add img2imgSegmindVegaRT
Browse files
pipelines/img2imgSegmindVegaRT.py
ADDED
@@ -0,0 +1,209 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from diffusers import (
|
2 |
+
AutoPipelineForImage2Image,
|
3 |
+
LCMScheduler,
|
4 |
+
AutoencoderTiny,
|
5 |
+
)
|
6 |
+
from compel import Compel, ReturnedEmbeddingsType
|
7 |
+
import torch
|
8 |
+
|
9 |
+
try:
|
10 |
+
import intel_extension_for_pytorch as ipex # type: ignore
|
11 |
+
except:
|
12 |
+
pass
|
13 |
+
|
14 |
+
import psutil
|
15 |
+
from config import Args
|
16 |
+
from pydantic import BaseModel, Field
|
17 |
+
from PIL import Image
|
18 |
+
import math
|
19 |
+
|
20 |
+
base_model = "segmind/Segmind-Vega"
|
21 |
+
lora_model = "segmind/Segmind-VegaRT"
|
22 |
+
taesd_model = "madebyollin/taesdxl"
|
23 |
+
|
24 |
+
default_prompt = "close-up photography of old man standing in the rain at night, in a street lit by lamps, leica 35mm summilux"
|
25 |
+
default_negative_prompt = "blurry, low quality, render, 3D, oversaturated"
|
26 |
+
page_content = """
|
27 |
+
<h1 class="text-3xl font-bold">Real-Time SDXL Turbo</h1>
|
28 |
+
<h3 class="text-xl font-bold">Image-to-Image</h3>
|
29 |
+
<p class="text-sm">
|
30 |
+
This demo showcases
|
31 |
+
<a
|
32 |
+
href="https://huggingface.co/stabilityai/sdxl-turbo"
|
33 |
+
target="_blank"
|
34 |
+
class="text-blue-500 underline hover:no-underline">SDXL Turbo</a>
|
35 |
+
Image to Image pipeline using
|
36 |
+
<a
|
37 |
+
href="https://huggingface.co/docs/diffusers/main/en/using-diffusers/sdxl_turbo"
|
38 |
+
target="_blank"
|
39 |
+
class="text-blue-500 underline hover:no-underline">Diffusers</a
|
40 |
+
> with a MJPEG stream server.
|
41 |
+
</p>
|
42 |
+
<p class="text-sm text-gray-500">
|
43 |
+
Change the prompt to generate different images, accepts <a
|
44 |
+
href="https://github.com/damian0815/compel/blob/main/doc/syntax.md"
|
45 |
+
target="_blank"
|
46 |
+
class="text-blue-500 underline hover:no-underline">Compel</a
|
47 |
+
> syntax.
|
48 |
+
</p>
|
49 |
+
"""
|
50 |
+
|
51 |
+
|
52 |
+
class Pipeline:
|
53 |
+
class Info(BaseModel):
|
54 |
+
name: str = "img2img"
|
55 |
+
title: str = "Image-to-Image Playground 256"
|
56 |
+
description: str = "Generates an image from a text prompt"
|
57 |
+
input_mode: str = "image"
|
58 |
+
page_content: str = page_content
|
59 |
+
|
60 |
+
class InputParams(BaseModel):
|
61 |
+
prompt: str = Field(
|
62 |
+
default_prompt,
|
63 |
+
title="Prompt",
|
64 |
+
field="textarea",
|
65 |
+
id="prompt",
|
66 |
+
)
|
67 |
+
negative_prompt: str = Field(
|
68 |
+
default_negative_prompt,
|
69 |
+
title="Negative Prompt",
|
70 |
+
field="textarea",
|
71 |
+
id="negative_prompt",
|
72 |
+
hide=True,
|
73 |
+
)
|
74 |
+
seed: int = Field(
|
75 |
+
2159232, min=0, title="Seed", field="seed", hide=True, id="seed"
|
76 |
+
)
|
77 |
+
steps: int = Field(
|
78 |
+
4, min=1, max=15, title="Steps", field="range", hide=True, id="steps"
|
79 |
+
)
|
80 |
+
width: int = Field(
|
81 |
+
1024, min=2, max=15, title="Width", disabled=True, hide=True, id="width"
|
82 |
+
)
|
83 |
+
height: int = Field(
|
84 |
+
1024, min=2, max=15, title="Height", disabled=True, hide=True, id="height"
|
85 |
+
)
|
86 |
+
guidance_scale: float = Field(
|
87 |
+
0.2,
|
88 |
+
min=0,
|
89 |
+
max=20,
|
90 |
+
step=0.001,
|
91 |
+
title="Guidance Scale",
|
92 |
+
field="range",
|
93 |
+
hide=True,
|
94 |
+
id="guidance_scale",
|
95 |
+
)
|
96 |
+
strength: float = Field(
|
97 |
+
0.5,
|
98 |
+
min=0.25,
|
99 |
+
max=1.0,
|
100 |
+
step=0.001,
|
101 |
+
title="Strength",
|
102 |
+
field="range",
|
103 |
+
hide=True,
|
104 |
+
id="strength",
|
105 |
+
)
|
106 |
+
|
107 |
+
def __init__(self, args: Args, device: torch.device, torch_dtype: torch.dtype):
|
108 |
+
if args.safety_checker:
|
109 |
+
self.pipe = AutoPipelineForImage2Image.from_pretrained(
|
110 |
+
base_model,
|
111 |
+
variant="fp16",
|
112 |
+
)
|
113 |
+
else:
|
114 |
+
self.pipe = AutoPipelineForImage2Image.from_pretrained(
|
115 |
+
base_model,
|
116 |
+
safety_checker=None,
|
117 |
+
variant="fp16",
|
118 |
+
)
|
119 |
+
if args.use_taesd:
|
120 |
+
self.pipe.vae = AutoencoderTiny.from_pretrained(
|
121 |
+
taesd_model, torch_dtype=torch_dtype, use_safetensors=True
|
122 |
+
).to(device)
|
123 |
+
|
124 |
+
self.pipe.load_lora_weights(lora_model)
|
125 |
+
self.pipe.fuse_lora()
|
126 |
+
self.pipe.scheduler = LCMScheduler.from_pretrained(
|
127 |
+
base_model, subfolder="scheduler"
|
128 |
+
)
|
129 |
+
self.pipe.set_progress_bar_config(disable=True)
|
130 |
+
self.pipe.to(device=device, dtype=torch_dtype)
|
131 |
+
if device.type != "mps":
|
132 |
+
self.pipe.unet.to(memory_format=torch.channels_last)
|
133 |
+
|
134 |
+
# check if computer has less than 64GB of RAM using sys or os
|
135 |
+
if psutil.virtual_memory().total < 64 * 1024**3:
|
136 |
+
self.pipe.enable_attention_slicing()
|
137 |
+
|
138 |
+
if args.torch_compile:
|
139 |
+
print("Running torch compile")
|
140 |
+
self.pipe.unet = torch.compile(
|
141 |
+
self.pipe.unet,
|
142 |
+
)
|
143 |
+
self.pipe.vae = torch.compile(
|
144 |
+
self.pipe.vae,
|
145 |
+
)
|
146 |
+
|
147 |
+
self.pipe(
|
148 |
+
prompt="warmup",
|
149 |
+
image=[Image.new("RGB", (768, 768))],
|
150 |
+
)
|
151 |
+
if args.compel:
|
152 |
+
self.pipe.compel_proc = Compel(
|
153 |
+
tokenizer=[self.pipe.tokenizer, self.pipe.tokenizer_2],
|
154 |
+
text_encoder=[self.pipe.text_encoder, self.pipe.text_encoder_2],
|
155 |
+
returned_embeddings_type=ReturnedEmbeddingsType.PENULTIMATE_HIDDEN_STATES_NON_NORMALIZED,
|
156 |
+
requires_pooled=[False, True],
|
157 |
+
)
|
158 |
+
|
159 |
+
def predict(self, params: "Pipeline.InputParams") -> Image.Image:
|
160 |
+
generator = torch.manual_seed(params.seed)
|
161 |
+
prompt = params.prompt
|
162 |
+
negative_prompt = params.negative_prompt
|
163 |
+
prompt_embeds = None
|
164 |
+
pooled_prompt_embeds = None
|
165 |
+
negative_prompt_embeds = None
|
166 |
+
negative_pooled_prompt_embeds = None
|
167 |
+
if hasattr(self.pipe, "compel_proc"):
|
168 |
+
prompt_embeds = self.pipe.compel_proc(
|
169 |
+
[params.prompt, params.negative_prompt]
|
170 |
+
)
|
171 |
+
prompt = None
|
172 |
+
negative_prompt = None
|
173 |
+
prompt_embeds = prompt_embeds[0:1]
|
174 |
+
pooled_prompt_embeds = pooled_prompt_embeds[0:1]
|
175 |
+
negative_prompt_embeds = prompt_embeds[1:2]
|
176 |
+
negative_pooled_prompt_embeds = pooled_prompt_embeds[1:2]
|
177 |
+
|
178 |
+
steps = params.steps
|
179 |
+
strength = params.strength
|
180 |
+
if int(steps * strength) < 1:
|
181 |
+
steps = math.ceil(1 / max(0.10, strength))
|
182 |
+
|
183 |
+
results = self.pipe(
|
184 |
+
image=params.image,
|
185 |
+
prompt=prompt,
|
186 |
+
negative_prompt=negative_prompt,
|
187 |
+
prompt_embeds=prompt_embeds,
|
188 |
+
pooled_prompt_embeds=pooled_prompt_embeds,
|
189 |
+
negative_prompt_embeds=negative_prompt_embeds,
|
190 |
+
negative_pooled_prompt_embeds=negative_pooled_prompt_embeds,
|
191 |
+
generator=generator,
|
192 |
+
strength=strength,
|
193 |
+
num_inference_steps=steps,
|
194 |
+
guidance_scale=params.guidance_scale,
|
195 |
+
width=params.width,
|
196 |
+
height=params.height,
|
197 |
+
output_type="pil",
|
198 |
+
)
|
199 |
+
|
200 |
+
nsfw_content_detected = (
|
201 |
+
results.nsfw_content_detected[0]
|
202 |
+
if "nsfw_content_detected" in results
|
203 |
+
else False
|
204 |
+
)
|
205 |
+
if nsfw_content_detected:
|
206 |
+
return None
|
207 |
+
result_image = results.images[0]
|
208 |
+
|
209 |
+
return result_image
|