File size: 5,951 Bytes
2951b6b
6a5df00
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2951b6b
6a5df00
 
 
 
46bd9ac
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
6a5df00
 
 
 
 
 
 
46bd9ac
6a5df00
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2951b6b
6a5df00
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
cf3ff1a
 
6a5df00
 
 
 
 
 
 
 
 
 
2951b6b
 
 
 
6a5df00
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
from diffusers import DiffusionPipeline, LCMScheduler, AutoencoderKL, AutoencoderTiny
from compel import Compel, ReturnedEmbeddingsType
import torch

try:
    import intel_extension_for_pytorch as ipex  # type: ignore
except:
    pass

import psutil
from config import Args
from pydantic import BaseModel, Field
from PIL import Image

model_id = "stabilityai/stable-diffusion-xl-base-1.0"
lcm_lora_id = "latent-consistency/lcm-lora-sdxl"
taesd_model = "madebyollin/taesdxl"


default_prompt = "close-up photography of old man standing in the rain at night, in a street lit by lamps, leica 35mm summilux"
default_negative_prompt = "blurry, low quality, render, 3D, oversaturated"
page_content = """
<h1 class="text-3xl font-bold">Real-Time Latent Consistency Model</h1>
<h3 class="text-xl font-bold">Text-to-Image SDXL + LCM + LoRA</h3>
<p class="text-sm">
    This demo showcases
    <a
    href="https://huggingface.co/blog/lcm_lora"
    target="_blank"
    class="text-blue-500 underline hover:no-underline">LCM LoRA</a
    >
    Text to Image pipeline using
    <a
    href="https://huggingface.co/docs/diffusers/main/en/using-diffusers/lcm#performing-inference-with-lcm"
    target="_blank"
    class="text-blue-500 underline hover:no-underline">Diffusers</a
    > with a MJPEG stream server.
</p>
<p class="text-sm text-gray-500">
    Change the prompt to generate different images, accepts <a
    href="https://github.com/damian0815/compel/blob/main/doc/syntax.md"
    target="_blank"
    class="text-blue-500 underline hover:no-underline">Compel</a
    > syntax.
</p>
"""


class Pipeline:
    class Info(BaseModel):
        name: str = "LCM+Lora+SDXL"
        title: str = "Text-to-Image SDXL + LCM + LoRA"
        description: str = "Generates an image from a text prompt"
        page_content: str = page_content
        input_mode: str = "text"

    class InputParams(BaseModel):
        prompt: str = Field(
            default_prompt,
            title="Prompt",
            field="textarea",
            id="prompt",
        )
        negative_prompt: str = Field(
            default_negative_prompt,
            title="Negative Prompt",
            field="textarea",
            id="negative_prompt",
            hide=True,
        )
        seed: int = Field(
            2159232, min=0, title="Seed", field="seed", hide=True, id="seed"
        )
        steps: int = Field(
            4, min=1, max=15, title="Steps", field="range", hide=True, id="steps"
        )
        width: int = Field(
            1024, min=2, max=15, title="Width", disabled=True, hide=True, id="width"
        )
        height: int = Field(
            1024, min=2, max=15, title="Height", disabled=True, hide=True, id="height"
        )
        guidance_scale: float = Field(
            1.0,
            min=0,
            max=20,
            step=0.001,
            title="Guidance Scale",
            field="range",
            hide=True,
            id="guidance_scale",
        )

    def __init__(self, args: Args, device: torch.device, torch_dtype: torch.dtype):
        vae = AutoencoderKL.from_pretrained(
            "madebyollin/sdxl-vae-fp16-fix", torch_dtype=torch_dtype
        )
        if args.safety_checker:
            self.pipe = DiffusionPipeline.from_pretrained(
                model_id,
                vae=vae,
            )
        else:
            self.pipe = DiffusionPipeline.from_pretrained(
                model_id,
                safety_checker=None,
                vae=vae,
            )
        # Load LCM LoRA
        self.pipe.load_lora_weights(lcm_lora_id, adapter_name="lcm")
        self.pipe.scheduler = LCMScheduler.from_config(self.pipe.scheduler.config)
        self.pipe.set_progress_bar_config(disable=True)
        self.pipe.to(device=device, dtype=torch_dtype).to(device)
        if device.type != "mps":
            self.pipe.unet.to(memory_format=torch.channels_last)

        if psutil.virtual_memory().total < 64 * 1024**3:
            self.pipe.enable_attention_slicing()

        self.pipe.compel_proc = Compel(
            tokenizer=[self.pipe.tokenizer, self.pipe.tokenizer_2],
            text_encoder=[self.pipe.text_encoder, self.pipe.text_encoder_2],
            returned_embeddings_type=ReturnedEmbeddingsType.PENULTIMATE_HIDDEN_STATES_NON_NORMALIZED,
            requires_pooled=[False, True],
        )
        if args.use_taesd:
            self.pipe.vae = AutoencoderTiny.from_pretrained(
                taesd_model, torch_dtype=torch_dtype, use_safetensors=True
            ).to(device)

        if args.torch_compile:
            self.pipe.unet = torch.compile(
                self.pipe.unet, mode="reduce-overhead", fullgraph=True
            )
            self.pipe.vae = torch.compile(
                self.pipe.vae, mode="reduce-overhead", fullgraph=True
            )
            self.pipe(
                prompt="warmup",
            )

    def predict(self, params: "Pipeline.InputParams") -> Image.Image:
        generator = torch.manual_seed(params.seed)

        prompt_embeds, pooled_prompt_embeds = self.pipe.compel_proc(
            [params.prompt, params.negative_prompt]
        )
        results = self.pipe(
            prompt_embeds=prompt_embeds[0:1],
            pooled_prompt_embeds=pooled_prompt_embeds[0:1],
            negative_prompt_embeds=prompt_embeds[1:2],
            negative_pooled_prompt_embeds=pooled_prompt_embeds[1:2],
            generator=generator,
            num_inference_steps=params.steps,
            guidance_scale=params.guidance_scale,
            width=params.width,
            height=params.height,
            output_type="pil",
        )

        nsfw_content_detected = (
            results.nsfw_content_detected[0]
            if "nsfw_content_detected" in results
            else False
        )
        if nsfw_content_detected:
            return None
        result_image = results.images[0]

        return result_image