Update app.py
Browse files
app.py
CHANGED
|
@@ -2,7 +2,7 @@
|
|
| 2 |
#und deutlich besserem Prompt-Verständnis - (Änderung Architektur).
|
| 3 |
#Eine deutsche Alternative zur Umsetzung von Text-Bild zu Bild ist Flux - mit einer völlig anderen Architektur als SD!
|
| 4 |
import gradio as gr
|
| 5 |
-
from diffusers import StableDiffusionPipeline, StableDiffusionImg2ImgPipeline
|
| 6 |
from diffusers import StableDiffusionInpaintPipeline
|
| 7 |
from controlnet_module import controlnet_processor
|
| 8 |
import torch
|
|
@@ -12,12 +12,10 @@ import os
|
|
| 12 |
import tempfile
|
| 13 |
import random
|
| 14 |
|
| 15 |
-
|
| 16 |
-
|
| 17 |
# === OPTIMIERTE EINSTELLUNGEN ===
|
| 18 |
device = "cuda" if torch.cuda.is_available() else "cpu"
|
| 19 |
torch_dtype = torch.float16 if device == "cuda" else torch.float32
|
| 20 |
-
IMG_SIZE =
|
| 21 |
|
| 22 |
print(f"Running on: {device}")
|
| 23 |
|
|
@@ -65,21 +63,44 @@ pipe_img2img = None
|
|
| 65 |
def load_txt2img():
|
| 66 |
global pipe_txt2img
|
| 67 |
if pipe_txt2img is None:
|
| 68 |
-
|
| 69 |
-
|
| 70 |
-
|
| 71 |
-
|
| 72 |
-
|
| 73 |
-
|
| 74 |
-
|
| 75 |
-
|
| 76 |
-
|
| 77 |
-
|
| 78 |
-
|
| 79 |
-
|
| 80 |
-
|
| 81 |
-
|
| 82 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 83 |
return pipe_txt2img
|
| 84 |
|
| 85 |
def load_img2img():
|
|
@@ -89,15 +110,13 @@ def load_img2img():
|
|
| 89 |
try:
|
| 90 |
pipe_img2img = StableDiffusionInpaintPipeline.from_pretrained(
|
| 91 |
"runwayml/stable-diffusion-inpainting",
|
| 92 |
-
#"stabilityai/stable-diffusion-2-inpainting", # Neues Modell
|
| 93 |
torch_dtype=torch_dtype,
|
| 94 |
-
|
| 95 |
-
allow_pickle=False,
|
| 96 |
safety_checker=None,
|
| 97 |
-
#clean_up_tokenization_spaces=False #benötigt neue Transformer-Version
|
| 98 |
).to(device)
|
| 99 |
except Exception as e:
|
| 100 |
-
print(f"Fehler beim Laden des Modells: {e}")
|
| 101 |
raise
|
| 102 |
|
| 103 |
|
|
@@ -111,7 +130,8 @@ def load_img2img():
|
|
| 111 |
|
| 112 |
pipe_img2img.enable_attention_slicing()
|
| 113 |
pipe_img2img.enable_vae_tiling()
|
| 114 |
-
pipe_img2img
|
|
|
|
| 115 |
|
| 116 |
return pipe_img2img
|
| 117 |
|
|
@@ -126,7 +146,7 @@ class TextToImageProgressCallback:
|
|
| 126 |
"""Neue Callback-Signatur für diffusers >= 1.0.0"""
|
| 127 |
self.current_step = step + 1
|
| 128 |
progress_percent = (step / self.total_steps) * 100
|
| 129 |
-
self.progress(progress_percent / 100, desc="Generierung läuft
|
| 130 |
return callback_kwargs
|
| 131 |
|
| 132 |
class ImageToImageProgressCallback:
|
|
@@ -152,7 +172,7 @@ class ImageToImageProgressCallback:
|
|
| 152 |
print(f"🎯 INTERNE STEP-AUSGABE: Strength {self.strength} → {self.actual_total_steps} tatsächliche Denoising-Schritte")
|
| 153 |
|
| 154 |
progress_percent = (step / self.actual_total_steps) * 100
|
| 155 |
-
self.progress(progress_percent / 100, desc="Generierung läuft
|
| 156 |
return callback_kwargs
|
| 157 |
|
| 158 |
# === NEUE FUNKTIONEN FÜR DIE FEATURES ===
|
|
@@ -239,9 +259,7 @@ def text_to_image(prompt, steps, guidance_scale, progress=gr.Progress()):
|
|
| 239 |
print(f"Starting generation for: {prompt}")
|
| 240 |
start_time = time.time()
|
| 241 |
|
| 242 |
-
|
| 243 |
-
progress(0, desc="Generierung läuft - CPU benötigt bis zu 20 Minuten!")
|
| 244 |
-
|
| 245 |
pipe = load_txt2img()
|
| 246 |
|
| 247 |
# ZUFÄLLIGER SEED für Variation
|
|
@@ -249,28 +267,27 @@ def text_to_image(prompt, steps, guidance_scale, progress=gr.Progress()):
|
|
| 249 |
generator = torch.Generator(device=device).manual_seed(seed)
|
| 250 |
print(f"Using seed: {seed}")
|
| 251 |
|
| 252 |
-
# NEUE Callback-Implementierung
|
| 253 |
callback = TextToImageProgressCallback(progress, steps)
|
| 254 |
|
|
|
|
| 255 |
image = pipe(
|
| 256 |
prompt=prompt,
|
| 257 |
-
height=IMG_SIZE
|
| 258 |
-
width=IMG_SIZE
|
| 259 |
num_inference_steps=int(steps),
|
| 260 |
guidance_scale=guidance_scale,
|
| 261 |
generator=generator,
|
| 262 |
-
callback_on_step_end=callback,
|
| 263 |
-
callback_on_step_end_tensor_inputs=[],
|
| 264 |
).images[0]
|
| 265 |
|
| 266 |
end_time = time.time()
|
| 267 |
print(f"Bild generiert in {end_time - start_time:.2f} Sekunden")
|
| 268 |
|
| 269 |
-
# Robuste Zwischenspeicherung
|
| 270 |
return image
|
| 271 |
|
| 272 |
except Exception as e:
|
| 273 |
-
print(f"Fehler: {e}")
|
| 274 |
import traceback
|
| 275 |
traceback.print_exc()
|
| 276 |
return None
|
|
@@ -299,12 +316,10 @@ def img_to_image(image, prompt, neg_prompt, strength, steps, guidance_scale,
|
|
| 299 |
|
| 300 |
# CONTROLNET-STRENGTH ANPASSEN ABHÄNGIG VOM MODUS
|
| 301 |
if face_preserve:
|
| 302 |
-
|
| 303 |
-
controlnet_strength = adj_strength * 0.8 # 80% für kombinierte OpenPose + Canny
|
| 304 |
print(f"🎯 ControlNet Modus: Umgebung beibehalten (Strength = {controlnet_strength:.3f})")
|
| 305 |
else:
|
| 306 |
-
|
| 307 |
-
controlnet_strength = adj_strength * 0.5 # 50% für OpenPose
|
| 308 |
print(f"🎯 ControlNet Modus: Person beibehalten (Strength = {controlnet_strength:.3f})")
|
| 309 |
|
| 310 |
controlnet_steps = min(25, int(steps * 0.8))
|
|
@@ -324,7 +339,7 @@ def img_to_image(image, prompt, neg_prompt, strength, steps, guidance_scale,
|
|
| 324 |
guidance_scale=guidance_scale,
|
| 325 |
controlnet_strength=controlnet_strength,
|
| 326 |
progress=progress,
|
| 327 |
-
keep_environment=face_preserve
|
| 328 |
)
|
| 329 |
|
| 330 |
print(f"✅ ControlNet Output erhalten: {type(controlnet_output)}")
|
|
@@ -335,10 +350,9 @@ def img_to_image(image, prompt, neg_prompt, strength, steps, guidance_scale,
|
|
| 335 |
# -------------------------------
|
| 336 |
progress(0.3, desc="ControlNet abgeschlossen – starte Inpaint...")
|
| 337 |
|
| 338 |
-
pipe = load_img2img()
|
| 339 |
|
| 340 |
-
|
| 341 |
-
img_resized = inpaint_input.convert("RGB").resize((512, 512)) # Bleibt bei 512 für Inpaint
|
| 342 |
|
| 343 |
adj_guidance = min(guidance_scale, 12.0)
|
| 344 |
seed = random.randint(0, 2**32 - 1)
|
|
@@ -351,7 +365,7 @@ def img_to_image(image, prompt, neg_prompt, strength, steps, guidance_scale,
|
|
| 351 |
mask = None
|
| 352 |
if bbox_x1 and bbox_y1 and bbox_x2 and bbox_y2:
|
| 353 |
orig_w, orig_h = image.size
|
| 354 |
-
scale_x, scale_y = 512 / orig_w, 512 / orig_h
|
| 355 |
bbox_coords = [
|
| 356 |
int(bbox_x1 * scale_x),
|
| 357 |
int(bbox_y1 * scale_y),
|
|
@@ -398,7 +412,6 @@ def img_to_image(image, prompt, neg_prompt, strength, steps, guidance_scale,
|
|
| 398 |
import traceback
|
| 399 |
traceback.print_exc()
|
| 400 |
return None
|
| 401 |
-
|
| 402 |
|
| 403 |
def update_bbox_from_image(image):
|
| 404 |
"""Aktualisiert die Bounding-Box-Koordinaten wenn ein Bild hochgeladen wird"""
|
|
@@ -480,7 +493,6 @@ def main_ui():
|
|
| 480 |
"""
|
| 481 |
) as demo:
|
| 482 |
|
| 483 |
-
# --- Hauptanwendungsbereich (zunächst versteckt) ---
|
| 484 |
with gr.Column(visible=True) as content_area:
|
| 485 |
with gr.Tab("Text zu Bild"):
|
| 486 |
gr.Markdown("**Beschreibe dein gewünschtes Bild:**")
|
|
@@ -524,7 +536,6 @@ def main_ui():
|
|
| 524 |
with gr.Tab("Bild zu Bild"):
|
| 525 |
gr.Markdown("**Lade ein Bild hoch und beschreibe die gewünschte Veränderung:**")
|
| 526 |
|
| 527 |
-
# NEUE ANORDNUNG: Eingabebild und Live-Vorschau nebeneinander
|
| 528 |
with gr.Row():
|
| 529 |
with gr.Column():
|
| 530 |
img_input = gr.Image(
|
|
@@ -542,7 +553,6 @@ def main_ui():
|
|
| 542 |
show_download_button=False
|
| 543 |
)
|
| 544 |
|
| 545 |
-
# DARUNTER: Checkbox Gesicht/Person oder Umgebung ändern
|
| 546 |
with gr.Row():
|
| 547 |
face_preserve = gr.Checkbox(
|
| 548 |
label="Schutz",
|
|
@@ -550,7 +560,6 @@ def main_ui():
|
|
| 550 |
info="🟢 Checkbox AN: Alles AUSSERHALB des gelben Rahmens verändern | 🔴 Checkbox AUS: Nur INNERHALB des gelben Rahmens verändern"
|
| 551 |
)
|
| 552 |
|
| 553 |
-
# DARUNTER: Bildelementbereich anpassen
|
| 554 |
with gr.Row():
|
| 555 |
gr.Markdown("**Bildelementbereich anpassen**")
|
| 556 |
|
|
@@ -581,7 +590,6 @@ def main_ui():
|
|
| 581 |
info="Untere Kante des Bildelementbereichs"
|
| 582 |
)
|
| 583 |
|
| 584 |
-
# DARUNTER: Prompt und Negativ-Prompt
|
| 585 |
with gr.Row():
|
| 586 |
with gr.Column():
|
| 587 |
img_prompt = gr.Textbox(
|
|
@@ -598,7 +606,6 @@ def main_ui():
|
|
| 598 |
info="Was soll vermieden werden? Unerwünschte Elemente auflisten."
|
| 599 |
)
|
| 600 |
|
| 601 |
-
# DARUNTER: Veränderungsstärke, Inferenzschritte, Promptstärke
|
| 602 |
with gr.Row():
|
| 603 |
with gr.Column():
|
| 604 |
strength_slider = gr.Slider(
|
|
@@ -628,7 +635,6 @@ def main_ui():
|
|
| 628 |
"• **Koordinaten nur bei erkennbaren Verzerrungen anpassen** (Bereiche leicht verschieben)"
|
| 629 |
)
|
| 630 |
|
| 631 |
-
|
| 632 |
transform_btn = gr.Button("Bild transformieren", variant="primary")
|
| 633 |
|
| 634 |
with gr.Row():
|
|
@@ -638,15 +644,12 @@ def main_ui():
|
|
| 638 |
type="pil"
|
| 639 |
)
|
| 640 |
|
| 641 |
-
# NEUE: Event-Handler für alle Live-Updates
|
| 642 |
-
# Bild-Upload: Auto-Koordinaten + Vorschau
|
| 643 |
img_input.change(
|
| 644 |
fn=process_image_upload,
|
| 645 |
inputs=[img_input],
|
| 646 |
outputs=[preview_output, bbox_x1, bbox_y1, bbox_x2, bbox_y2]
|
| 647 |
)
|
| 648 |
|
| 649 |
-
# Live-Updates bei Koordinaten-Änderungen
|
| 650 |
coordinate_inputs = [img_input, bbox_x1, bbox_y1, bbox_x2, bbox_y2, face_preserve]
|
| 651 |
|
| 652 |
bbox_x1.change(
|
|
@@ -673,14 +676,12 @@ def main_ui():
|
|
| 673 |
outputs=preview_output
|
| 674 |
)
|
| 675 |
|
| 676 |
-
# Live-Update bei Modus-Änderung
|
| 677 |
face_preserve.change(
|
| 678 |
fn=update_live_preview,
|
| 679 |
inputs=coordinate_inputs,
|
| 680 |
outputs=preview_output
|
| 681 |
)
|
| 682 |
|
| 683 |
-
# Transform-Button (UNVERÄNDERT - gibt OUTPUT zurück)
|
| 684 |
transform_btn.click(
|
| 685 |
fn=img_to_image,
|
| 686 |
inputs=[
|
|
@@ -697,7 +698,11 @@ def main_ui():
|
|
| 697 |
|
| 698 |
if __name__ == "__main__":
|
| 699 |
demo = main_ui()
|
| 700 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
| 701 |
demo.launch(
|
| 702 |
server_name="0.0.0.0",
|
| 703 |
server_port=7860,
|
|
|
|
| 2 |
#und deutlich besserem Prompt-Verständnis - (Änderung Architektur).
|
| 3 |
#Eine deutsche Alternative zur Umsetzung von Text-Bild zu Bild ist Flux - mit einer völlig anderen Architektur als SD!
|
| 4 |
import gradio as gr
|
| 5 |
+
from diffusers import StableDiffusionPipeline, StableDiffusionImg2ImgPipeline
|
| 6 |
from diffusers import StableDiffusionInpaintPipeline
|
| 7 |
from controlnet_module import controlnet_processor
|
| 8 |
import torch
|
|
|
|
| 12 |
import tempfile
|
| 13 |
import random
|
| 14 |
|
|
|
|
|
|
|
| 15 |
# === OPTIMIERTE EINSTELLUNGEN ===
|
| 16 |
device = "cuda" if torch.cuda.is_available() else "cpu"
|
| 17 |
torch_dtype = torch.float16 if device == "cuda" else torch.float32
|
| 18 |
+
IMG_SIZE = 512 # Jetzt 512x512 für Realistic Vision
|
| 19 |
|
| 20 |
print(f"Running on: {device}")
|
| 21 |
|
|
|
|
| 63 |
def load_txt2img():
|
| 64 |
global pipe_txt2img
|
| 65 |
if pipe_txt2img is None:
|
| 66 |
+
try:
|
| 67 |
+
print("Loading Realistic Vision V6.0 for high-quality 512x512...")
|
| 68 |
+
pipe_txt2img = StableDiffusionPipeline.from_pretrained(
|
| 69 |
+
"SG161222/Realistic_Vision_V6.0_B1",
|
| 70 |
+
torch_dtype=torch_dtype,
|
| 71 |
+
safety_checker=None,
|
| 72 |
+
requires_safety_checker=False,
|
| 73 |
+
add_watermarker=False,
|
| 74 |
+
use_safetensors=True, # Sicherheitsproblem behoben
|
| 75 |
+
variant="fp16" if torch_dtype == torch.float16 else None,
|
| 76 |
+
).to(device)
|
| 77 |
+
|
| 78 |
+
from diffusers import DPMSolverMultistepScheduler
|
| 79 |
+
pipe_txt2img.scheduler = DPMSolverMultistepScheduler.from_config(
|
| 80 |
+
pipe_txt2img.scheduler.config,
|
| 81 |
+
use_karras_sigmas=True,
|
| 82 |
+
algorithm_type="sde-dpmsolver++"
|
| 83 |
+
)
|
| 84 |
+
|
| 85 |
+
# T4 OPTIMIERUNGEN
|
| 86 |
+
pipe_txt2img.enable_attention_slicing()
|
| 87 |
+
pipe_txt2img.enable_vae_slicing()
|
| 88 |
+
if hasattr(pipe_txt2img, 'vae'):
|
| 89 |
+
pipe_txt2img.vae.enable_slicing()
|
| 90 |
+
|
| 91 |
+
print("✅ Realistic Vision V6.0 erfolgreich geladen")
|
| 92 |
+
|
| 93 |
+
except Exception as e:
|
| 94 |
+
print(f"❌ Fehler beim Laden von Realistic Vision: {e}")
|
| 95 |
+
print("🔄 Fallback auf SD 1.5...")
|
| 96 |
+
# Fallback auf Standard SD 1.5
|
| 97 |
+
pipe_txt2img = StableDiffusionPipeline.from_pretrained(
|
| 98 |
+
"runwayml/stable-diffusion-v1-5",
|
| 99 |
+
torch_dtype=torch_dtype,
|
| 100 |
+
use_safetensors=True,
|
| 101 |
+
).to(device)
|
| 102 |
+
pipe_txt2img.enable_attention_slicing()
|
| 103 |
+
|
| 104 |
return pipe_txt2img
|
| 105 |
|
| 106 |
def load_img2img():
|
|
|
|
| 110 |
try:
|
| 111 |
pipe_img2img = StableDiffusionInpaintPipeline.from_pretrained(
|
| 112 |
"runwayml/stable-diffusion-inpainting",
|
|
|
|
| 113 |
torch_dtype=torch_dtype,
|
| 114 |
+
use_safetensors=True, # Sicherheitsproblem behoben
|
| 115 |
+
allow_pickle=False,
|
| 116 |
safety_checker=None,
|
|
|
|
| 117 |
).to(device)
|
| 118 |
except Exception as e:
|
| 119 |
+
print(f"Fehler beim Laden des Inpainting-Modells: {e}")
|
| 120 |
raise
|
| 121 |
|
| 122 |
|
|
|
|
| 130 |
|
| 131 |
pipe_img2img.enable_attention_slicing()
|
| 132 |
pipe_img2img.enable_vae_tiling()
|
| 133 |
+
if hasattr(pipe_img2img, 'vae_slicing'):
|
| 134 |
+
pipe_img2img.vae_slicing = True
|
| 135 |
|
| 136 |
return pipe_img2img
|
| 137 |
|
|
|
|
| 146 |
"""Neue Callback-Signatur für diffusers >= 1.0.0"""
|
| 147 |
self.current_step = step + 1
|
| 148 |
progress_percent = (step / self.total_steps) * 100
|
| 149 |
+
self.progress(progress_percent / 100, desc="Generierung läuft...")
|
| 150 |
return callback_kwargs
|
| 151 |
|
| 152 |
class ImageToImageProgressCallback:
|
|
|
|
| 172 |
print(f"🎯 INTERNE STEP-AUSGABE: Strength {self.strength} → {self.actual_total_steps} tatsächliche Denoising-Schritte")
|
| 173 |
|
| 174 |
progress_percent = (step / self.actual_total_steps) * 100
|
| 175 |
+
self.progress(progress_percent / 100, desc="Generierung läuft...")
|
| 176 |
return callback_kwargs
|
| 177 |
|
| 178 |
# === NEUE FUNKTIONEN FÜR DIE FEATURES ===
|
|
|
|
| 259 |
print(f"Starting generation for: {prompt}")
|
| 260 |
start_time = time.time()
|
| 261 |
|
| 262 |
+
progress(0, desc="Lade Modell...")
|
|
|
|
|
|
|
| 263 |
pipe = load_txt2img()
|
| 264 |
|
| 265 |
# ZUFÄLLIGER SEED für Variation
|
|
|
|
| 267 |
generator = torch.Generator(device=device).manual_seed(seed)
|
| 268 |
print(f"Using seed: {seed}")
|
| 269 |
|
|
|
|
| 270 |
callback = TextToImageProgressCallback(progress, steps)
|
| 271 |
|
| 272 |
+
# NEUE: 512x512 für Realistic Vision
|
| 273 |
image = pipe(
|
| 274 |
prompt=prompt,
|
| 275 |
+
height=512, # ← 512 statt IMG_SIZE (1024)
|
| 276 |
+
width=512, # ← 512 statt IMG_SIZE (1024)
|
| 277 |
num_inference_steps=int(steps),
|
| 278 |
guidance_scale=guidance_scale,
|
| 279 |
generator=generator,
|
| 280 |
+
callback_on_step_end=callback,
|
| 281 |
+
callback_on_step_end_tensor_inputs=[],
|
| 282 |
).images[0]
|
| 283 |
|
| 284 |
end_time = time.time()
|
| 285 |
print(f"Bild generiert in {end_time - start_time:.2f} Sekunden")
|
| 286 |
|
|
|
|
| 287 |
return image
|
| 288 |
|
| 289 |
except Exception as e:
|
| 290 |
+
print(f"Fehler in text_to_image: {e}")
|
| 291 |
import traceback
|
| 292 |
traceback.print_exc()
|
| 293 |
return None
|
|
|
|
| 316 |
|
| 317 |
# CONTROLNET-STRENGTH ANPASSEN ABHÄNGIG VOM MODUS
|
| 318 |
if face_preserve:
|
| 319 |
+
controlnet_strength = adj_strength * 0.8
|
|
|
|
| 320 |
print(f"🎯 ControlNet Modus: Umgebung beibehalten (Strength = {controlnet_strength:.3f})")
|
| 321 |
else:
|
| 322 |
+
controlnet_strength = adj_strength * 0.5
|
|
|
|
| 323 |
print(f"🎯 ControlNet Modus: Person beibehalten (Strength = {controlnet_strength:.3f})")
|
| 324 |
|
| 325 |
controlnet_steps = min(25, int(steps * 0.8))
|
|
|
|
| 339 |
guidance_scale=guidance_scale,
|
| 340 |
controlnet_strength=controlnet_strength,
|
| 341 |
progress=progress,
|
| 342 |
+
keep_environment=face_preserve
|
| 343 |
)
|
| 344 |
|
| 345 |
print(f"✅ ControlNet Output erhalten: {type(controlnet_output)}")
|
|
|
|
| 350 |
# -------------------------------
|
| 351 |
progress(0.3, desc="ControlNet abgeschlossen – starte Inpaint...")
|
| 352 |
|
| 353 |
+
pipe = load_img2img()
|
| 354 |
|
| 355 |
+
img_resized = inpaint_input.convert("RGB").resize((512, 512))
|
|
|
|
| 356 |
|
| 357 |
adj_guidance = min(guidance_scale, 12.0)
|
| 358 |
seed = random.randint(0, 2**32 - 1)
|
|
|
|
| 365 |
mask = None
|
| 366 |
if bbox_x1 and bbox_y1 and bbox_x2 and bbox_y2:
|
| 367 |
orig_w, orig_h = image.size
|
| 368 |
+
scale_x, scale_y = 512 / orig_w, 512 / orig_h
|
| 369 |
bbox_coords = [
|
| 370 |
int(bbox_x1 * scale_x),
|
| 371 |
int(bbox_y1 * scale_y),
|
|
|
|
| 412 |
import traceback
|
| 413 |
traceback.print_exc()
|
| 414 |
return None
|
|
|
|
| 415 |
|
| 416 |
def update_bbox_from_image(image):
|
| 417 |
"""Aktualisiert die Bounding-Box-Koordinaten wenn ein Bild hochgeladen wird"""
|
|
|
|
| 493 |
"""
|
| 494 |
) as demo:
|
| 495 |
|
|
|
|
| 496 |
with gr.Column(visible=True) as content_area:
|
| 497 |
with gr.Tab("Text zu Bild"):
|
| 498 |
gr.Markdown("**Beschreibe dein gewünschtes Bild:**")
|
|
|
|
| 536 |
with gr.Tab("Bild zu Bild"):
|
| 537 |
gr.Markdown("**Lade ein Bild hoch und beschreibe die gewünschte Veränderung:**")
|
| 538 |
|
|
|
|
| 539 |
with gr.Row():
|
| 540 |
with gr.Column():
|
| 541 |
img_input = gr.Image(
|
|
|
|
| 553 |
show_download_button=False
|
| 554 |
)
|
| 555 |
|
|
|
|
| 556 |
with gr.Row():
|
| 557 |
face_preserve = gr.Checkbox(
|
| 558 |
label="Schutz",
|
|
|
|
| 560 |
info="🟢 Checkbox AN: Alles AUSSERHALB des gelben Rahmens verändern | 🔴 Checkbox AUS: Nur INNERHALB des gelben Rahmens verändern"
|
| 561 |
)
|
| 562 |
|
|
|
|
| 563 |
with gr.Row():
|
| 564 |
gr.Markdown("**Bildelementbereich anpassen**")
|
| 565 |
|
|
|
|
| 590 |
info="Untere Kante des Bildelementbereichs"
|
| 591 |
)
|
| 592 |
|
|
|
|
| 593 |
with gr.Row():
|
| 594 |
with gr.Column():
|
| 595 |
img_prompt = gr.Textbox(
|
|
|
|
| 606 |
info="Was soll vermieden werden? Unerwünschte Elemente auflisten."
|
| 607 |
)
|
| 608 |
|
|
|
|
| 609 |
with gr.Row():
|
| 610 |
with gr.Column():
|
| 611 |
strength_slider = gr.Slider(
|
|
|
|
| 635 |
"• **Koordinaten nur bei erkennbaren Verzerrungen anpassen** (Bereiche leicht verschieben)"
|
| 636 |
)
|
| 637 |
|
|
|
|
| 638 |
transform_btn = gr.Button("Bild transformieren", variant="primary")
|
| 639 |
|
| 640 |
with gr.Row():
|
|
|
|
| 644 |
type="pil"
|
| 645 |
)
|
| 646 |
|
|
|
|
|
|
|
| 647 |
img_input.change(
|
| 648 |
fn=process_image_upload,
|
| 649 |
inputs=[img_input],
|
| 650 |
outputs=[preview_output, bbox_x1, bbox_y1, bbox_x2, bbox_y2]
|
| 651 |
)
|
| 652 |
|
|
|
|
| 653 |
coordinate_inputs = [img_input, bbox_x1, bbox_y1, bbox_x2, bbox_y2, face_preserve]
|
| 654 |
|
| 655 |
bbox_x1.change(
|
|
|
|
| 676 |
outputs=preview_output
|
| 677 |
)
|
| 678 |
|
|
|
|
| 679 |
face_preserve.change(
|
| 680 |
fn=update_live_preview,
|
| 681 |
inputs=coordinate_inputs,
|
| 682 |
outputs=preview_output
|
| 683 |
)
|
| 684 |
|
|
|
|
| 685 |
transform_btn.click(
|
| 686 |
fn=img_to_image,
|
| 687 |
inputs=[
|
|
|
|
| 698 |
|
| 699 |
if __name__ == "__main__":
|
| 700 |
demo = main_ui()
|
| 701 |
+
# OPTIMIERTE WARTESCHLANGE FÜR T4
|
| 702 |
+
demo.queue(
|
| 703 |
+
max_size=3, # Max 3 Anfragen in Warteschlange
|
| 704 |
+
concurrency_count=1 # Nur 1 Generation gleichzeitig
|
| 705 |
+
)
|
| 706 |
demo.launch(
|
| 707 |
server_name="0.0.0.0",
|
| 708 |
server_port=7860,
|