Spaces:
Build error
Build error
initial app file
Browse files
app.py
ADDED
@@ -0,0 +1,328 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import streamlit as st
|
2 |
+
import torch
|
3 |
+
import random
|
4 |
+
import subprocess
|
5 |
+
import os
|
6 |
+
from PIL import Image
|
7 |
+
from diffusers import StableDiffusionPipeline, UNet2DConditionModel
|
8 |
+
# ... any other imports you need ...
|
9 |
+
|
10 |
+
# ------------------------------------------------------------------------------
|
11 |
+
# 1. Load your models ONCE in global scope (so they don't reload on every run).
|
12 |
+
# ------------------------------------------------------------------------------
|
13 |
+
|
14 |
+
@st.cache_resource
|
15 |
+
def load_sd_pipeline(base_model_path: str, fine_tuned_path: str):
|
16 |
+
# Safety checker dummy function for demonstration:
|
17 |
+
def dummy_safety_checker(images, clip_input):
|
18 |
+
return images, False
|
19 |
+
|
20 |
+
pipe = StableDiffusionPipeline.from_pretrained(
|
21 |
+
base_model_path,
|
22 |
+
torch_dtype=torch.float16
|
23 |
+
)
|
24 |
+
pipe.to("cuda")
|
25 |
+
|
26 |
+
# Load the fine-tuned UNet
|
27 |
+
unet = UNet2DConditionModel.from_pretrained(
|
28 |
+
fine_tuned_path,
|
29 |
+
subfolder="unet",
|
30 |
+
torch_dtype=torch.float16
|
31 |
+
).to('cuda')
|
32 |
+
|
33 |
+
pipe.unet = unet
|
34 |
+
pipe.safety_checker = dummy_safety_checker
|
35 |
+
|
36 |
+
return pipe
|
37 |
+
|
38 |
+
# Similarly, if you want to load Zero123++ or other pipelines:
|
39 |
+
@st.cache_resource
|
40 |
+
def load_zero123_pipeline():
|
41 |
+
from diffusers import DiffusionPipeline, EulerAncestralDiscreteScheduler
|
42 |
+
|
43 |
+
pipeline = DiffusionPipeline.from_pretrained(
|
44 |
+
"sudo-ai/zero123plus-v1.2",
|
45 |
+
custom_pipeline="sudo-ai/zero123plus-pipeline",
|
46 |
+
torch_dtype=torch.float16
|
47 |
+
)
|
48 |
+
pipeline.scheduler = EulerAncestralDiscreteScheduler.from_config(
|
49 |
+
pipeline.scheduler.config, timestep_spacing='trailing'
|
50 |
+
)
|
51 |
+
pipeline.to("cuda")
|
52 |
+
return pipeline
|
53 |
+
|
54 |
+
|
55 |
+
# Example placeholders for the SyncDreamer command or internal functions:
|
56 |
+
def run_syncdreamer(input_path: str, output_dir: str = "syncdreamer_output"):
|
57 |
+
"""Runs SyncDreamer on input_path and places results into output_dir."""
|
58 |
+
st.info("Running SyncDreamer... (placeholder)")
|
59 |
+
# This is where your actual command would go:
|
60 |
+
# subprocess.run([...], check=True)
|
61 |
+
os.makedirs(output_dir, exist_ok=True)
|
62 |
+
# (In a real scenario, you'd handle .jpg to .png conversion, etc.)
|
63 |
+
st.success(f"SyncDreamer completed. Results in: {output_dir}")
|
64 |
+
|
65 |
+
|
66 |
+
# Helper function for Zero123++ pipeline
|
67 |
+
def make_square_min_dim(image: Image.Image, min_side: int = 320) -> Image.Image:
|
68 |
+
w, h = image.size
|
69 |
+
scale = max(min_side / w, min_side / h, 1.0)
|
70 |
+
new_w, new_h = int(w * scale), int(h * scale)
|
71 |
+
image = image.resize((new_w, new_h), Image.LANCZOS)
|
72 |
+
|
73 |
+
side = max(new_w, new_h)
|
74 |
+
new_img = Image.new(mode="RGB", size=(side, side), color=(255, 255, 255))
|
75 |
+
offset_x = (side - new_w) // 2
|
76 |
+
offset_y = (side - new_h) // 2
|
77 |
+
new_img.paste(image, (offset_x, offset_y))
|
78 |
+
return new_img
|
79 |
+
|
80 |
+
|
81 |
+
# ------------------------------------------------------------------------------
|
82 |
+
# 2. Streamlit application.
|
83 |
+
# ------------------------------------------------------------------------------
|
84 |
+
|
85 |
+
def main():
|
86 |
+
st.title("Funko Generator Demo")
|
87 |
+
|
88 |
+
# Let’s load pipelines in the background:
|
89 |
+
base_model_path = "runwayml/stable-diffusion-v1-5"
|
90 |
+
fine_tuned_path = "/content/drive/MyDrive/CC_Project/checkpoint-3000" # adapt if needed
|
91 |
+
sd_pipe = load_sd_pipeline(base_model_path, fine_tuned_path)
|
92 |
+
|
93 |
+
zero123_pipe = load_zero123_pipeline() # For multi-view generation
|
94 |
+
|
95 |
+
# Session state to hold:
|
96 |
+
if "latest_image" not in st.session_state:
|
97 |
+
st.session_state["latest_image"] = None
|
98 |
+
if "original_prompt" not in st.session_state:
|
99 |
+
st.session_state["original_prompt"] = ""
|
100 |
+
|
101 |
+
# --------------------------------------------------------------------------
|
102 |
+
# A) Prompt input & initial generation
|
103 |
+
# --------------------------------------------------------------------------
|
104 |
+
st.subheader("1. Enter your Funko prompt")
|
105 |
+
|
106 |
+
# Show examples in the UI
|
107 |
+
with st.expander("Examples of valid prompts"):
|
108 |
+
st.write("""
|
109 |
+
- A standing plain human Funko in a blue shirt and blue pants with round black eyes with glasses with a belt.
|
110 |
+
- A sitting angry animal Funko with squint black eyes.
|
111 |
+
- A standing happy robot Funko in a brown shirt and grey pants with squint black eyes with cane and monocle.
|
112 |
+
- ...
|
113 |
+
""")
|
114 |
+
|
115 |
+
user_prompt = st.text_area("Type your Funko prompt here:",
|
116 |
+
value="A standing plain human Funko in a blue shirt and blue pants with round black eyes with glasses.")
|
117 |
+
generate_button = st.button("Generate Initial Funko")
|
118 |
+
|
119 |
+
if generate_button:
|
120 |
+
st.session_state["original_prompt"] = user_prompt
|
121 |
+
with st.spinner("Generating image..."):
|
122 |
+
with torch.autocast("cuda"):
|
123 |
+
image = sd_pipe(user_prompt, num_inference_steps=50).images[0]
|
124 |
+
st.session_state["latest_image"] = image
|
125 |
+
|
126 |
+
st.success("Image generated!")
|
127 |
+
|
128 |
+
if st.session_state["latest_image"] is not None:
|
129 |
+
st.image(st.session_state["latest_image"], caption="Latest Generated Image", use_column_width=True)
|
130 |
+
|
131 |
+
# --------------------------------------------------------------------------
|
132 |
+
# B) Change the Funko (attributes)
|
133 |
+
# --------------------------------------------------------------------------
|
134 |
+
st.subheader("2. Modify Funko Attributes")
|
135 |
+
st.write("Select new attributes below. If you choose 'none', that attribute will be ignored/omitted in the prompt.")
|
136 |
+
|
137 |
+
# Possible attributes (from your code) — including 'none'
|
138 |
+
characters = ['none', 'animal', 'human', 'robot']
|
139 |
+
eyes_shape = ['none', 'anime', 'black', 'closed', 'round', 'square', 'squint']
|
140 |
+
eyes_color = ['none', 'black', 'blue', 'brown', 'green', 'grey', 'orange', 'pink', 'purple', 'red', 'white', 'yellow']
|
141 |
+
eyewear = ['none', 'eyepatch', 'glasses', 'goggles', 'helmet', 'mask', 'sunglasses']
|
142 |
+
hair_color = ['none', 'black', 'blonde', 'blue', 'brown', 'green', 'grey', 'orange', 'pink', 'purple', 'red', 'white', 'yellow']
|
143 |
+
emotion = ['none', 'angry', 'happy', 'plain', 'sad']
|
144 |
+
shirt_color = ['none', 'black', 'blue', 'brown', 'green', 'grey', 'orange', 'pink', 'purple', 'red', 'white', 'yellow']
|
145 |
+
pants_color = ['none', 'black', 'blue', 'brown', 'green', 'grey', 'orange', 'pink', 'purple', 'red', 'white', 'yellow']
|
146 |
+
accessories = ['none', 'bag', 'ball', 'belt', 'bird', 'book', 'cape', 'guitar', 'hat', 'helmet', 'sword', 'wand', 'wings']
|
147 |
+
pose = ['none', 'sitting', 'standing']
|
148 |
+
|
149 |
+
# Create selection widgets:
|
150 |
+
chosen_char = st.selectbox("Character:", characters)
|
151 |
+
chosen_eyes_shape = st.selectbox("Eyes Shape:", eyes_shape)
|
152 |
+
chosen_eyes_color = st.selectbox("Eyes Color:", eyes_color)
|
153 |
+
chosen_eyewear = st.selectbox("Eyewear:", eyewear)
|
154 |
+
chosen_hair_color = st.selectbox("Hair Color:", hair_color)
|
155 |
+
chosen_emotion = st.selectbox("Emotion:", emotion)
|
156 |
+
chosen_shirt_color = st.selectbox("Shirt Color:", shirt_color)
|
157 |
+
chosen_pants_color = st.selectbox("Pants Color:", pants_color)
|
158 |
+
chosen_accessories = st.selectbox("Accessories:", accessories)
|
159 |
+
chosen_pose = st.selectbox("Pose:", pose)
|
160 |
+
|
161 |
+
# Now we form a modified prompt. For demonstration,
|
162 |
+
# let's do something simple: we take the original prompt, parse it, and
|
163 |
+
# replace only the attributes that are not 'none'.
|
164 |
+
def modify_prompt(base_prompt: str):
|
165 |
+
# A simple example: we can build a new prompt from scratch, ignoring the old text.
|
166 |
+
# In reality, you might parse the old text or do something more sophisticated.
|
167 |
+
new_prompt_segments = []
|
168 |
+
|
169 |
+
# Pose
|
170 |
+
if chosen_pose != 'none':
|
171 |
+
new_prompt_segments.append(f"A {chosen_pose}")
|
172 |
+
else:
|
173 |
+
new_prompt_segments.append("A standing") # default fallback
|
174 |
+
|
175 |
+
# Emotion + Character
|
176 |
+
if chosen_emotion != 'none':
|
177 |
+
new_prompt_segments.append(chosen_emotion)
|
178 |
+
else:
|
179 |
+
new_prompt_segments.append("plain") # fallback
|
180 |
+
|
181 |
+
if chosen_char != 'none':
|
182 |
+
new_prompt_segments.append(chosen_char + " Funko")
|
183 |
+
else:
|
184 |
+
new_prompt_segments.append("human Funko")
|
185 |
+
|
186 |
+
# Shirt color
|
187 |
+
if chosen_shirt_color != 'none':
|
188 |
+
new_prompt_segments.append(f"in a {chosen_shirt_color} shirt")
|
189 |
+
else:
|
190 |
+
new_prompt_segments.append("in a blue shirt")
|
191 |
+
|
192 |
+
# Pants color
|
193 |
+
if chosen_pants_color != 'none':
|
194 |
+
new_prompt_segments.append(f"and {chosen_pants_color} pants")
|
195 |
+
else:
|
196 |
+
new_prompt_segments.append("and blue pants")
|
197 |
+
|
198 |
+
# Eyes
|
199 |
+
eye_text = []
|
200 |
+
if chosen_eyes_shape != 'none':
|
201 |
+
eye_text.append(f"{chosen_eyes_shape}")
|
202 |
+
else:
|
203 |
+
eye_text.append("round")
|
204 |
+
if chosen_eyes_color != 'none':
|
205 |
+
eye_text.append(f"{chosen_eyes_color}")
|
206 |
+
else:
|
207 |
+
eye_text.append("black")
|
208 |
+
eye_text.append("eyes")
|
209 |
+
new_prompt_segments.append("with " + " ".join(eye_text))
|
210 |
+
|
211 |
+
# Eyewear
|
212 |
+
if chosen_eyewear != 'none':
|
213 |
+
new_prompt_segments.append(f"with {chosen_eyewear}")
|
214 |
+
|
215 |
+
# Hair
|
216 |
+
if chosen_hair_color != 'none':
|
217 |
+
new_prompt_segments.append(f"with {chosen_hair_color} hair")
|
218 |
+
|
219 |
+
# Accessories
|
220 |
+
if chosen_accessories != 'none':
|
221 |
+
new_prompt_segments.append(f"with a {chosen_accessories}")
|
222 |
+
|
223 |
+
return " ".join(new_prompt_segments) + "."
|
224 |
+
|
225 |
+
if st.button("Generate Modified Funko"):
|
226 |
+
if not st.session_state["original_prompt"]:
|
227 |
+
st.warning("Please generate an initial Funko (step 1) before modifying it.")
|
228 |
+
else:
|
229 |
+
new_prompt = modify_prompt(st.session_state["original_prompt"])
|
230 |
+
st.write(f"**New Prompt**: {new_prompt}")
|
231 |
+
|
232 |
+
with st.spinner("Generating modified image..."):
|
233 |
+
with torch.autocast("cuda"):
|
234 |
+
image = sd_pipe(new_prompt, num_inference_steps=50).images[0]
|
235 |
+
st.session_state["latest_image"] = image
|
236 |
+
|
237 |
+
st.image(st.session_state["latest_image"], caption="Modified Image", use_column_width=True)
|
238 |
+
|
239 |
+
# --------------------------------------------------------------------------
|
240 |
+
# C) Animate the Funko with SyncDreamer
|
241 |
+
# --------------------------------------------------------------------------
|
242 |
+
st.subheader("3. Animate the Funko (SyncDreamer)")
|
243 |
+
st.write("Click the button to run SyncDreamer on the last generated image. (Demo)")
|
244 |
+
|
245 |
+
if st.button("Animate with SyncDreamer"):
|
246 |
+
if st.session_state["latest_image"] is None:
|
247 |
+
st.warning("No image found. Please generate a Funko first.")
|
248 |
+
else:
|
249 |
+
# Save latest image locally so SyncDreamer can process it
|
250 |
+
input_path = "latest_funko.png"
|
251 |
+
st.session_state["latest_image"].save(input_path)
|
252 |
+
run_syncdreamer(input_path, output_dir="syncdreamer_output")
|
253 |
+
|
254 |
+
# Optionally display a placeholder or actual frames/GIF
|
255 |
+
# ...
|
256 |
+
st.success("SyncDreamer animation completed (placeholder).")
|
257 |
+
|
258 |
+
# --------------------------------------------------------------------------
|
259 |
+
# D) Multi-View 3D Funko (Zero123++)
|
260 |
+
# --------------------------------------------------------------------------
|
261 |
+
st.subheader("4. Generate Multi-View 3D Funko (Zero123++)")
|
262 |
+
|
263 |
+
if st.button("Generate Multi-View 3D"):
|
264 |
+
if st.session_state["latest_image"] is None:
|
265 |
+
st.warning("No image found. Please generate a Funko first.")
|
266 |
+
else:
|
267 |
+
# Save the last image as input for Zero123
|
268 |
+
input_path = "funko_for_zero123.png"
|
269 |
+
st.session_state["latest_image"].save(input_path)
|
270 |
+
|
271 |
+
# Make sure image is at least 320x320 and square
|
272 |
+
original_img = Image.open(input_path).convert("RGB")
|
273 |
+
cond = make_square_min_dim(original_img, min_side=320)
|
274 |
+
|
275 |
+
# Inference
|
276 |
+
st.info("Running Zero123++ pipeline... Please wait.")
|
277 |
+
with torch.autocast("cuda"):
|
278 |
+
result_grid = zero123_pipe(cond, num_inference_steps=50).images[0]
|
279 |
+
|
280 |
+
result_grid.save("zero123_grid.png")
|
281 |
+
st.image(result_grid, caption="Zero123++ Multi-View Grid (640x960)")
|
282 |
+
|
283 |
+
# Optionally crop and display sub-views
|
284 |
+
# Here we crop 6 sub-images of 320x320 from the 640x960 grid:
|
285 |
+
coords = [
|
286 |
+
(0, 0, 320, 320),
|
287 |
+
(320, 0, 640, 320),
|
288 |
+
(0, 320, 320, 640),
|
289 |
+
(320, 320, 640, 640),
|
290 |
+
(0, 640, 320, 960),
|
291 |
+
(320, 640, 640, 960),
|
292 |
+
]
|
293 |
+
st.write("### Generated Views:")
|
294 |
+
for i, (x1, y1, x2, y2) in enumerate(coords):
|
295 |
+
sub_img = result_grid.crop((x1, y1, x2, y2))
|
296 |
+
sub_path = f"zero123_view_{i}.png"
|
297 |
+
sub_img.save(sub_path)
|
298 |
+
st.image(sub_path, width=256)
|
299 |
+
|
300 |
+
# --------------------------------------------------------------------------
|
301 |
+
# E) Integrate a New Background
|
302 |
+
# --------------------------------------------------------------------------
|
303 |
+
st.subheader("5. Apply a New Background to Each View")
|
304 |
+
|
305 |
+
st.write("Upload a background image, then apply it to each previously generated view.")
|
306 |
+
bg_file = st.file_uploader("Upload Background Image", type=["png", "jpg", "jpeg"])
|
307 |
+
if bg_file is not None:
|
308 |
+
st.image(bg_file, caption="Selected Background", width=200)
|
309 |
+
|
310 |
+
if st.button("Apply Background to Multi-View"):
|
311 |
+
if bg_file is None:
|
312 |
+
st.warning("No background uploaded.")
|
313 |
+
else:
|
314 |
+
# Save background to disk:
|
315 |
+
bg_path = "background.png"
|
316 |
+
with open(bg_path, "wb") as f:
|
317 |
+
f.write(bg_file.read())
|
318 |
+
|
319 |
+
# In a real implementation, you would do the compositing described
|
320 |
+
# in your original code with threshold-based masking, etc.
|
321 |
+
# For demonstration, let's just say "Applied!"
|
322 |
+
st.success("Background compositing placeholder done. Check your images in the output folder.")
|
323 |
+
|
324 |
+
st.write("End of the Demo. Adjust code as needed for your pipeline paths and logic.")
|
325 |
+
|
326 |
+
|
327 |
+
if __name__ == "__main__":
|
328 |
+
main()
|