Update app.py
Browse files
app.py
CHANGED
|
@@ -29,7 +29,9 @@ from diffusers import (
|
|
| 29 |
StableDiffusionControlNetPipeline,
|
| 30 |
ControlNetModel,
|
| 31 |
StableDiffusionPipeline,
|
| 32 |
-
StableDiffusionXLPipeline
|
|
|
|
|
|
|
| 33 |
)
|
| 34 |
from diffusers import UniPCMultistepScheduler, DPMSolverMultistepScheduler, EulerAncestralDiscreteScheduler
|
| 35 |
from controlnet_aux import (
|
|
@@ -60,11 +62,14 @@ CURRENT_CONTROLNET_KEY = None
|
|
| 60 |
CURRENT_T2I_PIPE = None
|
| 61 |
CURRENT_T2I_MODEL = None
|
| 62 |
CURRENT_SDXL_REFINER = None
|
|
|
|
|
|
|
| 63 |
|
| 64 |
# Enhanced SDXL Models (including NSFW-capable)
|
| 65 |
SDXL_MODELS = [
|
| 66 |
"stabilityai/stable-diffusion-xl-base-1.0",
|
| 67 |
"stabilityai/stable-diffusion-xl-refiner-1.0",
|
|
|
|
| 68 |
"Laxhar/noobai-XL-1.1",
|
| 69 |
"RunDiffusion/Juggernaut-XL-v9",
|
| 70 |
"dataautogpt3/ProteusV0.4",
|
|
@@ -76,6 +81,12 @@ SDXL_MODELS = [
|
|
| 76 |
"digiplay/Pony_Diffusion_V6_XL"
|
| 77 |
]
|
| 78 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 79 |
# Enhanced SD1.5 Models (including NSFW-capable)
|
| 80 |
SD15_MODELS = [
|
| 81 |
# Original models
|
|
@@ -117,6 +128,11 @@ SD15_MODELS = [
|
|
| 117 |
"Fictiverse/Stable_Diffusion_VoxelArt_Model"
|
| 118 |
]
|
| 119 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 120 |
# Chinese Models
|
| 121 |
CHINESE_MODELS = [
|
| 122 |
"AI-Chen/Chinese-Stable-Diffusion",
|
|
@@ -144,7 +160,8 @@ CONTROLNET_MODELS_SD15 = {
|
|
| 144 |
CONTROLNET_MODELS_SDXL = {
|
| 145 |
"canny_sdxl": "diffusers/controlnet-canny-sdxl-1.0",
|
| 146 |
"depth_sdxl": "diffusers/controlnet-depth-sdxl-1.0",
|
| 147 |
-
"openpose_sdxl": "thibaud/controlnet-openpose-sdxl-1.0"
|
|
|
|
| 148 |
}
|
| 149 |
|
| 150 |
# แก้ไข LoRA models ให้ใช้ชื่อที่ไม่มีช่องว่าง
|
|
@@ -189,7 +206,9 @@ LORA_MODELS = {
|
|
| 189 |
"realistic-nsfw": "digiplay/RealisticNSFW_v1",
|
| 190 |
"anime-nsfw": "Linaqruf/anime-nsfw-lora",
|
| 191 |
"hentai-diffusion": "Deltaadams/Hentai-Diffusion",
|
| 192 |
-
"sexy-pose": "alvdansen/sexy-pose-lora"
|
|
|
|
|
|
|
| 193 |
}
|
| 194 |
|
| 195 |
# VAE models for better quality
|
|
@@ -197,7 +216,8 @@ VAE_MODELS = {
|
|
| 197 |
"None": None,
|
| 198 |
"SD1.5 VAE": "stabilityai/sd-vae-ft-mse",
|
| 199 |
"Anime VAE": "hakurei/waifu-diffusion-v1-4",
|
| 200 |
-
"SDXL VAE": "madebyollin/sdxl-vae-fp16-fix"
|
|
|
|
| 201 |
}
|
| 202 |
|
| 203 |
# Detector instances
|
|
@@ -207,6 +227,10 @@ def is_sdxl_model(model_name: str) -> bool:
|
|
| 207 |
"""Check if model is SDXL"""
|
| 208 |
return model_name in SDXL_MODELS or "xl" in model_name.lower() or "XL" in model_name
|
| 209 |
|
|
|
|
|
|
|
|
|
|
|
|
|
| 210 |
def load_detector(detector_type: str):
|
| 211 |
"""Lazy load detector"""
|
| 212 |
global DETECTORS
|
|
@@ -250,14 +274,19 @@ def load_detector(detector_type: str):
|
|
| 250 |
def get_controlnet_model(controlnet_type: str, is_sdxl: bool = False):
|
| 251 |
"""Get ControlNet model based on type"""
|
| 252 |
if is_sdxl:
|
|
|
|
|
|
|
| 253 |
return CONTROLNET_MODELS_SDXL[controlnet_type]
|
| 254 |
else:
|
|
|
|
|
|
|
| 255 |
return CONTROLNET_MODELS_SD15[controlnet_type]
|
| 256 |
|
| 257 |
def prepare_condition_image(image, controlnet_type, is_sdxl=False):
|
| 258 |
"""Prepare condition image for ControlNet"""
|
| 259 |
-
if
|
| 260 |
-
|
|
|
|
| 261 |
if detector:
|
| 262 |
result = detector(image, detect_resolution=512 if not is_sdxl else 1024, image_resolution=512 if not is_sdxl else 1024)
|
| 263 |
return Image.fromarray(result) if isinstance(result, np.ndarray) else result
|
|
@@ -312,11 +341,7 @@ def get_pipeline(model_name: str, controlnet_type: str = "lineart", lora_model:
|
|
| 312 |
|
| 313 |
try:
|
| 314 |
is_sdxl = is_sdxl_model(model_name)
|
| 315 |
-
|
| 316 |
-
if is_sdxl and controlnet_type not in CONTROLNET_MODELS_SDXL:
|
| 317 |
-
raise ValueError(f"SDXL model only supports: {list(CONTROLNET_MODELS_SDXL.keys())}")
|
| 318 |
-
elif not is_sdxl and controlnet_type not in CONTROLNET_MODELS_SD15:
|
| 319 |
-
raise ValueError(f"SD1.5 model only supports: {list(CONTROLNET_MODELS_SD15.keys())}")
|
| 320 |
|
| 321 |
controlnet_model_name = get_controlnet_model(controlnet_type, is_sdxl)
|
| 322 |
controlnet = ControlNetModel.from_pretrained(
|
|
@@ -325,15 +350,27 @@ def get_pipeline(model_name: str, controlnet_type: str = "lineart", lora_model:
|
|
| 325 |
).to(device)
|
| 326 |
|
| 327 |
if is_sdxl:
|
| 328 |
-
|
| 329 |
-
|
| 330 |
-
|
| 331 |
-
|
| 332 |
-
|
| 333 |
-
|
| 334 |
-
|
| 335 |
-
|
| 336 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 337 |
else:
|
| 338 |
pipe = StableDiffusionControlNetPipeline.from_pretrained(
|
| 339 |
model_name,
|
|
@@ -388,14 +425,23 @@ def get_pipeline(model_name: str, controlnet_type: str = "lineart", lora_model:
|
|
| 388 |
pass
|
| 389 |
pipe.enable_model_cpu_offload()
|
| 390 |
|
| 391 |
-
|
| 392 |
-
|
| 393 |
-
|
| 394 |
-
except:
|
| 395 |
try:
|
| 396 |
-
pipe.scheduler =
|
|
|
|
| 397 |
except:
|
| 398 |
pass
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 399 |
|
| 400 |
CURRENT_CONTROLNET_PIPE = pipe
|
| 401 |
CURRENT_CONTROLNET_KEY = key
|
|
@@ -410,29 +456,55 @@ def get_pipeline(model_name: str, controlnet_type: str = "lineart", lora_model:
|
|
| 410 |
def load_t2i_model(model_name: str, lora_model: str = None, lora_weight: float = 0.8,
|
| 411 |
vae_model: str = None):
|
| 412 |
"""Load text-to-image model with optional LoRA and VAE"""
|
| 413 |
-
global CURRENT_T2I_PIPE, CURRENT_T2I_MODEL, CURRENT_SDXL_REFINER
|
| 414 |
|
| 415 |
-
|
| 416 |
-
|
| 417 |
|
| 418 |
-
|
| 419 |
-
return
|
| 420 |
|
| 421 |
-
if
|
| 422 |
-
|
| 423 |
-
|
| 424 |
-
|
| 425 |
-
if
|
| 426 |
-
|
| 427 |
-
|
| 428 |
-
|
| 429 |
-
if
|
| 430 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 431 |
|
| 432 |
print(f"📥 Loading T2I model: {model_name}")
|
| 433 |
|
| 434 |
try:
|
| 435 |
-
if
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 436 |
if use_refiner:
|
| 437 |
CURRENT_T2I_PIPE = StableDiffusionXLPipeline.from_pretrained(
|
| 438 |
"stabilityai/stable-diffusion-xl-base-1.0",
|
|
@@ -462,6 +534,9 @@ def load_t2i_model(model_name: str, lora_model: str = None, lora_weight: float =
|
|
| 462 |
use_safetensors=True,
|
| 463 |
variant="fp16" if dtype == torch.float16 else None
|
| 464 |
).to(device)
|
|
|
|
|
|
|
|
|
|
| 465 |
else:
|
| 466 |
CURRENT_T2I_PIPE = StableDiffusionPipeline.from_pretrained(
|
| 467 |
model_name,
|
|
@@ -471,6 +546,9 @@ def load_t2i_model(model_name: str, lora_model: str = None, lora_weight: float =
|
|
| 471 |
use_safetensors=True,
|
| 472 |
variant="fp16" if dtype == torch.float16 else None
|
| 473 |
).to(device)
|
|
|
|
|
|
|
|
|
|
| 474 |
|
| 475 |
# Load custom VAE
|
| 476 |
if vae_model and vae_model != "None":
|
|
@@ -478,7 +556,7 @@ def load_t2i_model(model_name: str, lora_model: str = None, lora_weight: float =
|
|
| 478 |
from diffusers import AutoencoderKL
|
| 479 |
print(f"🔄 Loading custom VAE: {vae_model}")
|
| 480 |
vae = AutoencoderKL.from_pretrained(vae_model, torch_dtype=dtype).to(device)
|
| 481 |
-
|
| 482 |
except Exception as e:
|
| 483 |
print(f"⚠️ Error loading VAE: {e}")
|
| 484 |
|
|
@@ -488,43 +566,53 @@ def load_t2i_model(model_name: str, lora_model: str = None, lora_weight: float =
|
|
| 488 |
try:
|
| 489 |
if lora_model in LORA_MODELS:
|
| 490 |
lora_path = LORA_MODELS[lora_model]
|
| 491 |
-
|
| 492 |
-
|
| 493 |
else:
|
| 494 |
-
|
| 495 |
-
|
| 496 |
except Exception as e:
|
| 497 |
print(f"⚠️ Error loading LoRA: {e}")
|
| 498 |
|
| 499 |
# Optimizations
|
| 500 |
-
|
| 501 |
|
| 502 |
-
if hasattr(
|
| 503 |
-
|
| 504 |
else:
|
| 505 |
try:
|
| 506 |
-
|
| 507 |
except:
|
| 508 |
pass
|
| 509 |
|
| 510 |
if device.type == "cuda":
|
| 511 |
try:
|
| 512 |
-
|
| 513 |
except:
|
| 514 |
pass
|
| 515 |
-
|
| 516 |
-
|
| 517 |
-
try:
|
| 518 |
-
CURRENT_T2I_PIPE.scheduler = EulerAncestralDiscreteScheduler.from_config(CURRENT_T2I_PIPE.scheduler.config)
|
| 519 |
-
except:
|
| 520 |
-
pass
|
| 521 |
|
| 522 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 523 |
|
| 524 |
except Exception as e:
|
| 525 |
print(f"❌ Error loading T2I model: {e}")
|
| 526 |
-
|
| 527 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
| 528 |
raise
|
| 529 |
|
| 530 |
def colorize_sd15(sketch, base_model, controlnet_type, lora_model, lora_weight, vae_model,
|
|
@@ -617,6 +705,47 @@ def colorize_sdxl(sketch, base_model, controlnet_type, lora_model, lora_weight,
|
|
| 617 |
error_img = Image.new('RGB', (1024, 1024), color='red')
|
| 618 |
return error_img, Image.new('RGB', (1024, 1024), color='gray')
|
| 619 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 620 |
def t2i_sd15(prompt, negative_prompt, model, lora_model, lora_weight, vae_model,
|
| 621 |
seed, steps, scale, w, h):
|
| 622 |
"""Text-to-image for SD1.5 models"""
|
|
@@ -669,7 +798,7 @@ def t2i_sdxl(prompt, negative_prompt, model, lora_model, lora_weight, vae_model,
|
|
| 669 |
return error_img
|
| 670 |
|
| 671 |
model_to_load = model
|
| 672 |
-
if use_refiner and "refiner" not in model.lower():
|
| 673 |
model_to_load = "stabilityai/stable-diffusion-xl-refiner-1.0"
|
| 674 |
|
| 675 |
load_t2i_model(model_to_load, lora_model, lora_weight, vae_model)
|
|
@@ -683,7 +812,7 @@ def t2i_sdxl(prompt, negative_prompt, model, lora_model, lora_weight, vae_model,
|
|
| 683 |
gen = torch.Generator(device=device).manual_seed(int(seed))
|
| 684 |
|
| 685 |
with torch.inference_mode():
|
| 686 |
-
if use_refiner and CURRENT_SDXL_REFINER is not None:
|
| 687 |
image = CURRENT_T2I_PIPE(
|
| 688 |
prompt=prompt,
|
| 689 |
negative_prompt=negative_prompt,
|
|
@@ -704,12 +833,84 @@ def t2i_sdxl(prompt, negative_prompt, model, lora_model, lora_weight, vae_model,
|
|
| 704 |
generator=gen
|
| 705 |
).images[0]
|
| 706 |
else:
|
| 707 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 708 |
prompt,
|
| 709 |
negative_prompt=negative_prompt,
|
| 710 |
width=int(w),
|
| 711 |
height=int(h),
|
| 712 |
-
num_inference_steps=
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 713 |
guidance_scale=float(scale),
|
| 714 |
generator=gen
|
| 715 |
).images[0]
|
|
@@ -719,7 +920,7 @@ def t2i_sdxl(prompt, negative_prompt, model, lora_model, lora_weight, vae_model,
|
|
| 719 |
|
| 720 |
return result
|
| 721 |
except Exception as e:
|
| 722 |
-
print(f"❌ Error in
|
| 723 |
error_img = Image.new('RGB', (int(w), int(h)), color='red')
|
| 724 |
from PIL import ImageDraw, ImageFont
|
| 725 |
draw = ImageDraw.Draw(error_img)
|
|
@@ -734,6 +935,7 @@ def unload_all_models():
|
|
| 734 |
global CURRENT_CONTROLNET_PIPE, CURRENT_CONTROLNET_KEY
|
| 735 |
global DETECTORS
|
| 736 |
global CURRENT_T2I_PIPE, CURRENT_T2I_MODEL, CURRENT_SDXL_REFINER
|
|
|
|
| 737 |
|
| 738 |
print("🗑️ Unloading all models from memory...")
|
| 739 |
|
|
@@ -766,7 +968,15 @@ def unload_all_models():
|
|
| 766 |
except:
|
| 767 |
pass
|
| 768 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 769 |
CURRENT_T2I_MODEL = None
|
|
|
|
| 770 |
|
| 771 |
gc.collect()
|
| 772 |
if torch.cuda.is_available():
|
|
@@ -867,7 +1077,7 @@ with gr.Blocks(title="🎨 AI Image Generator Pro", theme=gr.themes.Soft()) as d
|
|
| 867 |
with gr.Tab("🎨 SDXL ControlNet"):
|
| 868 |
gr.Markdown("""
|
| 869 |
### Transform sketches/images using SDXL with ControlNet
|
| 870 |
-
- **Supports:** canny_sdxl, depth_sdxl, openpose_sdxl
|
| 871 |
- **Best Resolution:** 1024x1024
|
| 872 |
- **Higher quality, more VRAM required**
|
| 873 |
""")
|
|
@@ -884,7 +1094,7 @@ with gr.Blocks(title="🎨 AI Image Generator Pro", theme=gr.themes.Soft()) as d
|
|
| 884 |
)
|
| 885 |
controlnet_type_sdxl = gr.Dropdown(
|
| 886 |
choices=list(CONTROLNET_MODELS_SDXL.keys()),
|
| 887 |
-
value="
|
| 888 |
label="ControlNet Type"
|
| 889 |
)
|
| 890 |
|
|
@@ -898,7 +1108,7 @@ with gr.Blocks(title="🎨 AI Image Generator Pro", theme=gr.themes.Soft()) as d
|
|
| 898 |
lora_weight_sdxl = gr.Slider(0.1, 2.0, 0.8, step=0.1, label="LoRA Weight")
|
| 899 |
|
| 900 |
vae_model_sdxl = gr.Dropdown(
|
| 901 |
-
choices=["None", "SDXL VAE"],
|
| 902 |
value="None",
|
| 903 |
label="VAE Model (Optional)"
|
| 904 |
)
|
|
@@ -934,6 +1144,73 @@ with gr.Blocks(title="🎨 AI Image Generator Pro", theme=gr.themes.Soft()) as d
|
|
| 934 |
[out_sdxl, condition_out_sdxl]
|
| 935 |
)
|
| 936 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 937 |
with gr.Tab("🖼️ SD1.5 Text-to-Image"):
|
| 938 |
gr.Markdown("""
|
| 939 |
### Generate images from text descriptions using SD1.5
|
|
@@ -999,6 +1276,81 @@ with gr.Blocks(title="🎨 AI Image Generator Pro", theme=gr.themes.Soft()) as d
|
|
| 999 |
t2i_out_sd15
|
| 1000 |
)
|
| 1001 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1002 |
with gr.Tab("🖼️ SDXL Text-to-Image"):
|
| 1003 |
gr.Markdown("""
|
| 1004 |
### Generate images from text descriptions using SDXL
|
|
@@ -1010,7 +1362,7 @@ with gr.Blocks(title="🎨 AI Image Generator Pro", theme=gr.themes.Soft()) as d
|
|
| 1010 |
with gr.Column(scale=1):
|
| 1011 |
gr.Markdown("### Model Configuration")
|
| 1012 |
t2i_model_sdxl = gr.Dropdown(
|
| 1013 |
-
choices=SDXL_MODELS,
|
| 1014 |
value="stabilityai/stable-diffusion-xl-base-1.0",
|
| 1015 |
label="SDXL Base Model"
|
| 1016 |
)
|
|
@@ -1073,104 +1425,107 @@ with gr.Blocks(title="🎨 AI Image Generator Pro", theme=gr.themes.Soft()) as d
|
|
| 1073 |
gr.Markdown("""
|
| 1074 |
# Model & Feature Guide
|
| 1075 |
|
| 1076 |
-
## 🎯
|
| 1077 |
|
| 1078 |
### SD1.5 (Stable Diffusion 1.5)
|
| 1079 |
-
- **Pros:**
|
| 1080 |
-
- **Cons:**
|
| 1081 |
-
- **Best for:** Quick
|
| 1082 |
-
- **Resolution:** 512x512 optimal
|
| 1083 |
|
| 1084 |
### SDXL (Stable Diffusion XL)
|
| 1085 |
-
- **Pros:**
|
| 1086 |
-
- **Cons:**
|
| 1087 |
-
- **Best for:** Final quality
|
| 1088 |
-
|
| 1089 |
-
|
| 1090 |
-
|
| 1091 |
-
|
| 1092 |
-
|
| 1093 |
-
|
| 1094 |
-
-
|
| 1095 |
-
- **
|
| 1096 |
-
- **
|
| 1097 |
-
- **
|
| 1098 |
-
|
| 1099 |
-
|
| 1100 |
-
|
| 1101 |
-
|
| 1102 |
-
- **
|
| 1103 |
-
- **
|
| 1104 |
-
|
| 1105 |
-
|
| 1106 |
-
|
| 1107 |
-
|
| 1108 |
-
|
| 1109 |
-
|
| 1110 |
-
|
| 1111 |
-
|
| 1112 |
-
|
| 1113 |
-
|
| 1114 |
-
|
| 1115 |
-
-
|
| 1116 |
-
-
|
| 1117 |
-
|
| 1118 |
-
|
| 1119 |
-
|
| 1120 |
-
-
|
| 1121 |
-
|
| 1122 |
-
|
| 1123 |
-
|
| 1124 |
-
|
| 1125 |
-
|
| 1126 |
-
|
| 1127 |
-
-
|
| 1128 |
-
-
|
| 1129 |
-
-
|
| 1130 |
-
- Use
|
| 1131 |
-
|
| 1132 |
-
|
| 1133 |
-
|
| 1134 |
-
###
|
| 1135 |
-
|
| 1136 |
-
|
| 1137 |
-
|
| 1138 |
-
|
| 1139 |
-
|
| 1140 |
-
|
| 1141 |
-
|
| 1142 |
-
|
| 1143 |
-
|
| 1144 |
-
|
| 1145 |
-
|
| 1146 |
-
|
| 1147 |
-
|
| 1148 |
-
|
| 1149 |
-
|
| 1150 |
-
|
| 1151 |
-
- **12+:** Very strict
|
| 1152 |
-
|
| 1153 |
-
### Resolution
|
| 1154 |
-
- **SD1.5:** 512x512, 512x768, 768x512 (max 1024x1024)
|
| 1155 |
-
- **SDXL:** 1024x1024, 1024x1536, 1536x1024 (max 2048x2048)
|
| 1156 |
-
|
| 1157 |
-
## 🚀 Performance Tips
|
| 1158 |
-
|
| 1159 |
-
### For Low VRAM (<8GB)
|
| 1160 |
- Use SD1.5 models only
|
|
|
|
| 1161 |
- Enable attention slicing
|
| 1162 |
-
- Use lower resolutions (512x512)
|
| 1163 |
-
- Reduce steps (20-30)
|
| 1164 |
|
| 1165 |
-
###
|
| 1166 |
-
-
|
| 1167 |
-
-
|
| 1168 |
- Enable xFormers
|
| 1169 |
|
| 1170 |
-
###
|
| 1171 |
-
-
|
| 1172 |
-
- SDXL with refiner
|
| 1173 |
- Higher resolutions
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1174 |
""")
|
| 1175 |
|
| 1176 |
try:
|
|
|
|
| 29 |
StableDiffusionControlNetPipeline,
|
| 30 |
ControlNetModel,
|
| 31 |
StableDiffusionPipeline,
|
| 32 |
+
StableDiffusionXLPipeline,
|
| 33 |
+
StableDiffusionXLControlNetPipeline,
|
| 34 |
+
AutoPipelineForText2Image
|
| 35 |
)
|
| 36 |
from diffusers import UniPCMultistepScheduler, DPMSolverMultistepScheduler, EulerAncestralDiscreteScheduler
|
| 37 |
from controlnet_aux import (
|
|
|
|
| 62 |
CURRENT_T2I_PIPE = None
|
| 63 |
CURRENT_T2I_MODEL = None
|
| 64 |
CURRENT_SDXL_REFINER = None
|
| 65 |
+
CURRENT_TURBO_PIPE = None
|
| 66 |
+
CURRENT_TURBO_MODEL = None
|
| 67 |
|
| 68 |
# Enhanced SDXL Models (including NSFW-capable)
|
| 69 |
SDXL_MODELS = [
|
| 70 |
"stabilityai/stable-diffusion-xl-base-1.0",
|
| 71 |
"stabilityai/stable-diffusion-xl-refiner-1.0",
|
| 72 |
+
"stabilityai/sdxl-turbo", # เพิ่ม SDXL-Turbo
|
| 73 |
"Laxhar/noobai-XL-1.1",
|
| 74 |
"RunDiffusion/Juggernaut-XL-v9",
|
| 75 |
"dataautogpt3/ProteusV0.4",
|
|
|
|
| 81 |
"digiplay/Pony_Diffusion_V6_XL"
|
| 82 |
]
|
| 83 |
|
| 84 |
+
# Turbo models (fast generation)
|
| 85 |
+
TURBO_MODELS = [
|
| 86 |
+
"stabilityai/sdxl-turbo",
|
| 87 |
+
"stabilityai/sd-turbo"
|
| 88 |
+
]
|
| 89 |
+
|
| 90 |
# Enhanced SD1.5 Models (including NSFW-capable)
|
| 91 |
SD15_MODELS = [
|
| 92 |
# Original models
|
|
|
|
| 128 |
"Fictiverse/Stable_Diffusion_VoxelArt_Model"
|
| 129 |
]
|
| 130 |
|
| 131 |
+
# Specialized models
|
| 132 |
+
SPECIAL_MODELS = {
|
| 133 |
+
"waifu_colorize": "ShinoharaHare/Waifu-Colorize-XL" # โมเดลสำหรับ colorize โดยเฉพาะ
|
| 134 |
+
}
|
| 135 |
+
|
| 136 |
# Chinese Models
|
| 137 |
CHINESE_MODELS = [
|
| 138 |
"AI-Chen/Chinese-Stable-Diffusion",
|
|
|
|
| 160 |
CONTROLNET_MODELS_SDXL = {
|
| 161 |
"canny_sdxl": "diffusers/controlnet-canny-sdxl-1.0",
|
| 162 |
"depth_sdxl": "diffusers/controlnet-depth-sdxl-1.0",
|
| 163 |
+
"openpose_sdxl": "thibaud/controlnet-openpose-sdxl-1.0",
|
| 164 |
+
"lineart_sdxl": "ShermanG/ControlNet-Standard-Lineart-for-SDXL" # เพิ่ม ControlNet Lineart สำหรับ SDXL
|
| 165 |
}
|
| 166 |
|
| 167 |
# แก้ไข LoRA models ให้ใช้ชื่อที่ไม่มีช่องว่าง
|
|
|
|
| 206 |
"realistic-nsfw": "digiplay/RealisticNSFW_v1",
|
| 207 |
"anime-nsfw": "Linaqruf/anime-nsfw-lora",
|
| 208 |
"hentai-diffusion": "Deltaadams/Hentai-Diffusion",
|
| 209 |
+
"sexy-pose": "alvdansen/sexy-pose-lora",
|
| 210 |
+
# Colorize LoRAs
|
| 211 |
+
"colorize-xl": "ShinoharaHare/Waifu-Colorize-XL"
|
| 212 |
}
|
| 213 |
|
| 214 |
# VAE models for better quality
|
|
|
|
| 216 |
"None": None,
|
| 217 |
"SD1.5 VAE": "stabilityai/sd-vae-ft-mse",
|
| 218 |
"Anime VAE": "hakurei/waifu-diffusion-v1-4",
|
| 219 |
+
"SDXL VAE": "madebyollin/sdxl-vae-fp16-fix",
|
| 220 |
+
"Turbo VAE": "madebyollin/sdxl-vae-fp16-fix"
|
| 221 |
}
|
| 222 |
|
| 223 |
# Detector instances
|
|
|
|
| 227 |
"""Check if model is SDXL"""
|
| 228 |
return model_name in SDXL_MODELS or "xl" in model_name.lower() or "XL" in model_name
|
| 229 |
|
| 230 |
+
def is_turbo_model(model_name: str) -> bool:
|
| 231 |
+
"""Check if model is Turbo"""
|
| 232 |
+
return model_name in TURBO_MODELS or "turbo" in model_name.lower()
|
| 233 |
+
|
| 234 |
def load_detector(detector_type: str):
|
| 235 |
"""Lazy load detector"""
|
| 236 |
global DETECTORS
|
|
|
|
| 274 |
def get_controlnet_model(controlnet_type: str, is_sdxl: bool = False):
|
| 275 |
"""Get ControlNet model based on type"""
|
| 276 |
if is_sdxl:
|
| 277 |
+
if controlnet_type not in CONTROLNET_MODELS_SDXL:
|
| 278 |
+
raise ValueError(f"SDXL ControlNet type must be one of: {list(CONTROLNET_MODELS_SDXL.keys())}")
|
| 279 |
return CONTROLNET_MODELS_SDXL[controlnet_type]
|
| 280 |
else:
|
| 281 |
+
if controlnet_type not in CONTROLNET_MODELS_SD15:
|
| 282 |
+
raise ValueError(f"SD1.5 ControlNet type must be one of: {list(CONTROLNET_MODELS_SD15.keys())}")
|
| 283 |
return CONTROLNET_MODELS_SD15[controlnet_type]
|
| 284 |
|
| 285 |
def prepare_condition_image(image, controlnet_type, is_sdxl=False):
|
| 286 |
"""Prepare condition image for ControlNet"""
|
| 287 |
+
if "lineart" in controlnet_type:
|
| 288 |
+
detector_type = "lineart_anime" if "anime" in controlnet_type else "lineart"
|
| 289 |
+
detector = load_detector(detector_type)
|
| 290 |
if detector:
|
| 291 |
result = detector(image, detect_resolution=512 if not is_sdxl else 1024, image_resolution=512 if not is_sdxl else 1024)
|
| 292 |
return Image.fromarray(result) if isinstance(result, np.ndarray) else result
|
|
|
|
| 341 |
|
| 342 |
try:
|
| 343 |
is_sdxl = is_sdxl_model(model_name)
|
| 344 |
+
is_turbo = is_turbo_model(model_name)
|
|
|
|
|
|
|
|
|
|
|
|
|
| 345 |
|
| 346 |
controlnet_model_name = get_controlnet_model(controlnet_type, is_sdxl)
|
| 347 |
controlnet = ControlNetModel.from_pretrained(
|
|
|
|
| 350 |
).to(device)
|
| 351 |
|
| 352 |
if is_sdxl:
|
| 353 |
+
if is_turbo:
|
| 354 |
+
# สำหรับ Turbo models ใช้ pipeline ที่เหมาะสม
|
| 355 |
+
pipe = StableDiffusionXLControlNetPipeline.from_pretrained(
|
| 356 |
+
model_name,
|
| 357 |
+
controlnet=controlnet,
|
| 358 |
+
torch_dtype=dtype,
|
| 359 |
+
safety_checker=None,
|
| 360 |
+
requires_safety_checker=False,
|
| 361 |
+
use_safetensors=True,
|
| 362 |
+
variant="fp16" if dtype == torch.float16 else None
|
| 363 |
+
).to(device)
|
| 364 |
+
else:
|
| 365 |
+
pipe = StableDiffusionXLControlNetPipeline.from_pretrained(
|
| 366 |
+
model_name,
|
| 367 |
+
controlnet=controlnet,
|
| 368 |
+
torch_dtype=dtype,
|
| 369 |
+
safety_checker=None,
|
| 370 |
+
requires_safety_checker=False,
|
| 371 |
+
use_safetensors=True,
|
| 372 |
+
variant="fp16" if dtype == torch.float16 else None
|
| 373 |
+
).to(device)
|
| 374 |
else:
|
| 375 |
pipe = StableDiffusionControlNetPipeline.from_pretrained(
|
| 376 |
model_name,
|
|
|
|
| 425 |
pass
|
| 426 |
pipe.enable_model_cpu_offload()
|
| 427 |
|
| 428 |
+
# ใช้ scheduler ที่เหมาะสมสำหรับแต่ละโมเดล
|
| 429 |
+
if is_turbo:
|
| 430 |
+
# สำหรับ Turbo models ใช้ scheduler ที่เร็วขึ้น
|
|
|
|
| 431 |
try:
|
| 432 |
+
pipe.scheduler = UniPCMultistepScheduler.from_config(pipe.scheduler.config)
|
| 433 |
+
print("✅ Using UniPC scheduler for Turbo")
|
| 434 |
except:
|
| 435 |
pass
|
| 436 |
+
else:
|
| 437 |
+
try:
|
| 438 |
+
pipe.scheduler = EulerAncestralDiscreteScheduler.from_config(pipe.scheduler.config)
|
| 439 |
+
print("✅ Using Euler Ancestral scheduler")
|
| 440 |
+
except:
|
| 441 |
+
try:
|
| 442 |
+
pipe.scheduler = DPMSolverMultistepScheduler.from_config(pipe.scheduler.config)
|
| 443 |
+
except:
|
| 444 |
+
pass
|
| 445 |
|
| 446 |
CURRENT_CONTROLNET_PIPE = pipe
|
| 447 |
CURRENT_CONTROLNET_KEY = key
|
|
|
|
| 456 |
def load_t2i_model(model_name: str, lora_model: str = None, lora_weight: float = 0.8,
|
| 457 |
vae_model: str = None):
|
| 458 |
"""Load text-to-image model with optional LoRA and VAE"""
|
| 459 |
+
global CURRENT_T2I_PIPE, CURRENT_T2I_MODEL, CURRENT_SDXL_REFINER, CURRENT_TURBO_PIPE, CURRENT_TURBO_MODEL
|
| 460 |
|
| 461 |
+
is_turbo = is_turbo_model(model_name)
|
| 462 |
+
use_refiner = "refiner" in model_name.lower() and not is_turbo
|
| 463 |
|
| 464 |
+
key = (model_name, lora_model, lora_weight, vae_model, use_refiner, is_turbo)
|
|
|
|
| 465 |
|
| 466 |
+
if is_turbo:
|
| 467 |
+
if CURRENT_TURBO_MODEL == key and CURRENT_TURBO_PIPE is not None:
|
| 468 |
+
return
|
| 469 |
+
else:
|
| 470 |
+
if CURRENT_T2I_MODEL == key and CURRENT_T2I_PIPE is not None:
|
| 471 |
+
return
|
| 472 |
+
|
| 473 |
+
if is_turbo:
|
| 474 |
+
if CURRENT_TURBO_PIPE is not None:
|
| 475 |
+
print(f"🗑️ Unloading old Turbo model: {CURRENT_TURBO_MODEL}")
|
| 476 |
+
del CURRENT_TURBO_PIPE
|
| 477 |
+
CURRENT_TURBO_PIPE = None
|
| 478 |
+
else:
|
| 479 |
+
if CURRENT_T2I_PIPE is not None:
|
| 480 |
+
print(f"🗑️ Unloading old T2I model: {CURRENT_T2I_MODEL}")
|
| 481 |
+
del CURRENT_T2I_PIPE
|
| 482 |
+
CURRENT_T2I_PIPE = None
|
| 483 |
+
if CURRENT_SDXL_REFINER is not None:
|
| 484 |
+
del CURRENT_SDXL_REFINER
|
| 485 |
+
CURRENT_SDXL_REFINER = None
|
| 486 |
+
|
| 487 |
+
gc.collect()
|
| 488 |
+
if torch.cuda.is_available():
|
| 489 |
+
torch.cuda.empty_cache()
|
| 490 |
|
| 491 |
print(f"📥 Loading T2I model: {model_name}")
|
| 492 |
|
| 493 |
try:
|
| 494 |
+
if is_turbo:
|
| 495 |
+
# โหลด Turbo model
|
| 496 |
+
CURRENT_TURBO_PIPE = AutoPipelineForText2Image.from_pretrained(
|
| 497 |
+
model_name,
|
| 498 |
+
torch_dtype=dtype,
|
| 499 |
+
safety_checker=None,
|
| 500 |
+
requires_safety_checker=False,
|
| 501 |
+
use_safetensors=True,
|
| 502 |
+
variant="fp16" if dtype == torch.float16 else None
|
| 503 |
+
).to(device)
|
| 504 |
+
|
| 505 |
+
CURRENT_TURBO_MODEL = key
|
| 506 |
+
pipe = CURRENT_TURBO_PIPE
|
| 507 |
+
elif is_sdxl_model(model_name):
|
| 508 |
if use_refiner:
|
| 509 |
CURRENT_T2I_PIPE = StableDiffusionXLPipeline.from_pretrained(
|
| 510 |
"stabilityai/stable-diffusion-xl-base-1.0",
|
|
|
|
| 534 |
use_safetensors=True,
|
| 535 |
variant="fp16" if dtype == torch.float16 else None
|
| 536 |
).to(device)
|
| 537 |
+
|
| 538 |
+
CURRENT_T2I_MODEL = key
|
| 539 |
+
pipe = CURRENT_T2I_PIPE
|
| 540 |
else:
|
| 541 |
CURRENT_T2I_PIPE = StableDiffusionPipeline.from_pretrained(
|
| 542 |
model_name,
|
|
|
|
| 546 |
use_safetensors=True,
|
| 547 |
variant="fp16" if dtype == torch.float16 else None
|
| 548 |
).to(device)
|
| 549 |
+
|
| 550 |
+
CURRENT_T2I_MODEL = key
|
| 551 |
+
pipe = CURRENT_T2I_PIPE
|
| 552 |
|
| 553 |
# Load custom VAE
|
| 554 |
if vae_model and vae_model != "None":
|
|
|
|
| 556 |
from diffusers import AutoencoderKL
|
| 557 |
print(f"🔄 Loading custom VAE: {vae_model}")
|
| 558 |
vae = AutoencoderKL.from_pretrained(vae_model, torch_dtype=dtype).to(device)
|
| 559 |
+
pipe.vae = vae
|
| 560 |
except Exception as e:
|
| 561 |
print(f"⚠️ Error loading VAE: {e}")
|
| 562 |
|
|
|
|
| 566 |
try:
|
| 567 |
if lora_model in LORA_MODELS:
|
| 568 |
lora_path = LORA_MODELS[lora_model]
|
| 569 |
+
pipe.load_lora_weights(lora_path)
|
| 570 |
+
pipe.fuse_lora(lora_scale=lora_weight)
|
| 571 |
else:
|
| 572 |
+
pipe.load_lora_weights(lora_model)
|
| 573 |
+
pipe.fuse_lora(lora_scale=lora_weight)
|
| 574 |
except Exception as e:
|
| 575 |
print(f"⚠️ Error loading LoRA: {e}")
|
| 576 |
|
| 577 |
# Optimizations
|
| 578 |
+
pipe.enable_attention_slicing(slice_size="max")
|
| 579 |
|
| 580 |
+
if hasattr(pipe, 'vae') and hasattr(pipe.vae, 'enable_slicing'):
|
| 581 |
+
pipe.vae.enable_slicing()
|
| 582 |
else:
|
| 583 |
try:
|
| 584 |
+
pipe.enable_vae_slicing()
|
| 585 |
except:
|
| 586 |
pass
|
| 587 |
|
| 588 |
if device.type == "cuda":
|
| 589 |
try:
|
| 590 |
+
pipe.enable_xformers_memory_efficient_attention()
|
| 591 |
except:
|
| 592 |
pass
|
| 593 |
+
pipe.enable_model_cpu_offload()
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 594 |
|
| 595 |
+
# ตั้งค่า scheduler
|
| 596 |
+
if is_turbo:
|
| 597 |
+
try:
|
| 598 |
+
pipe.scheduler = UniPCMultistepScheduler.from_config(pipe.scheduler.config)
|
| 599 |
+
print("✅ Using UniPC scheduler for Turbo")
|
| 600 |
+
except:
|
| 601 |
+
pass
|
| 602 |
+
else:
|
| 603 |
+
try:
|
| 604 |
+
pipe.scheduler = EulerAncestralDiscreteScheduler.from_config(pipe.scheduler.config)
|
| 605 |
+
except:
|
| 606 |
+
pass
|
| 607 |
|
| 608 |
except Exception as e:
|
| 609 |
print(f"❌ Error loading T2I model: {e}")
|
| 610 |
+
if is_turbo:
|
| 611 |
+
CURRENT_TURBO_PIPE = None
|
| 612 |
+
CURRENT_TURBO_MODEL = None
|
| 613 |
+
else:
|
| 614 |
+
CURRENT_T2I_PIPE = None
|
| 615 |
+
CURRENT_T2I_MODEL = None
|
| 616 |
raise
|
| 617 |
|
| 618 |
def colorize_sd15(sketch, base_model, controlnet_type, lora_model, lora_weight, vae_model,
|
|
|
|
| 705 |
error_img = Image.new('RGB', (1024, 1024), color='red')
|
| 706 |
return error_img, Image.new('RGB', (1024, 1024), color='gray')
|
| 707 |
|
| 708 |
+
def colorize_waifu_xl(sketch, lora_weight, vae_model, prompt, negative_prompt, seed, steps, scale, cn_weight):
|
| 709 |
+
"""Colorize function specifically for Waifu-Colorize-XL model"""
|
| 710 |
+
try:
|
| 711 |
+
model_name = "stabilityai/stable-diffusion-xl-base-1.0"
|
| 712 |
+
controlnet_type = "lineart_sdxl"
|
| 713 |
+
lora_model = "colorize-xl"
|
| 714 |
+
|
| 715 |
+
pipe = get_pipeline(model_name, controlnet_type, lora_model, lora_weight, vae_model)
|
| 716 |
+
|
| 717 |
+
status_msg = f"🎨 Using: Waifu-Colorize-XL (Lineart ControlNet)"
|
| 718 |
+
print(status_msg)
|
| 719 |
+
|
| 720 |
+
condition_img = prepare_condition_image(sketch, controlnet_type, is_sdxl=True)
|
| 721 |
+
|
| 722 |
+
gen = torch.Generator(device=device).manual_seed(int(seed))
|
| 723 |
+
|
| 724 |
+
# สำหรับ Waifu-Colorize-XL ใช้ prompt พิเศษ
|
| 725 |
+
enhanced_prompt = f"colorized, vibrant colors, anime style, {prompt}" if prompt else "colorized, vibrant colors, anime style, masterpiece"
|
| 726 |
+
|
| 727 |
+
with torch.inference_mode():
|
| 728 |
+
out = pipe(
|
| 729 |
+
enhanced_prompt,
|
| 730 |
+
negative_prompt=negative_prompt,
|
| 731 |
+
image=condition_img,
|
| 732 |
+
num_inference_steps=int(steps),
|
| 733 |
+
guidance_scale=float(scale),
|
| 734 |
+
controlnet_conditioning_scale=float(cn_weight),
|
| 735 |
+
generator=gen,
|
| 736 |
+
height=1024,
|
| 737 |
+
width=1024
|
| 738 |
+
).images[0]
|
| 739 |
+
|
| 740 |
+
if device.type == "cuda":
|
| 741 |
+
torch.cuda.empty_cache()
|
| 742 |
+
|
| 743 |
+
return out, condition_img
|
| 744 |
+
except Exception as e:
|
| 745 |
+
print(f"❌ Error in colorize_waifu_xl: {e}")
|
| 746 |
+
error_img = Image.new('RGB', (1024, 1024), color='red')
|
| 747 |
+
return error_img, Image.new('RGB', (1024, 1024), color='gray')
|
| 748 |
+
|
| 749 |
def t2i_sd15(prompt, negative_prompt, model, lora_model, lora_weight, vae_model,
|
| 750 |
seed, steps, scale, w, h):
|
| 751 |
"""Text-to-image for SD1.5 models"""
|
|
|
|
| 798 |
return error_img
|
| 799 |
|
| 800 |
model_to_load = model
|
| 801 |
+
if use_refiner and "refiner" not in model.lower() and not is_turbo_model(model):
|
| 802 |
model_to_load = "stabilityai/stable-diffusion-xl-refiner-1.0"
|
| 803 |
|
| 804 |
load_t2i_model(model_to_load, lora_model, lora_weight, vae_model)
|
|
|
|
| 812 |
gen = torch.Generator(device=device).manual_seed(int(seed))
|
| 813 |
|
| 814 |
with torch.inference_mode():
|
| 815 |
+
if use_refiner and CURRENT_SDXL_REFINER is not None and not is_turbo_model(model):
|
| 816 |
image = CURRENT_T2I_PIPE(
|
| 817 |
prompt=prompt,
|
| 818 |
negative_prompt=negative_prompt,
|
|
|
|
| 833 |
generator=gen
|
| 834 |
).images[0]
|
| 835 |
else:
|
| 836 |
+
if is_turbo_model(model):
|
| 837 |
+
# สำหรับ Turbo models ใช้ steps น้อยลง
|
| 838 |
+
turbo_steps = max(1, min(10, int(steps)))
|
| 839 |
+
result = CURRENT_TURBO_PIPE(
|
| 840 |
+
prompt,
|
| 841 |
+
negative_prompt=negative_prompt,
|
| 842 |
+
width=int(w),
|
| 843 |
+
height=int(h),
|
| 844 |
+
num_inference_steps=turbo_steps,
|
| 845 |
+
guidance_scale=float(scale),
|
| 846 |
+
generator=gen
|
| 847 |
+
).images[0]
|
| 848 |
+
else:
|
| 849 |
+
result = CURRENT_T2I_PIPE(
|
| 850 |
+
prompt,
|
| 851 |
+
negative_prompt=negative_prompt,
|
| 852 |
+
width=int(w),
|
| 853 |
+
height=int(h),
|
| 854 |
+
num_inference_steps=int(steps),
|
| 855 |
+
guidance_scale=float(scale),
|
| 856 |
+
generator=gen
|
| 857 |
+
).images[0]
|
| 858 |
+
|
| 859 |
+
if device.type == "cuda":
|
| 860 |
+
torch.cuda.empty_cache()
|
| 861 |
+
|
| 862 |
+
return result
|
| 863 |
+
except Exception as e:
|
| 864 |
+
print(f"❌ Error in t2i_sdxl: {e}")
|
| 865 |
+
error_img = Image.new('RGB', (int(w), int(h)), color='red')
|
| 866 |
+
from PIL import ImageDraw, ImageFont
|
| 867 |
+
draw = ImageDraw.Draw(error_img)
|
| 868 |
+
try:
|
| 869 |
+
font = ImageFont.truetype("arial.ttf", 20)
|
| 870 |
+
except:
|
| 871 |
+
font = ImageFont.load_default()
|
| 872 |
+
draw.text((50, 50), f"Error: {str(e)[:50]}...", fill="white", font=font)
|
| 873 |
+
return error_img
|
| 874 |
+
|
| 875 |
+
def t2i_turbo(prompt, negative_prompt, model, lora_model, lora_weight, vae_model,
|
| 876 |
+
seed, steps, scale, w, h):
|
| 877 |
+
"""Text-to-image for Turbo models (fast generation)"""
|
| 878 |
+
try:
|
| 879 |
+
if model not in TURBO_MODELS:
|
| 880 |
+
error_img = Image.new('RGB', (int(w), int(h)), color='red')
|
| 881 |
+
return error_img
|
| 882 |
+
|
| 883 |
+
load_t2i_model(model, lora_model, lora_weight, vae_model)
|
| 884 |
+
|
| 885 |
+
print(f"⚡ Using Turbo model: {model}")
|
| 886 |
+
if lora_model and lora_model != "None":
|
| 887 |
+
print(f" with LoRA: {lora_model} (weight: {lora_weight})")
|
| 888 |
+
|
| 889 |
+
gen = torch.Generator(device=device).manual_seed(int(seed))
|
| 890 |
+
|
| 891 |
+
with torch.inference_mode():
|
| 892 |
+
# สำหรับ Turbo models ใช้ steps น้อยลง (1-10 steps)
|
| 893 |
+
turbo_steps = max(1, min(10, int(steps)))
|
| 894 |
+
|
| 895 |
+
if is_sdxl_model(model):
|
| 896 |
+
# SDXL-Turbo
|
| 897 |
+
result = CURRENT_TURBO_PIPE(
|
| 898 |
prompt,
|
| 899 |
negative_prompt=negative_prompt,
|
| 900 |
width=int(w),
|
| 901 |
height=int(h),
|
| 902 |
+
num_inference_steps=turbo_steps,
|
| 903 |
+
guidance_scale=float(scale),
|
| 904 |
+
generator=gen
|
| 905 |
+
).images[0]
|
| 906 |
+
else:
|
| 907 |
+
# SD-Turbo
|
| 908 |
+
result = CURRENT_TURBO_PIPE(
|
| 909 |
+
prompt,
|
| 910 |
+
negative_prompt=negative_prompt,
|
| 911 |
+
width=int(w),
|
| 912 |
+
height=int(h),
|
| 913 |
+
num_inference_steps=turbo_steps,
|
| 914 |
guidance_scale=float(scale),
|
| 915 |
generator=gen
|
| 916 |
).images[0]
|
|
|
|
| 920 |
|
| 921 |
return result
|
| 922 |
except Exception as e:
|
| 923 |
+
print(f"❌ Error in t2i_turbo: {e}")
|
| 924 |
error_img = Image.new('RGB', (int(w), int(h)), color='red')
|
| 925 |
from PIL import ImageDraw, ImageFont
|
| 926 |
draw = ImageDraw.Draw(error_img)
|
|
|
|
| 935 |
global CURRENT_CONTROLNET_PIPE, CURRENT_CONTROLNET_KEY
|
| 936 |
global DETECTORS
|
| 937 |
global CURRENT_T2I_PIPE, CURRENT_T2I_MODEL, CURRENT_SDXL_REFINER
|
| 938 |
+
global CURRENT_TURBO_PIPE, CURRENT_TURBO_MODEL
|
| 939 |
|
| 940 |
print("🗑️ Unloading all models from memory...")
|
| 941 |
|
|
|
|
| 968 |
except:
|
| 969 |
pass
|
| 970 |
|
| 971 |
+
try:
|
| 972 |
+
if CURRENT_TURBO_PIPE is not None:
|
| 973 |
+
del CURRENT_TURBO_PIPE
|
| 974 |
+
CURRENT_TURBO_PIPE = None
|
| 975 |
+
except:
|
| 976 |
+
pass
|
| 977 |
+
|
| 978 |
CURRENT_T2I_MODEL = None
|
| 979 |
+
CURRENT_TURBO_MODEL = None
|
| 980 |
|
| 981 |
gc.collect()
|
| 982 |
if torch.cuda.is_available():
|
|
|
|
| 1077 |
with gr.Tab("🎨 SDXL ControlNet"):
|
| 1078 |
gr.Markdown("""
|
| 1079 |
### Transform sketches/images using SDXL with ControlNet
|
| 1080 |
+
- **Supports:** canny_sdxl, depth_sdxl, openpose_sdxl, lineart_sdxl (new!)
|
| 1081 |
- **Best Resolution:** 1024x1024
|
| 1082 |
- **Higher quality, more VRAM required**
|
| 1083 |
""")
|
|
|
|
| 1094 |
)
|
| 1095 |
controlnet_type_sdxl = gr.Dropdown(
|
| 1096 |
choices=list(CONTROLNET_MODELS_SDXL.keys()),
|
| 1097 |
+
value="lineart_sdxl",
|
| 1098 |
label="ControlNet Type"
|
| 1099 |
)
|
| 1100 |
|
|
|
|
| 1108 |
lora_weight_sdxl = gr.Slider(0.1, 2.0, 0.8, step=0.1, label="LoRA Weight")
|
| 1109 |
|
| 1110 |
vae_model_sdxl = gr.Dropdown(
|
| 1111 |
+
choices=["None", "SDXL VAE", "Turbo VAE"],
|
| 1112 |
value="None",
|
| 1113 |
label="VAE Model (Optional)"
|
| 1114 |
)
|
|
|
|
| 1144 |
[out_sdxl, condition_out_sdxl]
|
| 1145 |
)
|
| 1146 |
|
| 1147 |
+
with gr.Tab("🌸 Waifu-Colorize-XL"):
|
| 1148 |
+
gr.Markdown("""
|
| 1149 |
+
### Specialized Anime Lineart Colorization
|
| 1150 |
+
- **Model:** ShinoharaHare/Waifu-Colorize-XL
|
| 1151 |
+
- **Specialized for:** Anime/manga lineart coloring
|
| 1152 |
+
- **Features:** Automatic colorization with vibrant anime colors
|
| 1153 |
+
- **Best Resolution:** 1024x1024
|
| 1154 |
+
""")
|
| 1155 |
+
|
| 1156 |
+
with gr.Row():
|
| 1157 |
+
with gr.Column(scale=1):
|
| 1158 |
+
inp_waifu = gr.Image(label="Input Lineart/Sketch", type="pil")
|
| 1159 |
+
|
| 1160 |
+
gr.Markdown("### Model Settings")
|
| 1161 |
+
gr.Markdown("**Using:** ShinoharaHare/Waifu-Colorize-XL (Specialized Anime Colorization)")
|
| 1162 |
+
|
| 1163 |
+
gr.Markdown("### Enhancement Options")
|
| 1164 |
+
with gr.Row():
|
| 1165 |
+
lora_weight_waifu = gr.Slider(0.1, 2.0, 1.0, step=0.1, label="LoRA Weight")
|
| 1166 |
+
|
| 1167 |
+
vae_model_waifu = gr.Dropdown(
|
| 1168 |
+
choices=["None", "SDXL VAE"],
|
| 1169 |
+
value="None",
|
| 1170 |
+
label="VAE Model (Optional)"
|
| 1171 |
+
)
|
| 1172 |
+
|
| 1173 |
+
with gr.Column(scale=1):
|
| 1174 |
+
out_waifu = gr.Image(label="Colorized Output")
|
| 1175 |
+
condition_out_waifu = gr.Image(label="Processed Lineart", type="pil")
|
| 1176 |
+
|
| 1177 |
+
gr.Markdown("### Generation Parameters")
|
| 1178 |
+
with gr.Row():
|
| 1179 |
+
prompt_waifu = gr.Textbox(
|
| 1180 |
+
label="Color Style Prompt (Optional)",
|
| 1181 |
+
placeholder="vibrant colors, anime style, beautiful coloring, masterpiece",
|
| 1182 |
+
lines=2
|
| 1183 |
+
)
|
| 1184 |
+
negative_prompt_waifu = gr.Textbox(
|
| 1185 |
+
label="Negative Prompt",
|
| 1186 |
+
placeholder="monochrome, grayscale, black and white, sketch, lineart only",
|
| 1187 |
+
lines=2,
|
| 1188 |
+
value="monochrome, grayscale, black and white, sketch, lineart only"
|
| 1189 |
+
)
|
| 1190 |
+
|
| 1191 |
+
with gr.Row():
|
| 1192 |
+
seed_waifu = gr.Number(value=-1, label="Seed (-1 for random)")
|
| 1193 |
+
steps_waifu = gr.Slider(10, 50, 25, step=1, label="Steps")
|
| 1194 |
+
scale_waifu = gr.Slider(1, 20, 7.5, step=0.5, label="CFG Scale")
|
| 1195 |
+
cn_weight_waifu = gr.Slider(0.1, 2.0, 1.2, step=0.1, label="ControlNet Weight")
|
| 1196 |
+
|
| 1197 |
+
gr.Markdown("### Tips for Best Results:")
|
| 1198 |
+
gr.Markdown("""
|
| 1199 |
+
1. Use clean lineart for best results
|
| 1200 |
+
2. Higher ControlNet weight (1.0-1.5) for better line following
|
| 1201 |
+
3. Lower CFG scale (5-8) for more natural coloring
|
| 1202 |
+
4. Add color hints in prompt (e.g., "blue hair, red eyes, pink dress")
|
| 1203 |
+
5. Keep prompts simple for this specialized model
|
| 1204 |
+
""")
|
| 1205 |
+
|
| 1206 |
+
run_waifu = gr.Button("🌸 Colorize with Waifu-Colorize-XL", variant="primary", size="lg")
|
| 1207 |
+
run_waifu.click(
|
| 1208 |
+
colorize_waifu_xl,
|
| 1209 |
+
[inp_waifu, lora_weight_waifu, vae_model_waifu,
|
| 1210 |
+
prompt_waifu, negative_prompt_waifu, seed_waifu, steps_waifu, scale_waifu, cn_weight_waifu],
|
| 1211 |
+
[out_waifu, condition_out_waifu]
|
| 1212 |
+
)
|
| 1213 |
+
|
| 1214 |
with gr.Tab("🖼️ SD1.5 Text-to-Image"):
|
| 1215 |
gr.Markdown("""
|
| 1216 |
### Generate images from text descriptions using SD1.5
|
|
|
|
| 1276 |
t2i_out_sd15
|
| 1277 |
)
|
| 1278 |
|
| 1279 |
+
with gr.Tab("⚡ SDXL-Turbo Text-to-Image"):
|
| 1280 |
+
gr.Markdown("""
|
| 1281 |
+
### Ultra-Fast Image Generation with SDXL-Turbo
|
| 1282 |
+
- **Model:** stabilityai/sdxl-turbo
|
| 1283 |
+
- **Features:** 1-4 steps generation, extremely fast
|
| 1284 |
+
- **Best Resolution:** 512x512 to 1024x1024
|
| 1285 |
+
- **Warning:** Lower quality than full SDXL but much faster
|
| 1286 |
+
""")
|
| 1287 |
+
|
| 1288 |
+
with gr.Row():
|
| 1289 |
+
with gr.Column(scale=1):
|
| 1290 |
+
gr.Markdown("### Model Configuration")
|
| 1291 |
+
t2i_model_turbo = gr.Dropdown(
|
| 1292 |
+
choices=TURBO_MODELS,
|
| 1293 |
+
value="stabilityai/sdxl-turbo",
|
| 1294 |
+
label="Turbo Model"
|
| 1295 |
+
)
|
| 1296 |
+
|
| 1297 |
+
gr.Markdown("### Enhancement Options")
|
| 1298 |
+
with gr.Row():
|
| 1299 |
+
t2i_lora_turbo = gr.Dropdown(
|
| 1300 |
+
choices=list(LORA_MODELS.keys()),
|
| 1301 |
+
value="None",
|
| 1302 |
+
label="LoRA Model"
|
| 1303 |
+
)
|
| 1304 |
+
t2i_lora_weight_turbo = gr.Slider(0.1, 2.0, 0.8, step=0.1, label="LoRA Weight")
|
| 1305 |
+
|
| 1306 |
+
t2i_vae_turbo = gr.Dropdown(
|
| 1307 |
+
choices=["None", "Turbo VAE", "SDXL VAE"],
|
| 1308 |
+
value="None",
|
| 1309 |
+
label="VAE Model"
|
| 1310 |
+
)
|
| 1311 |
+
|
| 1312 |
+
with gr.Column(scale=1):
|
| 1313 |
+
t2i_out_turbo = gr.Image(label="Generated Image", type="pil")
|
| 1314 |
+
|
| 1315 |
+
gr.Markdown("### Prompts")
|
| 1316 |
+
with gr.Row():
|
| 1317 |
+
t2i_prompt_turbo = gr.Textbox(
|
| 1318 |
+
label="Prompt",
|
| 1319 |
+
lines=4,
|
| 1320 |
+
placeholder="masterpiece, best quality, highly detailed, beautiful, 1girl"
|
| 1321 |
+
)
|
| 1322 |
+
t2i_negative_prompt_turbo = gr.Textbox(
|
| 1323 |
+
label="Negative Prompt",
|
| 1324 |
+
lines=4,
|
| 1325 |
+
placeholder="lowres, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality"
|
| 1326 |
+
)
|
| 1327 |
+
|
| 1328 |
+
gr.Markdown("### Generation Parameters (Turbo needs only 1-4 steps!)")
|
| 1329 |
+
with gr.Row():
|
| 1330 |
+
t2i_seed_turbo = gr.Number(value=-1, label="Seed (-1 for random)")
|
| 1331 |
+
t2i_steps_turbo = gr.Slider(1, 10, 4, step=1, label="Steps (1-4 recommended)")
|
| 1332 |
+
t2i_scale_turbo = gr.Slider(0, 10, 0.0, step=0.5, label="CFG Scale (0-2 recommended)")
|
| 1333 |
+
|
| 1334 |
+
with gr.Row():
|
| 1335 |
+
w_turbo = gr.Slider(256, 1024, 512, step=64, label="Width")
|
| 1336 |
+
h_turbo = gr.Slider(256, 1024, 512, step=64, label="Height")
|
| 1337 |
+
|
| 1338 |
+
gr.Markdown("### Turbo Model Tips:")
|
| 1339 |
+
gr.Markdown("""
|
| 1340 |
+
1. **Use very few steps:** 1-4 steps is enough!
|
| 1341 |
+
2. **Low CFG Scale:** 0.0-2.0 works best
|
| 1342 |
+
3. **Fast but lower quality:** For quick previews/testing
|
| 1343 |
+
4. **Works well with:** Simple prompts, concept testing
|
| 1344 |
+
""")
|
| 1345 |
+
|
| 1346 |
+
gen_btn_turbo = gr.Button("⚡ Generate with Turbo (Fast!)", variant="primary", size="lg")
|
| 1347 |
+
gen_btn_turbo.click(
|
| 1348 |
+
t2i_turbo,
|
| 1349 |
+
[t2i_prompt_turbo, t2i_negative_prompt_turbo, t2i_model_turbo, t2i_lora_turbo, t2i_lora_weight_turbo,
|
| 1350 |
+
t2i_vae_turbo, t2i_seed_turbo, t2i_steps_turbo, t2i_scale_turbo, w_turbo, h_turbo],
|
| 1351 |
+
t2i_out_turbo
|
| 1352 |
+
)
|
| 1353 |
+
|
| 1354 |
with gr.Tab("🖼️ SDXL Text-to-Image"):
|
| 1355 |
gr.Markdown("""
|
| 1356 |
### Generate images from text descriptions using SDXL
|
|
|
|
| 1362 |
with gr.Column(scale=1):
|
| 1363 |
gr.Markdown("### Model Configuration")
|
| 1364 |
t2i_model_sdxl = gr.Dropdown(
|
| 1365 |
+
choices=[m for m in SDXL_MODELS if m not in TURBO_MODELS], # ไม่รวม Turbo models
|
| 1366 |
value="stabilityai/stable-diffusion-xl-base-1.0",
|
| 1367 |
label="SDXL Base Model"
|
| 1368 |
)
|
|
|
|
| 1425 |
gr.Markdown("""
|
| 1426 |
# Model & Feature Guide
|
| 1427 |
|
| 1428 |
+
## 🎯 Model Comparison
|
| 1429 |
|
| 1430 |
### SD1.5 (Stable Diffusion 1.5)
|
| 1431 |
+
- **Pros:** Fast, low VRAM, many models
|
| 1432 |
+
- **Cons:** 512px max, lower quality
|
| 1433 |
+
- **Best for:** Quick tests, anime, lower-end hardware
|
|
|
|
| 1434 |
|
| 1435 |
### SDXL (Stable Diffusion XL)
|
| 1436 |
+
- **Pros:** 1024px+, high quality, better composition
|
| 1437 |
+
- **Cons:** High VRAM, slower
|
| 1438 |
+
- **Best for:** Final quality, professional work
|
| 1439 |
+
|
| 1440 |
+
### SDXL-Turbo
|
| 1441 |
+
- **Pros:** Extremely fast (1-4 steps!)
|
| 1442 |
+
- **Cons:** Lower quality than full SDXL
|
| 1443 |
+
- **Best for:** Quick previews, concept testing
|
| 1444 |
+
|
| 1445 |
+
### Waifu-Colorize-XL
|
| 1446 |
+
- **Pros:** Specialized for anime lineart coloring
|
| 1447 |
+
- **Cons:** Anime-only, requires clean lineart
|
| 1448 |
+
- **Best for:** Anime/manga colorization
|
| 1449 |
+
|
| 1450 |
+
## 🎨 New ControlNet for SDXL
|
| 1451 |
+
|
| 1452 |
+
### Lineart ControlNet for SDXL
|
| 1453 |
+
- **Model:** ShermanG/ControlNet-Standard-Lineart-for-SDXL
|
| 1454 |
+
- **Purpose:** Convert lineart to colored images
|
| 1455 |
+
- **Best with:** Clean lineart, anime/manga styles
|
| 1456 |
+
- **Used in:** Waifu-Colorize-XL Tab
|
| 1457 |
+
|
| 1458 |
+
## 💎 Recommended Workflows
|
| 1459 |
+
|
| 1460 |
+
### For Anime/Manga Artists
|
| 1461 |
+
1. Draw lineart
|
| 1462 |
+
2. Use **Waifu-Colorize-XL** Tab for automatic coloring
|
| 1463 |
+
3. Or use **SDXL ControlNet** with lineart_sdxl
|
| 1464 |
+
|
| 1465 |
+
### For Quick Concepts
|
| 1466 |
+
1. Use **SDXL-Turbo** Tab for 1-4 step generation
|
| 1467 |
+
2. Refine in **SDXL Text-to-Image** if needed
|
| 1468 |
+
|
| 1469 |
+
### For Professional Work
|
| 1470 |
+
1. Use **SDXL Text-to-Image** with Refiner
|
| 1471 |
+
2. High steps (40-50), CFG 7-9
|
| 1472 |
+
3. 1024x1024 resolution
|
| 1473 |
+
|
| 1474 |
+
## ⚡ Turbo Model Tips
|
| 1475 |
+
|
| 1476 |
+
### SDXL-Turbo Best Practices:
|
| 1477 |
+
- **Steps:** 1-4 only!
|
| 1478 |
+
- **CFG Scale:** 0.0-2.0
|
| 1479 |
+
- **Prompts:** Keep simple
|
| 1480 |
+
- **Resolution:** 512x512 to 1024x1024
|
| 1481 |
+
- **Use case:** Storyboarding, concept art, quick iterations
|
| 1482 |
+
|
| 1483 |
+
## 🌸 Waifu-Colorize-XL Tips
|
| 1484 |
+
|
| 1485 |
+
### For Best Results:
|
| 1486 |
+
1. **Clean lineart:** No stray marks
|
| 1487 |
+
2. **ControlNet weight:** 1.0-1.5
|
| 1488 |
+
3. **CFG Scale:** 5-8
|
| 1489 |
+
4. **Simple prompts:** "vibrant colors, anime style"
|
| 1490 |
+
5. **Resolution:** 1024x1024
|
| 1491 |
+
|
| 1492 |
+
### Example Workflow:
|
| 1493 |
+
1. Draw/sketch in your favorite app
|
| 1494 |
+
2. Export as clean lineart (black on white)
|
| 1495 |
+
3. Upload to Waifu-Colorize-XL Tab
|
| 1496 |
+
4. Adjust parameters as needed
|
| 1497 |
+
5. Generate and refine
|
| 1498 |
+
|
| 1499 |
+
## 🚀 Performance Optimization
|
| 1500 |
+
|
| 1501 |
+
### Low VRAM (<8GB)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1502 |
- Use SD1.5 models only
|
| 1503 |
+
- 512x512 resolution
|
| 1504 |
- Enable attention slicing
|
|
|
|
|
|
|
| 1505 |
|
| 1506 |
+
### Medium VRAM (8-12GB)
|
| 1507 |
+
- SD1.5 and SDXL (no refiner)
|
| 1508 |
+
- 1024x1024 for SDXL
|
| 1509 |
- Enable xFormers
|
| 1510 |
|
| 1511 |
+
### High VRAM (12GB+)
|
| 1512 |
+
- All models including SDXL with refiner
|
|
|
|
| 1513 |
- Higher resolutions
|
| 1514 |
+
- Multiple LoRAs
|
| 1515 |
+
|
| 1516 |
+
## 🔄 Memory Management
|
| 1517 |
+
|
| 1518 |
+
### When to Unload Models:
|
| 1519 |
+
1. Switching between SD1.5 and SDXL
|
| 1520 |
+
2. Getting "out of memory" errors
|
| 1521 |
+
3. Changing ControlNet types
|
| 1522 |
+
4. After long generation sessions
|
| 1523 |
+
|
| 1524 |
+
### Memory Saving Tips:
|
| 1525 |
+
1. Use "Unload All Models" button
|
| 1526 |
+
2. Generate in batches
|
| 1527 |
+
3. Lower resolution for testing
|
| 1528 |
+
4. Close other GPU applications
|
| 1529 |
""")
|
| 1530 |
|
| 1531 |
try:
|