File size: 7,934 Bytes
0f8ec45
 
 
 
 
2d30d63
0f8ec45
1fd4564
da74abb
2d30d63
 
0f8ec45
dd6d711
0f8ec45
dd6d711
 
 
 
 
da74abb
0f8ec45
ed2d2b6
0f8ec45
 
 
 
 
 
 
2d30d63
 
 
 
 
 
 
 
 
 
 
 
 
 
ed2d2b6
2d30d63
 
 
 
 
 
 
 
0f8ec45
ed2d2b6
0f8ec45
3142fb1
 
 
0f8ec45
 
ed2d2b6
0f8ec45
 
 
 
ad0113c
0f8ec45
 
 
afeabee
dd6d711
 
afeabee
 
 
da74abb
ed2d2b6
afeabee
 
 
 
 
 
 
dd6d711
 
 
 
9459bfb
d68c977
dd6d711
 
 
 
 
 
0f8ec45
ed2d2b6
0f8ec45
 
 
da74abb
0f8ec45
543cd64
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
0f8ec45
 
 
 
 
 
 
 
 
8f1a540
cf80990
8f1a540
 
cf80990
c1d1b49
1332b31
8f1a540
 
 
cf80990
1332b31
 
8f1a540
 
 
cf80990
1332b31
 
8f1a540
 
 
2d30d63
c1d1b49
1332b31
8f1a540
 
 
1332b31
8f1a540
1332b31
8f1a540
 
 
0f8ec45
 
 
 
 
 
582e489
0f8ec45
582e489
0f8ec45
 
 
 
 
 
 
56b6710
0f8ec45
 
 
 
 
 
 
 
 
8f1a540
cf80990
8f1a540
 
 
cd04efe
8f1a540
 
0f8ec45
 
 
6aef7b0
 
 
2d30d63
 
 
 
0f8ec45
 
 
 
 
 
 
 
 
 
 
 
 
 
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
from __future__ import annotations
import math
import random
import spaces
import gradio as gr
import numpy as np
import torch
from PIL import Image
from diffusers import StableDiffusionXLPipeline, EDMEulerScheduler, StableDiffusionXLInstructPix2PixPipeline, AutoencoderKL
from custom_pipeline import CosStableDiffusionXLInstructPix2PixPipeline
from huggingface_hub import hf_hub_download
from huggingface_hub import InferenceClient
from diffusers import StableDiffusion3Pipeline, SD3Transformer2DModel, FlowMatchEulerDiscreteScheduler

device = "cuda" if torch.cuda.is_available() else "cpu"
dtype = torch.float16

repo = "stabilityai/stable-diffusion-3-medium-diffusers"
pipe = StableDiffusion3Pipeline.from_pretrained(repo, torch_dtype=torch.float16).to(device)

help_text = """
To optimize image results:
- Adjust the **Image CFG weight** if the image isn't changing enough or is changing too much. Lower it to allow bigger changes, or raise it to preserve original details.
- Modify the **Text CFG weight** to influence how closely the edit follows text instructions. Increase it to adhere more to the text, or decrease it for subtler changes.
- Experiment with different **random seeds** and **CFG values** for varied outcomes.
- **Rephrase your instructions** for potentially better results.
- **Increase the number of steps** for enhanced edits.
"""

def set_timesteps_patched(self, num_inference_steps: int, device = None):
    self.num_inference_steps = num_inference_steps
    
    ramp = np.linspace(0, 1, self.num_inference_steps)
    sigmas = torch.linspace(math.log(self.config.sigma_min), math.log(self.config.sigma_max), len(ramp)).exp().flip(0)
    
    sigmas = (sigmas).to(dtype=torch.float32, device=device)
    self.timesteps = self.precondition_noise(sigmas)
    
    self.sigmas = torch.cat([sigmas, torch.zeros(1, device=sigmas.device)])
    self._step_index = None
    self._begin_index = None
    self.sigmas = self.sigmas.to("cpu") 

# Image Editor
edit_file = hf_hub_download(repo_id="stabilityai/cosxl", filename="cosxl_edit.safetensors")
EDMEulerScheduler.set_timesteps = set_timesteps_patched
vae = AutoencoderKL.from_pretrained("madebyollin/sdxl-vae-fp16-fix", torch_dtype=torch.float16)
pipe_edit = StableDiffusionXLInstructPix2PixPipeline.from_single_file(
    edit_file, num_in_channels=8, is_cosxl_edit=True, vae=vae, torch_dtype=torch.float16,
)
pipe_edit.scheduler = EDMEulerScheduler(sigma_min=0.002, sigma_max=120.0, sigma_data=1.0, prediction_type="v_prediction")
pipe_edit.to("cuda")

# Generator
@spaces.GPU(duration=30, queue=False)
def king(type ,
        input_image ,
        instruction: str ,
        steps: int = 8,
        randomize_seed: bool = False,
        seed: int = 25,
        text_cfg_scale: float = 7.3,
        image_cfg_scale: float = 1.7,
        width: int = 1024,
        height: int = 1024,
        guidance_scale: float = 6,
        use_resolution_binning: bool = True,
        progress=gr.Progress(track_tqdm=True),
    ):
    if type=="Image Editing" :
        if randomize_seed:
            seed = random.randint(0, 99999)
        text_cfg_scale = text_cfg_scale
        image_cfg_scale = image_cfg_scale
        input_image = input_image

        steps=steps
        generator = torch.manual_seed(seed)
        output_image = pipe_edit(
            instruction, image=input_image,
            guidance_scale=text_cfg_scale, image_guidance_scale=image_cfg_scale,
            num_inference_steps=steps, generator=generator).images[0]
        return seed, output_image
    else :
        if randomize_seed:
            seed = random.randint(0, 99999)
        generator = torch.Generator().manual_seed(seed)
        image = pipe(
            prompt = instruction,
            guidance_scale = 7, 
            num_inference_steps = steps, 
            width = width, 
            height = height,
            generator = generator
        ).images[0]         
        return seed, image

# Prompt classifier
def response(instruction, input_image=None):
    if input_image is None:
        output="Image Generation"
        yield output
    else:
        client = InferenceClient("mistralai/Mixtral-8x7B-Instruct-v0.1")
            
        generate_kwargs = dict(
                max_new_tokens=5,
            )
    
        system="[SYSTEM] You will be provided with text, and your task is to classify task is image generation or image editing answer with only task do not say anything else and stop as soon as possible. [TEXT]"
            
        formatted_prompt = system + instruction + "[TASK]"
        stream = client.text_generation(formatted_prompt, **generate_kwargs, stream=True, details=True, return_full_text=False)
        output = ""
        for response in stream:        
            if not response.token.text == "</s>":
                output += response.token.text
        if "editing" in output:
            output = "Image Editing"
        else:
            output = "Image Generation"
        yield output
    return output

css = '''
.gradio-container{max-width: 600px !important}
h1{text-align:center}
footer {
    visibility: hidden
}
'''

examples=[
        [
            "Image Generation",
            None,
            "A luxurious supercar with a unique design. The car should have a pearl white finish, and gold accents. 4k, realistic.",

        ],
        [
            "Image Editing",
            "./supercar.png",
            "make it red",

        ],
        [
            "Image Editing",
            "./red_car.png",
            "add some snow",

        ],
        [
            "Image Generation",
            None,
            "Ironman fighting with hulk, wall painting",

        ],
        [
            "Image Generation",
            None,
            "Beautiful Eiffel Tower at Night",

        ],
    ]

with gr.Blocks(css=css) as demo:
    gr.Markdown("# Image Generator Pro")
    with gr.Row():
        with gr.Column(scale=4):
            instruction = gr.Textbox(lines=1, label="Instruction", interactive=True)
        with gr.Column(scale=1):
            type = gr.Dropdown(["Image Generation","Image Editing"], label="Task", value="Image Generation",interactive=True, info="AI will select option based on your query, but if it selects wrong, please choose correct one.")
        with gr.Column(scale=1):
            generate_button = gr.Button("Generate")

    with gr.Row():
        input_image = gr.Image(label="Image", type="pil", interactive=True)

    with gr.Row():
        text_cfg_scale = gr.Number(value=7.3, step=0.1, label="Text CFG", interactive=True)
        image_cfg_scale = gr.Number(value=1.7, step=0.1,label="Image CFG", interactive=True)
        steps = gr.Number(value=25, precision=0, label="Steps", interactive=True)
        randomize_seed = gr.Radio(
                ["Fix Seed", "Randomize Seed"],
                value="Randomize Seed",
                type="index",
                show_label=False,
                interactive=True,
            )
        seed = gr.Number(value=1371, precision=0, label="Seed", interactive=True)

    gr.Examples(
        examples=examples,
        inputs=[type,input_image, instruction],
        fn=king,
        outputs=[input_image],
        cache_examples=False,
    )

    gr.Markdown(help_text)

    instruction.change(fn=response, inputs=[instruction,input_image], outputs=type, queue=False)

    input_image.upload(fn=response, inputs=[instruction,input_image], outputs=type, queue=False)
    
    gr.on(triggers=[
            generate_button.click,
            instruction.submit
        ],
            fn=king,
            inputs=[type,
                input_image,
                instruction,
                steps,
                randomize_seed,
                seed,
                text_cfg_scale,
                image_cfg_scale,
            ],
            outputs=[seed, input_image],
        )

demo.queue(max_size=99999).launch()