File size: 8,792 Bytes
52dfea4
0e0ee20
 
 
 
c90c9ea
0e0ee20
 
c4064b3
91464a3
2bee297
7039ded
607d766
e2c1d93
2bee297
c90c9ea
 
 
c4064b3
 
 
91464a3
 
 
 
 
c4064b3
 
abd810c
 
 
 
c4064b3
0e0ee20
 
 
 
c4064b3
0e0ee20
abd810c
40de9c5
c4064b3
2bee297
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
abd810c
 
c4064b3
abd810c
 
0e0ee20
f3e96f9
c59400c
e2c1d93
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
5ecece8
0e0ee20
 
 
 
5ecece8
 
 
 
 
 
 
0e0ee20
 
 
5ecece8
 
 
0e0ee20
 
0b93385
5511562
aad2ddd
5b82e60
 
72cad74
 
 
 
5511562
72cad74
 
 
 
 
c4064b3
72cad74
 
 
5511562
0e0ee20
 
 
 
 
 
 
 
b488680
e2c1d93
 
 
 
c59400c
0e0ee20
e2c1d93
 
 
fd8e800
5511562
aad2ddd
72cad74
 
0e0ee20
0b93385
 
1441e58
07d3eff
8648a3b
504da62
 
8648a3b
07d3eff
504da62
 
3c4321b
504da62
38d4ed7
 
 
877fae7
504da62
c6fd2a7
 
 
5511562
c6fd2a7
0e0ee20
d6802e8
5511562
38d4ed7
5511562
 
02302e4
07d3eff
0e0ee20
db98dea
1fff27d
0e0ee20
 
38d4ed7
0e0ee20
8648a3b
 
0e0ee20
1fff27d
db98dea
8dce9c7
0e0ee20
 
38d4ed7
2c6d128
 
40de9c5
756113f
2c6d128
 
5511562
 
2c6d128
 
 
 
5511562
0e0ee20
5ecece8
 
751429f
5ecece8
 
07d3eff
 
 
0e0ee20
5511562
7fb9e28
0e0ee20
 
38d4ed7
 
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
224
225
226
import os
import gradio as gr
import json
import logging
import torch
from os import path
from PIL import Image
import spaces
from diffusers import DiffusionPipeline, AutoPipelineForText2Image
from diffusers import StableDiffusion3Pipeline, FlowMatchEulerDiscreteScheduler, SD3Transformer2DModel # pip install diffusers>=0.31.0
from transformers import CLIPModel, CLIPProcessor, CLIPTextModel, CLIPTokenizer, CLIPConfig, T5EncoderModel, T5Tokenizer
import copy
import random
import time
from huggingface_hub import HfFileSystem, ModelCard
from huggingface_hub import login, hf_hub_download
import safetensors.torch
from safetensors.torch import load_file
hf_token = os.environ.get("HF_TOKEN")
login(token=hf_token)

cache_path = path.join(path.dirname(path.abspath(__file__)), "models")
os.environ["TRANSFORMERS_CACHE"] = cache_path
os.environ["HF_HUB_CACHE"] = cache_path
os.environ["HF_HOME"] = cache_path

torch.set_float32_matmul_precision("high")

#torch._inductor.config.conv_1x1_as_mm = True
#torch._inductor.config.coordinate_descent_tuning = True
#torch._inductor.config.epilogue_fusion = False
#torch._inductor.config.coordinate_descent_check_all_directions = True


# Load LoRAs from JSON file
with open('loras.json', 'r') as f:
    loras = json.load(f)
    
# Initialize the base model
#base_model = "stabilityai/stable-diffusion-3.5-large"
pipe = AutoPipelineForText2Image.from_pretrained("stabilityai/stable-diffusion-3.5-large-turbo", torch_dtype=torch.bfloat16)

clipmodel = 'norm'
if clipmodel == "long":
    model_id = "zer0int/LongCLIP-GmP-ViT-L-14"
    config = CLIPConfig.from_pretrained(model_id)
    maxtokens = 248
if clipmodel == "norm":
    model_id = "zer0int/CLIP-GmP-ViT-L-14"
    config = CLIPConfig.from_pretrained(model_id)
    maxtokens = 77
clip_model = CLIPModel.from_pretrained(model_id, torch_dtype=torch.bfloat16, config=config, ignore_mismatched_sizes=True).to("cuda")
clip_processor = CLIPProcessor.from_pretrained(model_id, padding="max_length", max_length=maxtokens, ignore_mismatched_sizes=True, return_tensors="pt", truncation=True)

pipe.tokenizer = clip_processor.tokenizer
pipe.text_encoder = clip_model.text_model
pipe.tokenizer_max_length = maxtokens
pipe.text_encoder.dtype = torch.bfloat16


#pipe.transformer.to(memory_format=torch.channels_last)
#pipe.vae.to(memory_format=torch.channels_last)

#pipe.transformer = torch.compile(pipe.transformer, mode="max-autotune", fullgraph=True)
#pipe.vae.decode = torch.compile(pipe.vae.decode, mode="max-autotune", fullgraph=True)

MAX_SEED = 2**32-1

class calculateDuration:
    def __init__(self, activity_name=""):
        self.activity_name = activity_name

    def __enter__(self):
        self.start_time = time.time()
        return self
    
    def __exit__(self, exc_type, exc_value, traceback):
        self.end_time = time.time()
        self.elapsed_time = self.end_time - self.start_time
        if self.activity_name:
            print(f"Elapsed time for {self.activity_name}: {self.elapsed_time:.6f} seconds")
        else:
            print(f"Elapsed time: {self.elapsed_time:.6f} seconds")


def update_selection(evt: gr.SelectData, width, height):
    selected_lora = loras[evt.index]
    new_placeholder = f"Type a prompt for {selected_lora['title']}"
    lora_repo = selected_lora["repo"]
    updated_text = f"### Selected: [{lora_repo}](https://huggingface.co/{lora_repo}) ✨"
    if "aspect" in selected_lora:
        if selected_lora["aspect"] == "portrait":
            width = 768
            height = 1024
        elif selected_lora["aspect"] == "landscape":
            width = 1024
            height = 768
    return (
        gr.update(placeholder=new_placeholder),
        updated_text,
        evt.index,
        width,
        height,
    )

@spaces.GPU(duration=70)
def infer(prompt, negative_prompt, trigger_word, steps, seed, cfg_scale, width, height, lora_scale, progress):
    pipe.to("cuda")
    generator = torch.Generator(device="cuda").manual_seed(seed)
    
    with calculateDuration("Generating image"):
        # Generate image
        image = pipe(
            prompt=f"{prompt} {trigger_word}",
            negative_prompt=negative_prompt,
            num_inference_steps=steps,
            guidance_scale=cfg_scale,
            width=width,
            height=height,
            generator=generator,
            joint_attention_kwargs={"scale": lora_scale},
        ).images[0]
    return image

def run_lora(prompt, negative_prompt, cfg_scale, steps, selected_index, randomize_seed, seed, width, height, lora_scale, progress=gr.Progress(track_tqdm=True)):
    if selected_index is None:
        raise gr.Error("You must select a LoRA before proceeding.")

    selected_lora = loras[selected_index]
    lora_path = selected_lora["repo"]
    trigger_word = selected_lora["trigger_word"]

    # Load LoRA weights
    with calculateDuration(f"Loading LoRA weights for {selected_lora['title']}"):
        if "weights" in selected_lora:
            pipe.load_lora_weights(lora_path, weight_name=selected_lora["weights"])
        else:
            pipe.load_lora_weights(lora_path)
        
    # Set random seed for reproducibility
    with calculateDuration("Randomizing seed"):
        if randomize_seed:
            seed = random.randint(0, MAX_SEED)
    
    image = infer(prompt, negative_prompt, trigger_word, steps, seed, cfg_scale, width, height, lora_scale, progress)
    pipe.to("cpu")
    pipe.unload_lora_weights()
    return image, seed  

run_lora.zerogpu = True

css = '''
#gen_btn{height: 100%}
#title{text-align: center}
#title h1{font-size: 3em; display:inline-flex; align-items:center}
#title img{width: 100px; margin-right: 0.5em}
#gallery .grid-wrap{height: 10vh}
'''
with gr.Blocks(theme=gr.themes.Soft(), css=css) as app:
    title = gr.HTML(
        """<h1><img src="https://huggingface.co/AlekseyCalvin/StabledHSTorY_SD3.5_LoRA_V2_rank256/resolve/main/acs62v.png" alt="LoRA">Stabled HSTorY S.D.3.5 LoRAs</h1>""",
        elem_id="title",
    )
    	    # Info blob stating what the app is running
    info_blob = gr.HTML(
        """<div id="info_blob">SOON®'s curated LoRa Gallery & Art Manufactory.|Now testing HST-triggered historic photo-trained LoRAs for Stable Diffusion 3.5.</div>"""
    )

        # Info blob stating what the app is running
    info_blob = gr.HTML(
        """<div id="info_blob">Prephrase prompts w/: "HST style autochrome photo" </div>"""
    )
    selected_index = gr.State(None)
    with gr.Row():
        with gr.Column(scale=2):
            prompt = gr.Textbox(label="Prompt", lines=1, placeholder="Select LoRa/Style & type prompt!")
        with gr.Column(scale=2):
            negative_prompt = gr.Textbox(label="Negative Prompt", lines=1, placeholder="What to exclude!")
        with gr.Column(scale=1, elem_id="gen_column"):
            generate_button = gr.Button("Generate", variant="primary", elem_id="gen_btn")
    with gr.Row():
        with gr.Column(scale=3):
            selected_info = gr.Markdown("")
            gallery = gr.Gallery(
                [(item["image"], item["title"]) for item in loras],
                label="LoRA Inventory",
                allow_preview=False,
                columns=3,
                elem_id="gallery"
            )
            
        with gr.Column(scale=4):
            result = gr.Image(label="Generated Image")

    with gr.Row():
        with gr.Accordion("Advanced Settings", open=True):
            with gr.Column():
                with gr.Row():
                    cfg_scale = gr.Slider(label="CFG Scale", minimum=0, maximum=20, step=.1, value=1.0)
                    steps = gr.Slider(label="Steps", minimum=1, maximum=50, step=1, value=7)
                
                with gr.Row():
                    width = gr.Slider(label="Width", minimum=256, maximum=1536, step=64, value=1024)
                    height = gr.Slider(label="Height", minimum=256, maximum=1536, step=64, value=1024)
                
                with gr.Row():
                    randomize_seed = gr.Checkbox(True, label="Randomize seed")
                    seed = gr.Slider(label="Seed", minimum=0, maximum=MAX_SEED, step=1, value=0, randomize=True)
                    lora_scale = gr.Slider(label="LoRA Scale", minimum=0, maximum=3.0, step=0.01, value=1.0)

    gallery.select(
        update_selection,
        inputs=[width, height],
        outputs=[prompt, selected_info, selected_index, width, height]
    )

    gr.on(
        triggers=[generate_button.click, prompt.submit],
        fn=run_lora,
        inputs=[prompt, negative_prompt, cfg_scale, steps, selected_index, randomize_seed, seed, width, height, lora_scale],
        outputs=[result, seed]
    )

app.queue(default_concurrency_limit=2).launch(show_error=True)
app.launch()