"unsafe.png")
25
+ for i, image in enumerate(images_list["sample"]):
26
+ if(images_list["nsfw_content_detected"][i]):
27
+ images.append(safe_image)
28
+ else:
29
+ images.append(image)
30
+ return images
31
+
32
+
33
+ description = \
34
+ """Generate variations on an input image using a fine-tuned version of Stable Diffision.
35
+ Trained by [Justin Pinkney](https://www.justinpinkney.com) ([@Buntworthy](https://twitter.com/Buntworthy)) at [Lambda](https://lambdalabs.com/)
36
+
37
+ __Get the [code](https://github.com/justinpinkney/stable-diffusion) and [model](https://huggingface.co/lambdalabs/stable-diffusion-image-conditioned).__
38
+
39
+ ![](https://raw.githubusercontent.com/justinpinkney/stable-diffusion/main/assets/im-vars-thin.jpg)
40
+
41
+ """
42
+
43
+ article = \
44
+ """
45
+ ## How does this work?
46
+
47
+ The normal Stable Diffusion model is trained to be conditioned on text input. This version has had the original text encoder (from CLIP) removed, and replaced with
48
+ the CLIP _image_ encoder instead. So instead of generating images based a text input, images are generated to match CLIP's embedding of the image.
49
+ This creates images which have the same rough style and content, but different details, in particular the composition is generally quite different.
50
+ This is a totally different approach to the img2img script of the original Stable Diffusion and gives very different results.
51
+
52
+ The model was fine tuned on the [LAION aethetics v2 6+ dataset](https://laion.ai/blog/laion-aesthetics/) to accept the new conditioning.
53
+ Training was done on 4xA6000 GPUs on [Lambda GPU Cloud](https://lambdalabs.com/service/gpu-cloud).
54
+ More details on the method and training will come in a future blog post.
55
+ """
56
+
57
+ device = "cpu"
58
+ pipe = StableDiffusionImageEmbedPipeline.from_pretrained(
59
+ "lambdalabs/sd-image-variations-diffusers",
60
+ revision="273115e88df42350019ef4d628265b8c29ef4af5",
61
+ )
62
+ pipe = pipe.to(device)
63
+
64
+ inputs = [
65
+ gr.Image(),
66
+ gr.Slider(0, 25, value=3, step=1, label="Guidance scale"),
67
+ gr.Slider(1, 4, value=1, step=1, label="Number images"),
68
+ gr.Slider(5, 50, value=25, step=5, label="Steps"),
69
+ gr.Slider(
70
+ label="Seed",
71
+ minimum=0,
72
+ maximum=2147483647,
73
+ step=1,
74
+ randomize=True,
75
+ )
76
+ ]
77
+ output = gr.Gallery(label="Generated variations")
78
+ output.style(grid=2)
79
+
80
+ examples = [
81
+ ["assets/im-examples/vermeer.jpg", 3, 1, True, 25],
82
+ ["assets/im-examples/matisse.jpg", 3, 1, True, 25],
83
+ ]
84
+
85
+ demo = gr.Interface(
86
+ fn=main,
87
+ title="Stable Diffusion Image Variations",
88
+ description=description,
89
+ article=article,
90
+ inputs=inputs,
91
+ outputs=output,
92
+ examples=examples,
93
+ )
94
+ demo.launch()
requirements.txt ADDED
@@ -0,0 +1 @@
 
 
1
+ git+git://github.com/LambdaLabsML/lambda-diffusers.git#egg=lambda-diffusers
unsafe.png ADDED