Spaces:
Running
on
Zero
Running
on
Zero
update app
Browse files
app.py
CHANGED
|
@@ -1,6 +1,7 @@
|
|
| 1 |
import os
|
| 2 |
import gradio as gr
|
| 3 |
import numpy as np
|
|
|
|
| 4 |
import torch
|
| 5 |
import random
|
| 6 |
from PIL import Image
|
|
@@ -8,18 +9,6 @@ from typing import Iterable
|
|
| 8 |
from gradio.themes import Soft
|
| 9 |
from gradio.themes.utils import colors, fonts, sizes
|
| 10 |
|
| 11 |
-
# --- Handle optional 'spaces' import for local compatibility ---
|
| 12 |
-
try:
|
| 13 |
-
import spaces
|
| 14 |
-
except ImportError:
|
| 15 |
-
class spaces:
|
| 16 |
-
@staticmethod
|
| 17 |
-
def GPU(duration=60):
|
| 18 |
-
def decorator(func):
|
| 19 |
-
return func
|
| 20 |
-
return decorator
|
| 21 |
-
|
| 22 |
-
# --- Custom Theme Setup ---
|
| 23 |
colors.steel_blue = colors.Color(
|
| 24 |
name="steel_blue",
|
| 25 |
c50="#EBF3F8",
|
|
@@ -88,17 +77,14 @@ class SteelBlueTheme(Soft):
|
|
| 88 |
|
| 89 |
steel_blue_theme = SteelBlueTheme()
|
| 90 |
|
| 91 |
-
# --- Hardware Setup ---
|
| 92 |
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
| 93 |
dtype = torch.bfloat16
|
| 94 |
|
| 95 |
-
# --- Imports for Custom Pipeline ---
|
| 96 |
from diffusers import FlowMatchEulerDiscreteScheduler
|
| 97 |
from qwenimage.pipeline_qwenimage_edit_plus import QwenImageEditPlusPipeline
|
| 98 |
from qwenimage.transformer_qwenimage import QwenImageTransformer2DModel
|
| 99 |
from qwenimage.qwen_fa3_processor import QwenDoubleStreamAttnProcessorFA3
|
| 100 |
|
| 101 |
-
# --- Model Initialization ---
|
| 102 |
print("Loading Qwen Image Edit Pipeline...")
|
| 103 |
pipe = QwenImageEditPlusPipeline.from_pretrained(
|
| 104 |
"Qwen/Qwen-Image-Edit-2509",
|
|
@@ -111,32 +97,26 @@ pipe = QwenImageEditPlusPipeline.from_pretrained(
|
|
| 111 |
torch_dtype=dtype
|
| 112 |
).to(device)
|
| 113 |
|
| 114 |
-
# 1. Load and Fuse Lightning (for speed)
|
| 115 |
print("Loading and Fusing Lightning LoRA...")
|
| 116 |
pipe.load_lora_weights("lightx2v/Qwen-Image-Lightning",
|
| 117 |
weight_name="Qwen-Image-Lightning-4steps-V2.0-bf16.safetensors",
|
| 118 |
adapter_name="lightning")
|
| 119 |
pipe.fuse_lora(adapter_names=["lightning"], lora_scale=1.0)
|
| 120 |
|
| 121 |
-
# 2. Load Task Specific LoRAs
|
| 122 |
print("Loading Task Adapters...")
|
| 123 |
|
| 124 |
-
# Texture
|
| 125 |
pipe.load_lora_weights("tarn59/apply_texture_qwen_image_edit_2509",
|
| 126 |
weight_name="apply_texture_v2_qwen_image_edit_2509.safetensors",
|
| 127 |
adapter_name="texture")
|
| 128 |
|
| 129 |
-
# Fusion (Fuse-Objects)
|
| 130 |
pipe.load_lora_weights("dx8152/Qwen-Image-Edit-2509-Fusion",
|
| 131 |
weight_name="溶图.safetensors",
|
| 132 |
adapter_name="fusion")
|
| 133 |
|
| 134 |
-
# Face Swap
|
| 135 |
pipe.load_lora_weights("Alissonerdx/BFS-Best-Face-Swap",
|
| 136 |
weight_name="bfs_head_v3_qwen_image_edit_2509.safetensors",
|
| 137 |
adapter_name="faceswap")
|
| 138 |
|
| 139 |
-
# Attempt Flash Attention 3
|
| 140 |
try:
|
| 141 |
pipe.transformer.set_attn_processor(QwenDoubleStreamAttnProcessorFA3())
|
| 142 |
print("Flash Attention 3 Processor set successfully.")
|
|
@@ -166,7 +146,7 @@ def update_dimensions_on_upload(image):
|
|
| 166 |
|
| 167 |
return new_width, new_height
|
| 168 |
|
| 169 |
-
@spaces.GPU(duration=
|
| 170 |
def infer(
|
| 171 |
image_1,
|
| 172 |
image_2,
|
|
@@ -182,7 +162,6 @@ def infer(
|
|
| 182 |
raise gr.Error("Please upload both images for Fusion/Texture/FaceSwap tasks.")
|
| 183 |
|
| 184 |
if not prompt:
|
| 185 |
-
# Add default prompts based on mode if user leaves empty (optional helper)
|
| 186 |
if lora_adapter == "Face-Swap":
|
| 187 |
prompt = "Swap the face."
|
| 188 |
elif lora_adapter == "Texture Edit":
|
|
@@ -190,8 +169,6 @@ def infer(
|
|
| 190 |
elif lora_adapter == "Fuse-Objects":
|
| 191 |
prompt = "Fuse object into background."
|
| 192 |
|
| 193 |
-
# Switch Adapters
|
| 194 |
-
# Note: Lightning is already fused, so we just enable the style adapter
|
| 195 |
adapters_map = {
|
| 196 |
"Texture Edit": "texture",
|
| 197 |
"Fuse-Objects": "fusion",
|
|
@@ -211,15 +188,13 @@ def infer(
|
|
| 211 |
generator = torch.Generator(device=device).manual_seed(seed)
|
| 212 |
negative_prompt = "worst quality, low quality, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, jpeg artifacts, signature, watermark, username, blurry"
|
| 213 |
|
| 214 |
-
# Prepare Images
|
| 215 |
img1_pil = image_1.convert("RGB")
|
| 216 |
img2_pil = image_2.convert("RGB")
|
| 217 |
|
| 218 |
-
# Calculate dimensions based on the primary image (Image 1)
|
| 219 |
width, height = update_dimensions_on_upload(img1_pil)
|
| 220 |
|
| 221 |
result = pipe(
|
| 222 |
-
image=[img1_pil, img2_pil],
|
| 223 |
prompt=prompt,
|
| 224 |
negative_prompt=negative_prompt,
|
| 225 |
height=height,
|
|
@@ -231,9 +206,8 @@ def infer(
|
|
| 231 |
|
| 232 |
return result, seed
|
| 233 |
|
| 234 |
-
@spaces.GPU(duration=
|
| 235 |
def infer_example(image_1, image_2, prompt, lora_adapter):
|
| 236 |
-
# Wrapper for examples that sets defaults
|
| 237 |
if image_1 is None or image_2 is None:
|
| 238 |
return None, 0
|
| 239 |
result, seed = infer(
|
|
@@ -241,10 +215,10 @@ def infer_example(image_1, image_2, prompt, lora_adapter):
|
|
| 241 |
image_2.convert("RGB"),
|
| 242 |
prompt,
|
| 243 |
lora_adapter,
|
| 244 |
-
0,
|
| 245 |
-
True,
|
| 246 |
-
1.0,
|
| 247 |
-
4
|
| 248 |
)
|
| 249 |
return result, seed
|
| 250 |
|
|
@@ -261,7 +235,7 @@ with gr.Blocks(css=css, theme=steel_blue_theme) as demo:
|
|
| 261 |
gr.Markdown("# **Qwen-Image-Edit-2509-LoRAs-Fast-Fusion**", elem_id="main-title")
|
| 262 |
gr.Markdown("Perform diverse image edits using specialized [LoRA](https://huggingface.co/models?other=base_model:adapter:Qwen/Qwen-Image-Edit-2509) adapters for the [Qwen-Image-Edit](https://huggingface.co/Qwen/Qwen-Image-Edit-2509) model.")
|
| 263 |
with gr.Row(equal_height=True):
|
| 264 |
-
|
| 265 |
with gr.Column(scale=1):
|
| 266 |
with gr.Row():
|
| 267 |
image_1 = gr.Image(label="Base / Background / Body", type="pil", height=290)
|
|
@@ -301,7 +275,7 @@ with gr.Blocks(css=css, theme=steel_blue_theme) as demo:
|
|
| 301 |
outputs=[output_image, seed],
|
| 302 |
fn=infer_example,
|
| 303 |
cache_examples=False,
|
| 304 |
-
label="Examples
|
| 305 |
)
|
| 306 |
|
| 307 |
run_button.click(
|
|
|
|
| 1 |
import os
|
| 2 |
import gradio as gr
|
| 3 |
import numpy as np
|
| 4 |
+
import spaces
|
| 5 |
import torch
|
| 6 |
import random
|
| 7 |
from PIL import Image
|
|
|
|
| 9 |
from gradio.themes import Soft
|
| 10 |
from gradio.themes.utils import colors, fonts, sizes
|
| 11 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 12 |
colors.steel_blue = colors.Color(
|
| 13 |
name="steel_blue",
|
| 14 |
c50="#EBF3F8",
|
|
|
|
| 77 |
|
| 78 |
steel_blue_theme = SteelBlueTheme()
|
| 79 |
|
|
|
|
| 80 |
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
| 81 |
dtype = torch.bfloat16
|
| 82 |
|
|
|
|
| 83 |
from diffusers import FlowMatchEulerDiscreteScheduler
|
| 84 |
from qwenimage.pipeline_qwenimage_edit_plus import QwenImageEditPlusPipeline
|
| 85 |
from qwenimage.transformer_qwenimage import QwenImageTransformer2DModel
|
| 86 |
from qwenimage.qwen_fa3_processor import QwenDoubleStreamAttnProcessorFA3
|
| 87 |
|
|
|
|
| 88 |
print("Loading Qwen Image Edit Pipeline...")
|
| 89 |
pipe = QwenImageEditPlusPipeline.from_pretrained(
|
| 90 |
"Qwen/Qwen-Image-Edit-2509",
|
|
|
|
| 97 |
torch_dtype=dtype
|
| 98 |
).to(device)
|
| 99 |
|
|
|
|
| 100 |
print("Loading and Fusing Lightning LoRA...")
|
| 101 |
pipe.load_lora_weights("lightx2v/Qwen-Image-Lightning",
|
| 102 |
weight_name="Qwen-Image-Lightning-4steps-V2.0-bf16.safetensors",
|
| 103 |
adapter_name="lightning")
|
| 104 |
pipe.fuse_lora(adapter_names=["lightning"], lora_scale=1.0)
|
| 105 |
|
|
|
|
| 106 |
print("Loading Task Adapters...")
|
| 107 |
|
|
|
|
| 108 |
pipe.load_lora_weights("tarn59/apply_texture_qwen_image_edit_2509",
|
| 109 |
weight_name="apply_texture_v2_qwen_image_edit_2509.safetensors",
|
| 110 |
adapter_name="texture")
|
| 111 |
|
|
|
|
| 112 |
pipe.load_lora_weights("dx8152/Qwen-Image-Edit-2509-Fusion",
|
| 113 |
weight_name="溶图.safetensors",
|
| 114 |
adapter_name="fusion")
|
| 115 |
|
|
|
|
| 116 |
pipe.load_lora_weights("Alissonerdx/BFS-Best-Face-Swap",
|
| 117 |
weight_name="bfs_head_v3_qwen_image_edit_2509.safetensors",
|
| 118 |
adapter_name="faceswap")
|
| 119 |
|
|
|
|
| 120 |
try:
|
| 121 |
pipe.transformer.set_attn_processor(QwenDoubleStreamAttnProcessorFA3())
|
| 122 |
print("Flash Attention 3 Processor set successfully.")
|
|
|
|
| 146 |
|
| 147 |
return new_width, new_height
|
| 148 |
|
| 149 |
+
@spaces.GPU(duration=30)
|
| 150 |
def infer(
|
| 151 |
image_1,
|
| 152 |
image_2,
|
|
|
|
| 162 |
raise gr.Error("Please upload both images for Fusion/Texture/FaceSwap tasks.")
|
| 163 |
|
| 164 |
if not prompt:
|
|
|
|
| 165 |
if lora_adapter == "Face-Swap":
|
| 166 |
prompt = "Swap the face."
|
| 167 |
elif lora_adapter == "Texture Edit":
|
|
|
|
| 169 |
elif lora_adapter == "Fuse-Objects":
|
| 170 |
prompt = "Fuse object into background."
|
| 171 |
|
|
|
|
|
|
|
| 172 |
adapters_map = {
|
| 173 |
"Texture Edit": "texture",
|
| 174 |
"Fuse-Objects": "fusion",
|
|
|
|
| 188 |
generator = torch.Generator(device=device).manual_seed(seed)
|
| 189 |
negative_prompt = "worst quality, low quality, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, jpeg artifacts, signature, watermark, username, blurry"
|
| 190 |
|
|
|
|
| 191 |
img1_pil = image_1.convert("RGB")
|
| 192 |
img2_pil = image_2.convert("RGB")
|
| 193 |
|
|
|
|
| 194 |
width, height = update_dimensions_on_upload(img1_pil)
|
| 195 |
|
| 196 |
result = pipe(
|
| 197 |
+
image=[img1_pil, img2_pil],
|
| 198 |
prompt=prompt,
|
| 199 |
negative_prompt=negative_prompt,
|
| 200 |
height=height,
|
|
|
|
| 206 |
|
| 207 |
return result, seed
|
| 208 |
|
| 209 |
+
@spaces.GPU(duration=30)
|
| 210 |
def infer_example(image_1, image_2, prompt, lora_adapter):
|
|
|
|
| 211 |
if image_1 is None or image_2 is None:
|
| 212 |
return None, 0
|
| 213 |
result, seed = infer(
|
|
|
|
| 215 |
image_2.convert("RGB"),
|
| 216 |
prompt,
|
| 217 |
lora_adapter,
|
| 218 |
+
0,
|
| 219 |
+
True,
|
| 220 |
+
1.0,
|
| 221 |
+
4
|
| 222 |
)
|
| 223 |
return result, seed
|
| 224 |
|
|
|
|
| 235 |
gr.Markdown("# **Qwen-Image-Edit-2509-LoRAs-Fast-Fusion**", elem_id="main-title")
|
| 236 |
gr.Markdown("Perform diverse image edits using specialized [LoRA](https://huggingface.co/models?other=base_model:adapter:Qwen/Qwen-Image-Edit-2509) adapters for the [Qwen-Image-Edit](https://huggingface.co/Qwen/Qwen-Image-Edit-2509) model.")
|
| 237 |
with gr.Row(equal_height=True):
|
| 238 |
+
|
| 239 |
with gr.Column(scale=1):
|
| 240 |
with gr.Row():
|
| 241 |
image_1 = gr.Image(label="Base / Background / Body", type="pil", height=290)
|
|
|
|
| 275 |
outputs=[output_image, seed],
|
| 276 |
fn=infer_example,
|
| 277 |
cache_examples=False,
|
| 278 |
+
label="Examples"
|
| 279 |
)
|
| 280 |
|
| 281 |
run_button.click(
|