Spaces:
Running
on
Zero
Running
on
Zero
Create app7.py
Browse files
app7.py
ADDED
@@ -0,0 +1,213 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import os
|
2 |
+
import gradio as gr
|
3 |
+
import json
|
4 |
+
import logging
|
5 |
+
import torch
|
6 |
+
from PIL import Image
|
7 |
+
import spaces
|
8 |
+
from diffusers import DiffusionPipeline, AutoencoderTiny, AutoencoderKL, AutoPipelineForText2Image
|
9 |
+
import copy
|
10 |
+
import random
|
11 |
+
import time
|
12 |
+
from diffusers.models.transformers import FluxTransformer2DModel
|
13 |
+
import safetensors.torch
|
14 |
+
from transformers import CLIPModel, CLIPProcessor, CLIPTextModel, CLIPTokenizer, CLIPConfig, T5EncoderModel, T5Tokenizer
|
15 |
+
from live_preview_helpers import calculate_shift, retrieve_timesteps, flux_pipe_call_that_returns_an_iterable_of_images
|
16 |
+
from huggingface_hub import HfFileSystem, ModelCard
|
17 |
+
from safetensors.torch import load_file
|
18 |
+
from huggingface_hub import login
|
19 |
+
|
20 |
+
hf_token = os.environ.get("HF_TOKEN")
|
21 |
+
login(token=hf_token)
|
22 |
+
|
23 |
+
torch.set_float32_matmul_precision("medium")
|
24 |
+
|
25 |
+
# Load LoRAs from JSON file
|
26 |
+
with open('loras.json', 'r') as f:
|
27 |
+
loras = json.load(f)
|
28 |
+
|
29 |
+
# Initialize the base model
|
30 |
+
dtype = torch.bfloat16
|
31 |
+
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
|
32 |
+
base_model = "John6666/hyper-flux1-dev-fp8-flux"
|
33 |
+
taef1 = AutoencoderTiny.from_pretrained("madebyollin/taef1", torch_dtype=dtype).to(device)
|
34 |
+
good_vae = AutoencoderKL.from_pretrained("black-forest-labs/FLUX.1-dev", subfolder="vae", torch_dtype=dtype).to(device)
|
35 |
+
pipe = DiffusionPipeline.from_pretrained(base_model, torch_dtype=dtype, vae=good_vae).to(device)
|
36 |
+
|
37 |
+
model_id = ("zer0int/LongCLIP-GmP-ViT-L-14")
|
38 |
+
config = CLIPConfig.from_pretrained(model_id)
|
39 |
+
config.text_config.max_position_embeddings = 248
|
40 |
+
clip_model = CLIPModel.from_pretrained(model_id, torch_dtype=torch.bfloat16, config=config, ignore_mismatched_sizes=True)
|
41 |
+
clip_processor = CLIPProcessor.from_pretrained(model_id, padding="max_length", max_length=248)
|
42 |
+
pipe.tokenizer = clip_processor.tokenizer
|
43 |
+
pipe.text_encoder = clip_model.text_model
|
44 |
+
pipe.tokenizer_max_length = 248
|
45 |
+
pipe.text_encoder.dtype = torch.bfloat16
|
46 |
+
|
47 |
+
MAX_SEED = 2**32-1
|
48 |
+
|
49 |
+
class calculateDuration:
|
50 |
+
def __init__(self, activity_name=""):
|
51 |
+
self.activity_name = activity_name
|
52 |
+
|
53 |
+
def __enter__(self):
|
54 |
+
self.start_time = time.time()
|
55 |
+
return self
|
56 |
+
|
57 |
+
def __exit__(self, exc_type, exc_value, traceback):
|
58 |
+
self.end_time = time.time()
|
59 |
+
self.elapsed_time = self.end_time - self.start_time
|
60 |
+
if self.activity_name:
|
61 |
+
print(f"Elapsed time for {self.activity_name}: {self.elapsed_time:.6f} seconds")
|
62 |
+
else:
|
63 |
+
print(f"Elapsed time: {self.elapsed_time:.6f} seconds")
|
64 |
+
|
65 |
+
|
66 |
+
def update_selection(evt: gr.SelectData, width, height):
|
67 |
+
selected_lora = loras[evt.index]
|
68 |
+
new_placeholder = f"Prompt with activator word(s): '{selected_lora['trigger_word']}'! "
|
69 |
+
lora_repo = selected_lora["repo"]
|
70 |
+
lora_trigger = selected_lora['trigger_word']
|
71 |
+
updated_text = f"### Selected: [{lora_repo}](https://huggingface.co/{lora_repo}). Prompt using: '{lora_trigger}'!"
|
72 |
+
if "aspect" in selected_lora:
|
73 |
+
if selected_lora["aspect"] == "portrait":
|
74 |
+
width = 768
|
75 |
+
height = 1024
|
76 |
+
elif selected_lora["aspect"] == "landscape":
|
77 |
+
width = 1024
|
78 |
+
height = 768
|
79 |
+
return (
|
80 |
+
gr.update(placeholder=new_placeholder),
|
81 |
+
updated_text,
|
82 |
+
evt.index,
|
83 |
+
width,
|
84 |
+
height,
|
85 |
+
)
|
86 |
+
|
87 |
+
@spaces.GPU()
|
88 |
+
def generate_image(prompt, trigger_word, steps, seed, cfg_scale, width, height, lora_scale, progress):
|
89 |
+
pipe.to("cuda")
|
90 |
+
generator = torch.Generator(device="cuda").manual_seed(seed)
|
91 |
+
|
92 |
+
with calculateDuration("Generating image"):
|
93 |
+
# Generate image
|
94 |
+
image = pipe(
|
95 |
+
prompt=f"{prompt} {trigger_word}",
|
96 |
+
num_inference_steps=steps,
|
97 |
+
guidance_scale=cfg_scale,
|
98 |
+
width=width,
|
99 |
+
height=height,
|
100 |
+
generator=generator,
|
101 |
+
joint_attention_kwargs={"scale": lora_scale},
|
102 |
+
).images[0]
|
103 |
+
return image
|
104 |
+
|
105 |
+
def run_lora(prompt, cfg_scale, steps, selected_index, randomize_seed, seed, width, height, lora_scale, progress=gr.Progress(track_tqdm=True)):
|
106 |
+
if selected_index is None:
|
107 |
+
raise gr.Error("You must select a LoRA before proceeding.")
|
108 |
+
|
109 |
+
selected_lora = loras[selected_index]
|
110 |
+
lora_path = selected_lora["repo"]
|
111 |
+
trigger_word = selected_lora['trigger_word']
|
112 |
+
if(trigger_word):
|
113 |
+
if "trigger_position" in selected_lora:
|
114 |
+
if selected_lora["trigger_position"] == "prepend":
|
115 |
+
prompt_mash = f"{trigger_word} {prompt}"
|
116 |
+
else:
|
117 |
+
prompt_mash = f"{prompt} {trigger_word}"
|
118 |
+
else:
|
119 |
+
prompt_mash = f"{trigger_word} {prompt}"
|
120 |
+
else:
|
121 |
+
prompt_mash = prompt
|
122 |
+
|
123 |
+
# Load LoRA weights
|
124 |
+
with calculateDuration(f"Loading LoRA weights for {selected_lora['title']}"):
|
125 |
+
if "weights" in selected_lora:
|
126 |
+
pipe.load_lora_weights(lora_path, weight_name=selected_lora["weights"])
|
127 |
+
else:
|
128 |
+
pipe.load_lora_weights(lora_path)
|
129 |
+
|
130 |
+
# Set random seed for reproducibility
|
131 |
+
with calculateDuration("Randomizing seed"):
|
132 |
+
if randomize_seed:
|
133 |
+
seed = random.randint(0, MAX_SEED)
|
134 |
+
|
135 |
+
image = generate_image(prompt, trigger_word, steps, seed, cfg_scale, width, height, lora_scale, progress)
|
136 |
+
pipe.to("cpu")
|
137 |
+
pipe.unload_lora_weights()
|
138 |
+
return image, seed
|
139 |
+
|
140 |
+
run_lora.zerogpu = True
|
141 |
+
|
142 |
+
css = '''
|
143 |
+
#gen_btn{height: 100%}
|
144 |
+
#title{text-align: center}
|
145 |
+
#title h1{font-size: 3em; display:inline-flex; align-items:center}
|
146 |
+
#title img{width: 100px; margin-right: 0.5em}
|
147 |
+
#gallery .grid-wrap{height: 10vh}
|
148 |
+
'''
|
149 |
+
with gr.Blocks(theme=gr.themes.Soft(), css=css) as app:
|
150 |
+
title = gr.HTML(
|
151 |
+
"""<h1><img src="https://huggingface.co/spaces/multimodalart/flux-lora-the-explorer/resolve/main/flux_lora.png" alt="LoRA"> SOONfactory </h1>""",
|
152 |
+
elem_id="title",
|
153 |
+
)
|
154 |
+
# Info blob stating what the app is running
|
155 |
+
info_blob = gr.HTML(
|
156 |
+
"""<div id="info_blob"> Generative Models Celebrating the Unique Style & Sensibility of the Bay Area-based artist Jacqueline Trosclair (known to her friends as "Jax", "Starlic Jorca", & in countless ways)... </div>"""
|
157 |
+
)
|
158 |
+
|
159 |
+
# Info blob stating what the app is running
|
160 |
+
info_blob = gr.HTML(
|
161 |
+
"""<div id="info_blob"> To create new arts via a generative model variant inspired by a Jax's artworks, or a model merging in the style of her favorite artist Unica Zürn, choose a version below. </div>"""
|
162 |
+
)
|
163 |
+
selected_index = gr.State(None)
|
164 |
+
with gr.Row():
|
165 |
+
with gr.Column(scale=3):
|
166 |
+
prompt = gr.Textbox(label="Prompt", lines=1, placeholder="Select LoRa/Style & type prompt!")
|
167 |
+
with gr.Column(scale=1, elem_id="gen_column"):
|
168 |
+
generate_button = gr.Button("Generate", variant="primary", elem_id="gen_btn")
|
169 |
+
with gr.Row():
|
170 |
+
with gr.Column(scale=3):
|
171 |
+
selected_info = gr.Markdown("")
|
172 |
+
gallery = gr.Gallery(
|
173 |
+
[(item["image"], item["title"]) for item in loras],
|
174 |
+
label="LoRA Inventory",
|
175 |
+
allow_preview=False,
|
176 |
+
columns=3,
|
177 |
+
elem_id="gallery"
|
178 |
+
)
|
179 |
+
|
180 |
+
with gr.Column(scale=4):
|
181 |
+
result = gr.Image(label="Generated Image")
|
182 |
+
|
183 |
+
with gr.Row():
|
184 |
+
with gr.Accordion("Advanced Settings", open=True):
|
185 |
+
with gr.Column():
|
186 |
+
with gr.Row():
|
187 |
+
cfg_scale = gr.Slider(label="CFG Scale", minimum=0, maximum=20, step=0.5, value=3.0)
|
188 |
+
steps = gr.Slider(label="Steps", minimum=1, maximum=50, step=1, value=12)
|
189 |
+
|
190 |
+
with gr.Row():
|
191 |
+
width = gr.Slider(label="Width", minimum=256, maximum=1536, step=64, value=1024)
|
192 |
+
height = gr.Slider(label="Height", minimum=256, maximum=1536, step=64, value=1088)
|
193 |
+
|
194 |
+
with gr.Row():
|
195 |
+
randomize_seed = gr.Checkbox(True, label="Randomize seed")
|
196 |
+
seed = gr.Slider(label="Seed", minimum=0, maximum=MAX_SEED, step=1, value=0, randomize=True)
|
197 |
+
lora_scale = gr.Slider(label="LoRA Scale", minimum=0, maximum=2.0, step=0.01, value=1.05)
|
198 |
+
|
199 |
+
gallery.select(
|
200 |
+
update_selection,
|
201 |
+
inputs=[width, height],
|
202 |
+
outputs=[prompt, selected_info, selected_index, width, height]
|
203 |
+
)
|
204 |
+
|
205 |
+
gr.on(
|
206 |
+
triggers=[generate_button.click, prompt.submit],
|
207 |
+
fn=run_lora,
|
208 |
+
inputs=[prompt, cfg_scale, steps, selected_index, randomize_seed, seed, width, height, lora_scale],
|
209 |
+
outputs=[result, seed]
|
210 |
+
)
|
211 |
+
|
212 |
+
app.queue(default_concurrency_limit=2).launch(show_error=True)
|
213 |
+
app.launch()
|