File size: 9,801 Bytes
54238f9
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
from io import BytesIO
import torch
import numpy as np
from PIL import Image
from einops import rearrange
from torch import autocast
from contextlib import nullcontext
import requests
import functools

from ldm.models.diffusion.ddim import DDIMSampler
from ldm.models.diffusion.plms import PLMSSampler
from ldm.extras import load_model_from_config, load_training_dir
import clip

from PIL import Image

from huggingface_hub import hf_hub_download
ckpt = hf_hub_download(repo_id="lambdalabs/image-mixer", filename="image-mixer-pruned.ckpt")
config = hf_hub_download(repo_id="lambdalabs/image-mixer", filename="image-mixer-config.yaml")

device = "cuda:0"
model = load_model_from_config(config, ckpt, device=device, verbose=False)
model = model.to(device).half()

clip_model, preprocess = clip.load("ViT-L/14", device=device)

n_inputs = 5

torch.cuda.empty_cache()

@functools.lru_cache()
def get_url_im(t):
    user_agent = {'User-agent': 'gradio-app'}
    response = requests.get(t, headers=user_agent)
    return Image.open(BytesIO(response.content))

@torch.no_grad()
def get_im_c(im_path, clip_model):
    # im = Image.open(im_path).convert("RGB")
    prompts = preprocess(im_path).to(device).unsqueeze(0)
    return clip_model.encode_image(prompts).float()

@torch.no_grad()
def get_txt_c(txt, clip_model):
    text = clip.tokenize([txt,]).to(device)
    return clip_model.encode_text(text)

def get_txt_diff(txt1, txt2, clip_model):
    return get_txt_c(txt1, clip_model) - get_txt_c(txt2, clip_model)

def to_im_list(x_samples_ddim):
    x_samples_ddim = torch.clamp((x_samples_ddim + 1.0) / 2.0, min=0.0, max=1.0)
    ims = []
    for x_sample in x_samples_ddim:
        x_sample = 255. * rearrange(x_sample.cpu().numpy(), 'c h w -> h w c')
        ims.append(Image.fromarray(x_sample.astype(np.uint8)))
    return ims

@torch.no_grad()
def sample(sampler, model, c, uc, scale, start_code, h=512, w=512, precision="autocast",ddim_steps=50):
    ddim_eta=0.0
    precision_scope = autocast if precision=="autocast" else nullcontext
    with precision_scope("cuda"):
        shape = [4, h // 8, w // 8]
        samples_ddim, _ = sampler.sample(S=ddim_steps,
                                            conditioning=c,
                                            batch_size=c.shape[0],
                                            shape=shape,
                                            verbose=False,
                                            unconditional_guidance_scale=scale,
                                            unconditional_conditioning=uc,
                                            eta=ddim_eta,
                                            x_T=start_code)

        x_samples_ddim = model.decode_first_stage(samples_ddim)
    return to_im_list(x_samples_ddim)

def run(*args):

    inps = []
    for i in range(0, len(args)-4, n_inputs):
        inps.append(args[i:i+n_inputs])

    scale, n_samples, seed, steps = args[-4:]
    h = w = 640

    sampler = DDIMSampler(model)
    # sampler = PLMSSampler(model)

    torch.manual_seed(seed)
    start_code = torch.randn(n_samples, 4, h//8, w//8, device=device)
    conds = []

    for b, t, im, s in zip(*inps):
        if b == "Image":
            this_cond = s*get_im_c(im, clip_model)
        elif b == "Text/URL":
            if t.startswith("http"):
                im = get_url_im(t)
                this_cond = s*get_im_c(im, clip_model)
            else:
                this_cond = s*get_txt_c(t, clip_model)
        else:
            this_cond = torch.zeros((1, 768), device=device)
        conds.append(this_cond)
    conds = torch.cat(conds, dim=0).unsqueeze(0)
    conds = conds.tile(n_samples, 1, 1)

    ims = sample(sampler, model, conds, 0*conds, scale, start_code, ddim_steps=steps)
    # return make_row(ims)

    # Clear GPU memory cache so less likely to OOM
    torch.cuda.empty_cache()
    return ims


import gradio as gr
from functools import partial
from itertools import chain

def change_visible(txt1, im1, val):
    outputs = {}
    if val == "Image":
        outputs[im1] = gr.update(visible=True)
        outputs[txt1] = gr.update(visible=False)
    elif val == "Text/URL":
        outputs[im1] = gr.update(visible=False)
        outputs[txt1] = gr.update(visible=True)
    elif val == "Nothing":
        outputs[im1] = gr.update(visible=False)
        outputs[txt1] = gr.update(visible=False)
    return outputs


with gr.Blocks(title="Image Mixer", css=".gr-box {border-color: #8136e2}") as demo:

    gr.Markdown("")
    gr.Markdown(
"""
# Image Mixer

_Created by [Justin Pinkney](https://www.justinpinkney.com) at [Lambda Labs](https://lambdalabs.com/)_

To skip the queue you can try it on <a href="https://gradio.659760a2fca34f7581e6586fa8d5d615.lambdaspaces.com/" style="display:inline-block;position: relative;"><img style="margin-top: 0;margin-bottom: 0;margin-left: .25em;" src="https://img.shields.io/badge/-Lambda%20Cloud-blueviolet"></a>, or <a href="https://huggingface.co/spaces/lambdalabs/image-mixer-demo?duplicate=true" style="display:inline-block;position: relative;"><img style="margin-top: 0;margin-bottom: 0;margin-left: .25em;" src="https://bit.ly/3gLdBN6"></a>

### __Provide one or more images to be mixed together by a fine-tuned Stable Diffusion model (see tips and advice below👇).__

![banner-large.jpeg](https://s3.amazonaws.com/moonup/production/uploads/1674039767068-62bd5f951e22ec84279820e8.jpeg)

""")

    btns = []
    txts = []
    ims = []
    strengths = []

    with gr.Row():
        for i in range(n_inputs):
            with gr.Box():
                with gr.Column():
                    btn1 = gr.Radio(
                        choices=["Image", "Text/URL", "Nothing"],
                        label=f"Input {i} type",
                        interactive=True,
                        value="Nothing",
                        )
                    txt1 = gr.Textbox(label="Text or Image URL", visible=False, interactive=True)
                    im1 = gr.Image(label="Image", interactive=True, visible=False, type="pil")
                    strength = gr.Slider(label="Strength", minimum=0, maximum=5, step=0.05, value=1, interactive=True)
    
                    fn = partial(change_visible, txt1, im1)
                    btn1.change(fn=fn, inputs=[btn1], outputs=[txt1, im1], queue=False)
    
                    btns.append(btn1)
                    txts.append(txt1)
                    ims.append(im1)
                    strengths.append(strength)
    with gr.Row():
        cfg_scale = gr.Slider(label="CFG scale", value=3, minimum=1, maximum=10, step=0.5)
        n_samples = gr.Slider(label="Num samples", value=1, minimum=1, maximum=1, step=1)
        seed = gr.Slider(label="Seed", value=0, minimum=0, maximum=10000, step=1)
        steps = gr.Slider(label="Steps", value=30, minimum=10, maximum=100, step=5)

    with gr.Row():
        submit = gr.Button("Generate")
    output = gr.Gallery().style(grid=[1,2], height="640px")

    inps = list(chain(btns, txts, ims, strengths))
    inps.extend([cfg_scale,n_samples,seed, steps,])
    submit.click(fn=run, inputs=inps, outputs=[output])

    ex = gr.Examples([
    [
        "Image", "Image", "Text/URL", "Nothing", "Nothing",
        "","","central symmetric figure detailed artwork","","",
        "gainsborough.jpeg","blonder.jpeg","blonder.jpeg","blonder.jpeg","blonder.jpeg",
        1,1.35,1.4,1,1,
        3.0, 1, 0, 30,
    ],
    [
        "Image", "Image", "Text/URL", "Image", "Nothing",
        "","","flowers","","",
        "ex2-1.jpeg","ex2-2.jpeg","blonder.jpeg","ex2-3.jpeg","blonder.jpeg",
        1,1,1.5,1.25,1,
        3.0, 1, 0, 30,
    ],
    [
        "Image", "Image", "Image", "Nothing", "Nothing",
        "","","","","",
        "ex1-1.jpeg","ex1-2.jpeg","ex1-3.jpeg","blonder.jpeg","blonder.jpeg",
        1.1,1,1.4,1,1,
        3.0, 1, 0, 30,
    ],
                     ],
    fn=run, inputs=inps, outputs=[output], cache_examples=True)

    gr.Markdown(
"""

## Tips

- You can provide between 1 and 5 inputs, these can either be an uploaded image a text prompt or a url to an image file.
- The order of the inputs shouldn't matter, any images will be centre cropped before use.
- Each input has an individual strength parameter which controls how big an influence it has on the output.
- The model was not trained using text and can not interpret complex text prompts.
- Using only text prompts doesn't work well, make sure there is at least one image or URL to an image.
- The parameters on the bottom row such as cfg scale do the same as for a normal Stable Diffusion model.
- Balancing the different inputs requires tweaking of the strengths, I suggest getting the right balance for a small number of samples and with few steps until you're
happy with the result then increase the steps for better quality.
- Outputs are 640x640 by default.
- If you want to run locally see the instruction on the [Model Card](https://huggingface.co/lambdalabs/image-mixer).

## How does this work?

This model is based on the [Stable Diffusion Image Variations model](https://huggingface.co/lambdalabs/sd-image-variations-diffusers)
but it has been fined tuned to take multiple CLIP image embeddings. During training, up to 5 random crops were taken from the training images and
the CLIP image embeddings were computed, these were then concatenated and used as the conditioning for the model. At inference time we can combine the image
embeddings from multiple images to mix their concepts (and we can also use the text encoder to add text concepts too).

The model was trained on a subset of LAION Improved Aesthetics at a resolution of 640x640 and was trained using 8xA100 GPUs on [Lambda GPU Cloud](https://lambdalabs.com/service/gpu-cloud).

""")

demo.launch()