Spaces:
Running
on
Zero
Running
on
Zero
bugfix
Browse files
app.py
CHANGED
@@ -2,51 +2,30 @@ import os
|
|
2 |
import gradio as gr
|
3 |
import numpy as np
|
4 |
import random
|
5 |
-
import spaces
|
6 |
-
from diffusers import DiffusionPipeline
|
7 |
import torch
|
8 |
import json
|
9 |
import logging
|
10 |
-
from diffusers import DiffusionPipeline
|
11 |
-
from live_preview_helpers import calculate_shift, retrieve_timesteps, flux_pipe_call_that_returns_an_iterable_of_images
|
12 |
from huggingface_hub import login
|
13 |
-
from huggingface_hub import hf_hub_download, HfFileSystem, ModelCard, snapshot_download
|
14 |
-
import copy
|
15 |
-
import random
|
16 |
import time
|
17 |
-
import boto3
|
18 |
-
from io import BytesIO
|
19 |
from datetime import datetime
|
20 |
-
from
|
21 |
-
|
22 |
-
|
23 |
-
|
24 |
-
|
25 |
-
|
26 |
HF_TOKEN = os.environ.get("HF_TOKEN")
|
27 |
-
|
28 |
login(token=HF_TOKEN)
|
29 |
|
30 |
-
#
|
31 |
-
dtype = torch.
|
32 |
device = "cuda" if torch.cuda.is_available() else "cpu"
|
33 |
-
base_model = "black-forest-labs/FLUX.1-dev"
|
34 |
-
|
35 |
-
# unet = UNet2DConditionModel.from_pretrained(
|
36 |
-
# base_model,
|
37 |
-
# torch_dtype=torch.float16,
|
38 |
-
# use_safetensors=True,
|
39 |
-
# variant="fp16",
|
40 |
-
# subfolder="unet",
|
41 |
-
# # ).to("cuda")
|
42 |
-
# tokenizer = AutoTokenizer.from_pretrained("openai/clip-vit-base-patch32")
|
43 |
-
|
44 |
|
|
|
45 |
pipe = DiffusionPipeline.from_pretrained(base_model, torch_dtype=dtype).to(device)
|
46 |
|
47 |
-
|
48 |
-
|
49 |
-
MAX_SEED = 2**32-1
|
50 |
|
51 |
class calculateDuration:
|
52 |
def __init__(self, activity_name=""):
|
@@ -64,133 +43,285 @@ class calculateDuration:
|
|
64 |
else:
|
65 |
print(f"Elapsed time: {self.elapsed_time:.6f} seconds")
|
66 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
67 |
|
68 |
-
|
69 |
-
|
70 |
-
|
71 |
-
|
72 |
-
|
73 |
-
|
74 |
-
|
75 |
-
|
76 |
-
|
77 |
-
|
78 |
-
|
79 |
-
|
80 |
-
current_time = datetime.now().strftime("%Y/%m/%d/%H%M%S")
|
81 |
-
image_file = f"generated_images/{current_time}_{random.randint(0, MAX_SEED)}.png"
|
82 |
-
buffer = BytesIO()
|
83 |
-
image.save(buffer, "PNG")
|
84 |
-
buffer.seek(0)
|
85 |
-
s3.upload_fileobj(buffer, bucket_name, image_file)
|
86 |
-
print("upload finish", image_file)
|
87 |
-
return image_file
|
88 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
89 |
|
90 |
-
|
91 |
-
def
|
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 |
generator = torch.Generator(device="cuda").manual_seed(seed)
|
|
|
122 |
with calculateDuration("Generating image"):
|
123 |
# Generate image
|
124 |
-
|
125 |
-
|
126 |
num_inference_steps=steps,
|
127 |
guidance_scale=cfg_scale,
|
128 |
width=width,
|
129 |
height=height,
|
130 |
generator=generator,
|
131 |
-
joint_attention_kwargs=joint_attention_kwargs
|
132 |
).images[0]
|
133 |
-
|
134 |
progress(99, "Generate success!")
|
135 |
-
return
|
136 |
-
|
137 |
-
# 在 Transformer 中,自定义注意力处理器
|
138 |
-
class CustomAttentionProcessor(torch.nn.Module):
|
139 |
-
def __init__(self, attention_control, adapter_name):
|
140 |
-
super().__init__()
|
141 |
-
self.attention_control = attention_control
|
142 |
-
self.adapter_name = adapter_name
|
143 |
-
|
144 |
-
def forward(self, attention_probs):
|
145 |
-
# 调用自定义的注意力控制函数
|
146 |
-
attention_probs = self.attention_control(attention_probs, self.adapter_name)
|
147 |
-
return attention_probs
|
148 |
|
|
|
149 |
|
150 |
-
|
151 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
152 |
|
153 |
-
|
154 |
# Load LoRA weights
|
155 |
-
|
156 |
-
|
157 |
-
|
158 |
-
|
159 |
-
|
160 |
-
|
161 |
-
|
162 |
-
|
163 |
-
|
164 |
-
|
165 |
-
|
166 |
-
|
167 |
-
|
168 |
-
|
169 |
-
|
170 |
-
|
171 |
-
|
172 |
-
|
173 |
-
|
174 |
-
|
175 |
-
|
|
|
|
|
176 |
# Set random seed for reproducibility
|
177 |
if randomize_seed:
|
178 |
-
|
179 |
-
|
180 |
-
final_image = generate_image(prompt, steps, seed, cfg_scale, width, height, progress)
|
181 |
|
182 |
-
|
183 |
-
|
184 |
-
|
185 |
-
|
186 |
-
|
187 |
-
result = {"status": "success", "message": "Image generated but not uploaded"}
|
188 |
|
189 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
190 |
|
191 |
-
|
192 |
|
|
|
193 |
|
|
|
194 |
css="""
|
195 |
#col-container {
|
196 |
margin: 0 auto;
|
@@ -199,12 +330,16 @@ css="""
|
|
199 |
"""
|
200 |
|
201 |
with gr.Blocks(css=css) as demo:
|
202 |
-
gr.Markdown("Flux with
|
203 |
with gr.Row():
|
204 |
|
205 |
with gr.Column():
|
206 |
-
|
207 |
-
|
|
|
|
|
|
|
|
|
208 |
run_button = gr.Button("Run", scale=0)
|
209 |
|
210 |
with gr.Accordion("Advanced Settings", open=False):
|
@@ -215,11 +350,11 @@ with gr.Blocks(css=css) as demo:
|
|
215 |
lora_scale = gr.Slider(label="LoRA Scale", minimum=0, maximum=1, step=0.01, value=0.5)
|
216 |
|
217 |
with gr.Row():
|
218 |
-
width = gr.Slider(label="Width", minimum=256, maximum=1536, step=64, value=
|
219 |
-
height = gr.Slider(label="Height", minimum=256, maximum=1536, step=64, value=
|
220 |
|
221 |
with gr.Row():
|
222 |
-
cfg_scale = gr.Slider(label="CFG Scale", minimum=1, maximum=20, step=0.5, value=
|
223 |
steps = gr.Slider(label="Steps", minimum=1, maximum=50, step=1, value=28)
|
224 |
|
225 |
upload_to_r2 = gr.Checkbox(label="Upload to R2", value=False)
|
@@ -231,13 +366,35 @@ with gr.Blocks(css=css) as demo:
|
|
231 |
|
232 |
with gr.Column():
|
233 |
result = gr.Image(label="Result", show_label=False)
|
234 |
-
|
235 |
-
|
236 |
-
|
237 |
-
|
238 |
-
|
239 |
-
|
240 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
241 |
)
|
242 |
|
243 |
-
demo.queue().launch()
|
|
|
2 |
import gradio as gr
|
3 |
import numpy as np
|
4 |
import random
|
|
|
|
|
5 |
import torch
|
6 |
import json
|
7 |
import logging
|
8 |
+
from diffusers import DiffusionPipeline
|
|
|
9 |
from huggingface_hub import login
|
|
|
|
|
|
|
10 |
import time
|
|
|
|
|
11 |
from datetime import datetime
|
12 |
+
from io import BytesIO
|
13 |
+
from diffusers.models.attention_processor import AttentionProcessor
|
14 |
+
import re
|
15 |
+
import json
|
16 |
+
# 登录 Hugging Face Hub
|
|
|
17 |
HF_TOKEN = os.environ.get("HF_TOKEN")
|
|
|
18 |
login(token=HF_TOKEN)
|
19 |
|
20 |
+
# 初始化
|
21 |
+
dtype = torch.float16 # 您可以根据需要调整数据类型
|
22 |
device = "cuda" if torch.cuda.is_available() else "cpu"
|
23 |
+
base_model = "black-forest-labs/FLUX.1-dev" # 替换为您的模型
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
24 |
|
25 |
+
# 加载管道
|
26 |
pipe = DiffusionPipeline.from_pretrained(base_model, torch_dtype=dtype).to(device)
|
27 |
|
28 |
+
MAX_SEED = 2**32 - 1
|
|
|
|
|
29 |
|
30 |
class calculateDuration:
|
31 |
def __init__(self, activity_name=""):
|
|
|
43 |
else:
|
44 |
print(f"Elapsed time: {self.elapsed_time:.6f} seconds")
|
45 |
|
46 |
+
# 定义位置、偏移和区域的映射
|
47 |
+
valid_locations = { # x, y in 90*90
|
48 |
+
'in the center': (45, 45),
|
49 |
+
'on the left': (15, 45),
|
50 |
+
'on the right': (75, 45),
|
51 |
+
'on the top': (45, 15),
|
52 |
+
'on the bottom': (45, 75),
|
53 |
+
'on the top-left': (15, 15),
|
54 |
+
'on the top-right': (75, 15),
|
55 |
+
'on the bottom-left': (15, 75),
|
56 |
+
'on the bottom-right': (75, 75)
|
57 |
+
}
|
58 |
|
59 |
+
valid_offsets = { # x, y in 90*90
|
60 |
+
'no offset': (0, 0),
|
61 |
+
'slightly to the left': (-10, 0),
|
62 |
+
'slightly to the right': (10, 0),
|
63 |
+
'slightly to the upper': (0, -10),
|
64 |
+
'slightly to the lower': (0, 10),
|
65 |
+
'slightly to the upper-left': (-10, -10),
|
66 |
+
'slightly to the upper-right': (10, -10),
|
67 |
+
'slightly to the lower-left': (-10, 10),
|
68 |
+
'slightly to the lower-right': (10, 10)
|
69 |
+
}
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
70 |
|
71 |
+
valid_areas = { # w, h in 90*90
|
72 |
+
"a small square area": (50, 50),
|
73 |
+
"a small vertical area": (40, 60),
|
74 |
+
"a small horizontal area": (60, 40),
|
75 |
+
"a medium-sized square area": (60, 60),
|
76 |
+
"a medium-sized vertical area": (50, 80),
|
77 |
+
"a medium-sized horizontal area": (80, 50),
|
78 |
+
"a large square area": (70, 70),
|
79 |
+
"a large vertical area": (60, 90),
|
80 |
+
"a large horizontal area": (90, 60)
|
81 |
+
}
|
82 |
|
83 |
+
# 解析角色位置的函数
|
84 |
+
def parse_character_position(character_position):
|
85 |
+
# 定义正则表达式模式
|
86 |
+
location_pattern = '|'.join(re.escape(key) for key in valid_locations.keys())
|
87 |
+
offset_pattern = '|'.join(re.escape(key) for key in valid_offsets.keys())
|
88 |
+
area_pattern = '|'.join(re.escape(key) for key in valid_areas.keys())
|
89 |
+
|
90 |
+
# 提取位置
|
91 |
+
location_match = re.search(location_pattern, character_position, re.IGNORECASE)
|
92 |
+
location = location_match.group(0) if location_match else 'in the center'
|
93 |
+
|
94 |
+
# 提取偏移
|
95 |
+
offset_match = re.search(offset_pattern, character_position, re.IGNORECASE)
|
96 |
+
offset = offset_match.group(0) if offset_match else 'no offset'
|
97 |
+
|
98 |
+
# 提取区域
|
99 |
+
area_match = re.search(area_pattern, character_position, re.IGNORECASE)
|
100 |
+
area = area_match.group(0) if area_match else 'a medium-sized square area'
|
101 |
+
|
102 |
+
return {
|
103 |
+
'location': location,
|
104 |
+
'offset': offset,
|
105 |
+
'area': area
|
106 |
+
}
|
107 |
|
108 |
+
# 创建掩码的函数
|
109 |
+
def create_attention_mask(image_width, image_height, location, offset, area):
|
110 |
+
# 图像在生成时通常会被缩放为 90x90,因此先定义一个基础尺寸
|
111 |
+
base_size = 90
|
112 |
+
|
113 |
+
# 获取位置坐标
|
114 |
+
loc_x, loc_y = valid_locations.get(location, (45, 45))
|
115 |
+
# 获取偏移量
|
116 |
+
offset_x, offset_y = valid_offsets.get(offset, (0, 0))
|
117 |
+
# 获取区域大小
|
118 |
+
area_width, area_height = valid_areas.get(area, (60, 60))
|
119 |
+
|
120 |
+
# 计算最终位置
|
121 |
+
final_x = loc_x + offset_x
|
122 |
+
final_y = loc_y + offset_y
|
123 |
+
|
124 |
+
# 将坐标和尺寸映射到实际图像尺寸
|
125 |
+
scale_x = image_width / base_size
|
126 |
+
scale_y = image_height / base_size
|
127 |
+
|
128 |
+
center_x = final_x * scale_x
|
129 |
+
center_y = final_y * scale_y
|
130 |
+
width = area_width * scale_x
|
131 |
+
height = area_height * scale_y
|
132 |
+
|
133 |
+
# 计算左上角和右下角坐标
|
134 |
+
x_start = int(max(center_x - width / 2, 0))
|
135 |
+
y_start = int(max(center_y - height / 2, 0))
|
136 |
+
x_end = int(min(center_x + width / 2, image_width))
|
137 |
+
y_end = int(min(center_y + height / 2, image_height))
|
138 |
+
|
139 |
+
# 创建掩码
|
140 |
+
mask = torch.zeros((image_height, image_width), dtype=torch.float32, device="cuda")
|
141 |
+
mask[y_start:y_end, x_start:x_end] = 1.0
|
142 |
+
|
143 |
+
# 展平成一维
|
144 |
+
mask_flat = mask.view(-1) # 形状为 (image_height * image_width,)
|
145 |
+
return mask_flat
|
146 |
+
|
147 |
+
# 自定义注意力处理器
|
148 |
+
class CustomCrossAttentionProcessor(AttentionProcessor):
|
149 |
+
def __init__(self, masks, embeddings, adapter_names):
|
150 |
+
super().__init__()
|
151 |
+
self.masks = masks # 列表,包含每个角色的掩码
|
152 |
+
self.embeddings = embeddings # 列表,包含每个角色的嵌入
|
153 |
+
self.adapter_names = adapter_names # 列表,包含每个角色的 LoRA 适配器名称
|
154 |
+
|
155 |
+
def __call__(self, attn, hidden_states, encoder_hidden_states=None, attention_mask=None, **kwargs):
|
156 |
+
# 获取当前的 adapter_name
|
157 |
+
adapter_name = getattr(attn, 'adapter_name', None)
|
158 |
+
if adapter_name is None or adapter_name not in self.adapter_names:
|
159 |
+
# 如果没有 adapter_name,直接执行默认的注意力计算
|
160 |
+
return super().__call__(attn, hidden_states, encoder_hidden_states, attention_mask, **kwargs)
|
161 |
|
162 |
+
# 查找 adapter_name 对应的索引
|
163 |
+
idx = self.adapter_names.index(adapter_name)
|
164 |
+
mask = self.masks[idx]
|
165 |
|
166 |
+
# 标准的注意力计算
|
167 |
+
batch_size, sequence_length, _ = hidden_states.shape
|
168 |
+
|
169 |
+
query = attn.to_q(hidden_states)
|
170 |
+
key = attn.to_k(encoder_hidden_states)
|
171 |
+
value = attn.to_v(encoder_hidden_states)
|
172 |
+
|
173 |
+
# 重塑以适应多头注意力
|
174 |
+
query = query.view(batch_size, -1, attn.heads, attn.head_dim).transpose(1, 2)
|
175 |
+
key = key.view(batch_size, -1, attn.heads, attn.head_dim).transpose(1, 2)
|
176 |
+
value = value.view(batch_size, -1, attn.heads, attn.head_dim).transpose(1, 2)
|
177 |
+
|
178 |
+
# 计算注意力得分
|
179 |
+
attention_scores = torch.matmul(query, key.transpose(-1, -2)) * attn.scale
|
180 |
+
|
181 |
+
# 应用掩码调整注意力得分
|
182 |
+
# 将 mask 调整为与 attention_scores 兼容的形状
|
183 |
+
# 假设 key_len 与 mask 的长度一致
|
184 |
+
mask_expanded = mask.unsqueeze(0).unsqueeze(0).unsqueeze(0) # (1, 1, 1, key_len)
|
185 |
+
# 将掩码应用于 attention_scores
|
186 |
+
attention_scores += mask_expanded * 1e6 # 增强对应位置的注意力
|
187 |
+
|
188 |
+
# 计算注意力概率
|
189 |
+
attention_probs = torch.nn.functional.softmax(attention_scores, dim=-1)
|
190 |
+
|
191 |
+
# 计算上下文向量
|
192 |
+
context = torch.matmul(attention_probs, value)
|
193 |
+
|
194 |
+
# 重塑回原始形状
|
195 |
+
context = context.transpose(1, 2).reshape(batch_size, -1, attn.heads * attn.head_dim)
|
196 |
+
|
197 |
+
# 输出投影
|
198 |
+
hidden_states = attn.to_out(context)
|
199 |
+
return hidden_states
|
200 |
+
|
201 |
+
# 替换注意力处理器的函数
|
202 |
+
def replace_attention_processors(pipe, masks, embeddings, adapter_names):
|
203 |
+
custom_processor = CustomCrossAttentionProcessor(masks, embeddings, adapter_names)
|
204 |
+
for name, module in pipe.unet.named_modules():
|
205 |
+
if hasattr(module, 'attn2'):
|
206 |
+
# 设置 adapter_name 为模块的属性
|
207 |
+
module.attn2.adapter_name = getattr(module, 'adapter_name', None)
|
208 |
+
module.attn2.processor = custom_processor
|
209 |
+
|
210 |
+
# 生成图像的函数
|
211 |
+
@spaces.GPU
|
212 |
+
@torch.inference_mode()
|
213 |
+
def generate_image_with_embeddings(prompt_embeddings, steps, seed, cfg_scale, width, height, progress):
|
214 |
+
pipe.to("cuda")
|
215 |
generator = torch.Generator(device="cuda").manual_seed(seed)
|
216 |
+
|
217 |
with calculateDuration("Generating image"):
|
218 |
# Generate image
|
219 |
+
generated_image = pipe(
|
220 |
+
prompt_embeds=prompt_embeddings,
|
221 |
num_inference_steps=steps,
|
222 |
guidance_scale=cfg_scale,
|
223 |
width=width,
|
224 |
height=height,
|
225 |
generator=generator,
|
|
|
226 |
).images[0]
|
227 |
+
|
228 |
progress(99, "Generate success!")
|
229 |
+
return generated_image
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
230 |
|
231 |
+
# 主函数
|
232 |
|
233 |
+
def run_lora(prompt_bg, character_prompts_json, character_positions_json, lora_strings_json, prompt_details, cfg_scale, steps, randomize_seed, seed, width, height, lora_scale, upload_to_r2, account_id, access_key, secret_key, bucket, progress=gr.Progress(track_tqdm=True)):
|
234 |
+
|
235 |
+
# 解析角色提示词、位置和 LoRA 字符串
|
236 |
+
try:
|
237 |
+
character_prompts = json.loads(character_prompts_json)
|
238 |
+
character_positions = json.loads(character_positions_json)
|
239 |
+
lora_strings = json.loads(lora_strings_json)
|
240 |
+
except json.JSONDecodeError as e:
|
241 |
+
raise ValueError(f"Invalid JSON input: {e}")
|
242 |
+
|
243 |
+
# 确保提示词、位置和 LoRA 字符串的数量一致
|
244 |
+
if len(character_prompts) != len(character_positions) or len(character_prompts) != len(lora_strings):
|
245 |
+
raise ValueError("The number of character prompts, positions, and LoRA strings must be the same.")
|
246 |
+
|
247 |
+
# 角色的数量
|
248 |
+
num_characters = len(character_prompts)
|
249 |
|
|
|
250 |
# Load LoRA weights
|
251 |
+
with calculateDuration("Loading LoRA weights"):
|
252 |
+
pipe.unload_lora_weights()
|
253 |
+
adapter_names = []
|
254 |
+
for lora_info in lora_strings:
|
255 |
+
lora_repo = lora_info.get("repo")
|
256 |
+
weights = lora_info.get("weights")
|
257 |
+
adapter_name = lora_info.get("adapter_name")
|
258 |
+
if lora_repo and weights and adapter_name:
|
259 |
+
# 调用 pipe.load_lora_weights() 方法加载权重
|
260 |
+
pipe.load_lora_weights(lora_repo, weight_name=weights, adapter_name=adapter_name)
|
261 |
+
adapter_names.append(adapter_name)
|
262 |
+
# 将 adapter_name 设置为模型的属性
|
263 |
+
setattr(pipe.unet, 'adapter_name', adapter_name)
|
264 |
+
else:
|
265 |
+
raise ValueError("Invalid LoRA string format. Each item must have 'repo', 'weights', and 'adapter_name' keys.")
|
266 |
+
adapter_weights = [lora_scale] * len(adapter_names)
|
267 |
+
# 调用 pipeline.set_adapters 方法设置 adapter 和对应权重
|
268 |
+
pipe.set_adapters(adapter_names, adapter_weights=adapter_weights)
|
269 |
+
|
270 |
+
# 确保 adapter_names 的数量与角色数量匹配
|
271 |
+
if len(adapter_names) != num_characters:
|
272 |
+
raise ValueError("The number of LoRA adapters must match the number of characters.")
|
273 |
+
|
274 |
# Set random seed for reproducibility
|
275 |
if randomize_seed:
|
276 |
+
with calculateDuration("Set random seed"):
|
277 |
+
seed = random.randint(0, MAX_SEED)
|
|
|
278 |
|
279 |
+
# 编码提示词
|
280 |
+
with calculateDuration("Encoding prompts"):
|
281 |
+
# 编码背景提示词
|
282 |
+
bg_text_input = pipe.tokenizer(prompt_bg, return_tensors="pt").to("cuda")
|
283 |
+
bg_embeddings = pipe.text_encoder(bg_text_input.input_ids.to(device))[0]
|
|
|
284 |
|
285 |
+
# 编码角色提示词
|
286 |
+
character_embeddings = []
|
287 |
+
for prompt in character_prompts:
|
288 |
+
char_text_input = pipe.tokenizer(prompt, return_tensors="pt").to("cuda")
|
289 |
+
char_embeddings = pipe.text_encoder(char_text_input.input_ids.to(device))[0]
|
290 |
+
character_embeddings.append(char_embeddings)
|
291 |
+
|
292 |
+
# 编码互动细节提示词
|
293 |
+
details_text_input = pipe.tokenizer(prompt_details, return_tensors="pt").to("cuda")
|
294 |
+
details_embeddings = pipe.text_encoder(details_text_input.input_ids.to(device))[0]
|
295 |
+
|
296 |
+
# 合并背景和互动细节的嵌入
|
297 |
+
prompt_embeddings = torch.cat([bg_embeddings, details_embeddings], dim=1)
|
298 |
+
|
299 |
+
# 解析角色位置
|
300 |
+
character_infos = []
|
301 |
+
for position_str in character_positions:
|
302 |
+
info = parse_character_position(position_str)
|
303 |
+
character_infos.append(info)
|
304 |
+
|
305 |
+
# 创建角色的掩码
|
306 |
+
masks = []
|
307 |
+
for info in character_infos:
|
308 |
+
mask = create_attention_mask(width, height, info['location'], info['offset'], info['area'])
|
309 |
+
masks.append(mask)
|
310 |
+
|
311 |
+
# 替换注意力处理器
|
312 |
+
replace_attention_processors(pipe, masks, character_embeddings, adapter_names)
|
313 |
+
|
314 |
+
# Generate image
|
315 |
+
final_image = generate_image_with_embeddings(prompt_embeddings, steps, seed, cfg_scale, width, height, progress)
|
316 |
+
|
317 |
+
# 您可以在此处添加上传图片的代码
|
318 |
+
result = {"status": "success", "message": "Image generated"}
|
319 |
|
320 |
+
progress(100, "Completed!")
|
321 |
|
322 |
+
return final_image, seed, json.dumps(result)
|
323 |
|
324 |
+
# Gradio 界面
|
325 |
css="""
|
326 |
#col-container {
|
327 |
margin: 0 auto;
|
|
|
330 |
"""
|
331 |
|
332 |
with gr.Blocks(css=css) as demo:
|
333 |
+
gr.Markdown("Flux with LoRA")
|
334 |
with gr.Row():
|
335 |
|
336 |
with gr.Column():
|
337 |
+
|
338 |
+
prompt_bg = gr.Text(label="Background Prompt", placeholder="Enter background/scene prompt", lines=2)
|
339 |
+
character_prompts = gr.Text(label="Character Prompts (JSON List)", placeholder='["Character 1 prompt", "Character 2 prompt"]', lines=5)
|
340 |
+
character_positions = gr.Text(label="Character Positions (JSON List)", placeholder='["Character 1 position", "Character 2 position"]', lines=5)
|
341 |
+
lora_strings_json = gr.Text(label="LoRA Strings (JSON List)", placeholder='[{"repo": "lora_repo1", "weights": "weights1", "adapter_name": "adapter_name1"}, {"repo": "lora_repo2", "weights": "weights2", "adapter_name": "adapter_name2"}]', lines=5)
|
342 |
+
prompt_details = gr.Text(label="Interaction Details", placeholder="Enter interaction details between characters", lines=2)
|
343 |
run_button = gr.Button("Run", scale=0)
|
344 |
|
345 |
with gr.Accordion("Advanced Settings", open=False):
|
|
|
350 |
lora_scale = gr.Slider(label="LoRA Scale", minimum=0, maximum=1, step=0.01, value=0.5)
|
351 |
|
352 |
with gr.Row():
|
353 |
+
width = gr.Slider(label="Width", minimum=256, maximum=1536, step=64, value=512)
|
354 |
+
height = gr.Slider(label="Height", minimum=256, maximum=1536, step=64, value=512)
|
355 |
|
356 |
with gr.Row():
|
357 |
+
cfg_scale = gr.Slider(label="CFG Scale", minimum=1, maximum=20, step=0.5, value=7.5)
|
358 |
steps = gr.Slider(label="Steps", minimum=1, maximum=50, step=1, value=28)
|
359 |
|
360 |
upload_to_r2 = gr.Checkbox(label="Upload to R2", value=False)
|
|
|
366 |
|
367 |
with gr.Column():
|
368 |
result = gr.Image(label="Result", show_label=False)
|
369 |
+
seed_output = gr.Text(label="Seed")
|
370 |
+
json_text = gr.Text(label="Result JSON")
|
371 |
+
|
372 |
+
inputs = [
|
373 |
+
prompt_bg,
|
374 |
+
character_prompts,
|
375 |
+
character_positions,
|
376 |
+
lora_strings_json,
|
377 |
+
prompt_details,
|
378 |
+
cfg_scale,
|
379 |
+
steps,
|
380 |
+
randomize_seed,
|
381 |
+
seed,
|
382 |
+
width,
|
383 |
+
height,
|
384 |
+
lora_scale,
|
385 |
+
upload_to_r2,
|
386 |
+
account_id,
|
387 |
+
access_key,
|
388 |
+
secret_key,
|
389 |
+
bucket
|
390 |
+
]
|
391 |
+
|
392 |
+
outputs = [result, seed_output, json_text]
|
393 |
+
|
394 |
+
run_button.click(
|
395 |
+
fn=run_lora,
|
396 |
+
inputs=inputs,
|
397 |
+
outputs=outputs
|
398 |
)
|
399 |
|
400 |
+
demo.queue().launch()
|