Spaces:
Running
Running
Update app.py
Browse files
app.py
CHANGED
@@ -8,8 +8,47 @@ from pathlib import Path
|
|
8 |
from PIL import Image
|
9 |
import os
|
10 |
import time
|
|
|
|
|
11 |
import spaces
|
12 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
13 |
DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
|
14 |
print(f"Using device: {DEVICE}")
|
15 |
|
@@ -18,12 +57,10 @@ DEFAULT_WIDTH = 1024
|
|
18 |
DEFAULT_GUIDANCE_SCALE = 3.5
|
19 |
DEFAULT_NUM_INFERENCE_STEPS = 15
|
20 |
DEFAULT_MAX_SEQUENCE_LENGTH = 512
|
21 |
-
GENERATION_SEED = 0 # could use a random number generator to set this, for more variety
|
22 |
HF_TOKEN = os.environ.get("HF_ACCESS_TOKEN")
|
23 |
|
24 |
CACHED_PIPES = {}
|
25 |
def load_bf16_pipeline():
|
26 |
-
"""Loads the original FLUX.1-dev pipeline in BF16 precision."""
|
27 |
print("Loading BF16 pipeline...")
|
28 |
MODEL_ID = "black-forest-labs/FLUX.1-dev"
|
29 |
if MODEL_ID in CACHED_PIPES:
|
@@ -36,7 +73,6 @@ def load_bf16_pipeline():
|
|
36 |
token=HF_TOKEN
|
37 |
)
|
38 |
pipe.to(DEVICE)
|
39 |
-
# pipe.enable_model_cpu_offload()
|
40 |
end_time = time.time()
|
41 |
mem_reserved = torch.cuda.memory_reserved(0)/1024**3 if DEVICE == "cuda" else 0
|
42 |
print(f"BF16 Pipeline loaded in {end_time - start_time:.2f}s. Memory reserved: {mem_reserved:.2f} GB")
|
@@ -44,10 +80,9 @@ def load_bf16_pipeline():
|
|
44 |
return pipe
|
45 |
except Exception as e:
|
46 |
print(f"Error loading BF16 pipeline: {e}")
|
47 |
-
raise
|
48 |
|
49 |
def load_bnb_8bit_pipeline():
|
50 |
-
"""Loads the FLUX.1-dev pipeline with 8-bit quantized components."""
|
51 |
print("Loading 8-bit BNB pipeline...")
|
52 |
MODEL_ID = "derekl35/FLUX.1-dev-bnb-8bit"
|
53 |
if MODEL_ID in CACHED_PIPES:
|
@@ -70,7 +105,6 @@ def load_bnb_8bit_pipeline():
|
|
70 |
raise
|
71 |
|
72 |
def load_bnb_4bit_pipeline():
|
73 |
-
"""Loads the FLUX.1-dev pipeline with 4-bit quantized components."""
|
74 |
print("Loading 4-bit BNB pipeline...")
|
75 |
MODEL_ID = "derekl35/FLUX.1-dev-nf4"
|
76 |
if MODEL_ID in CACHED_PIPES:
|
@@ -89,20 +123,17 @@ def load_bnb_4bit_pipeline():
|
|
89 |
CACHED_PIPES[MODEL_ID] = pipe
|
90 |
return pipe
|
91 |
except Exception as e:
|
92 |
-
print(f"4-bit BNB pipeline: {e}")
|
93 |
raise
|
94 |
|
95 |
@spaces.GPU(duration=240)
|
96 |
def generate_images(prompt, quantization_choice, progress=gr.Progress(track_tqdm=True)):
|
97 |
-
"""Loads original and selected quantized model, generates one image each, shuffles results."""
|
98 |
if not prompt:
|
99 |
-
return None, {}, gr.update(value="Please enter a prompt.", interactive=False), gr.update(choices=[], value=None)
|
100 |
|
101 |
if not quantization_choice:
|
102 |
-
|
103 |
-
return None, {}, gr.update(value="Please select a quantization method.", interactive=False), gr.update(choices=[], value=None)
|
104 |
|
105 |
-
# Determine which quantized model to load
|
106 |
if quantization_choice == "8-bit":
|
107 |
quantized_load_func = load_bnb_8bit_pipeline
|
108 |
quantized_label = "Quantized (8-bit)"
|
@@ -110,12 +141,11 @@ def generate_images(prompt, quantization_choice, progress=gr.Progress(track_tqdm
|
|
110 |
quantized_load_func = load_bnb_4bit_pipeline
|
111 |
quantized_label = "Quantized (4-bit)"
|
112 |
else:
|
113 |
-
|
114 |
-
return None, {}, gr.update(value="Invalid quantization choice.", interactive=False), gr.update(choices=[], value=None)
|
115 |
|
116 |
model_configs = [
|
117 |
("Original", load_bf16_pipeline),
|
118 |
-
(quantized_label, quantized_load_func),
|
119 |
]
|
120 |
|
121 |
results = []
|
@@ -128,8 +158,6 @@ def generate_images(prompt, quantization_choice, progress=gr.Progress(track_tqdm
|
|
128 |
"max_sequence_length": DEFAULT_MAX_SEQUENCE_LENGTH,
|
129 |
}
|
130 |
|
131 |
-
current_pipe = None # Keep track of the current pipe for cleanup
|
132 |
-
|
133 |
seed = random.getrandbits(64)
|
134 |
print(f"Using seed: {seed}")
|
135 |
|
@@ -147,7 +175,6 @@ def generate_images(prompt, quantization_choice, progress=gr.Progress(track_tqdm
|
|
147 |
gen_start_time = time.time()
|
148 |
image_list = current_pipe(**pipe_kwargs, generator=torch.manual_seed(seed)).images
|
149 |
image = image_list[0]
|
150 |
-
# image.save(f"{load_start_time}.png")
|
151 |
gen_end_time = time.time()
|
152 |
results.append({"label": label, "image": image})
|
153 |
print(f"--- Finished Generation with {label} Model in {gen_end_time - gen_start_time:.2f} seconds ---")
|
@@ -156,64 +183,42 @@ def generate_images(prompt, quantization_choice, progress=gr.Progress(track_tqdm
|
|
156 |
|
157 |
except Exception as e:
|
158 |
print(f"Error during {label} model processing: {e}")
|
159 |
-
|
160 |
-
return None, {}, gr.update(value=f"Error processing {label} model: {e}", interactive=False), gr.update(choices=[], value=None)
|
161 |
|
162 |
-
# No finally block needed here, cleanup happens before next load or after loop
|
163 |
|
164 |
if len(results) != len(model_configs):
|
165 |
-
|
166 |
-
# Update all outputs
|
167 |
-
return None, {}, gr.update(value="Failed to generate images for all model types.", interactive=False), gr.update(choices=[], value=None)
|
168 |
|
169 |
-
# Shuffle the results for display
|
170 |
shuffled_results = results.copy()
|
171 |
random.shuffle(shuffled_results)
|
172 |
-
|
173 |
-
# Create the gallery data: [(image, caption), (image, caption)]
|
174 |
shuffled_data_for_gallery = [(res["image"], f"Image {i+1}") for i, res in enumerate(shuffled_results)]
|
175 |
-
|
176 |
-
# Create the mapping: display_index -> correct_label (e.g., {0: 'Original', 1: 'Quantized (8-bit)'})
|
177 |
correct_mapping = {i: res["label"] for i, res in enumerate(shuffled_results)}
|
178 |
print("Correct mapping (hidden):", correct_mapping)
|
179 |
|
180 |
-
|
181 |
|
182 |
-
# Return shuffled images, the correct mapping state, status message, and update the guess radio
|
183 |
-
return shuffled_data_for_gallery, correct_mapping, gr.update(value="Generation complete! Make your guess.", interactive=False), guess_radio_update
|
184 |
|
185 |
-
|
186 |
-
# --- Guess Verification Function ---
|
187 |
def check_guess(user_guess, correct_mapping_state):
|
188 |
-
"""Compares the user's guess with the correct mapping stored in the state."""
|
189 |
-
|
190 |
if not isinstance(correct_mapping_state, dict) or not correct_mapping_state:
|
191 |
return "Please generate images first (state is empty or invalid)."
|
192 |
-
|
193 |
if user_guess is None:
|
194 |
return "Please select which image you think is quantized."
|
195 |
|
196 |
-
# Find which display index (0 or 1) corresponds to the quantized image
|
197 |
quantized_image_index = -1
|
198 |
quantized_label_actual = ""
|
199 |
for index, label in correct_mapping_state.items():
|
200 |
-
if "Quantized" in label:
|
201 |
quantized_image_index = index
|
202 |
-
quantized_label_actual = label
|
203 |
break
|
204 |
-
|
205 |
if quantized_image_index == -1:
|
206 |
-
# This shouldn't happen if generation was successful
|
207 |
return "Error: Could not find the quantized image in the mapping data."
|
208 |
|
209 |
-
|
210 |
-
correct_guess_label = f"Image {quantized_image_index + 1}" # "Image 1" or "Image 2"
|
211 |
-
|
212 |
if user_guess == correct_guess_label:
|
213 |
feedback = f"Correct! {correct_guess_label} used the {quantized_label_actual} model."
|
214 |
else:
|
215 |
feedback = f"Incorrect. The quantized image ({quantized_label_actual}) was {correct_guess_label}."
|
216 |
-
|
217 |
return feedback
|
218 |
|
219 |
EXAMPLE_DIR = Path(__file__).parent / "examples"
|
@@ -221,12 +226,14 @@ EXAMPLES = [
|
|
221 |
{
|
222 |
"prompt": "A photorealistic portrait of an astronaut on Mars",
|
223 |
"files": ["astronauts_seed_6456306350371904162.png", "astronauts_bnb_8bit.png"],
|
224 |
-
"quantized_idx": 1,
|
|
|
225 |
},
|
226 |
{
|
227 |
"prompt": "Water-color painting of a cat wearing sunglasses",
|
228 |
"files": ["watercolor_cat_bnb_8bit.png", "watercolor_cat_seed_14269059182221286790.png"],
|
229 |
"quantized_idx": 0,
|
|
|
230 |
},
|
231 |
# {
|
232 |
# "prompt": "Neo-tokyo cyberpunk cityscape at night, rain-soaked streets, 8-K",
|
@@ -236,97 +243,304 @@ EXAMPLES = [
|
|
236 |
]
|
237 |
|
238 |
def load_example(idx):
|
239 |
-
"""Return [(PIL.Image, caption)...], mapping dict, and feedback string"""
|
240 |
ex = EXAMPLES[idx]
|
241 |
imgs = [Image.open(EXAMPLE_DIR / f) for f in ex["files"]]
|
242 |
gallery_items = [(img, f"Image {i+1}") for i, img in enumerate(imgs)]
|
243 |
-
mapping = {i: ("Quantized" if i == ex["quantized_idx"] else "Original")
|
244 |
for i in range(2)}
|
245 |
return gallery_items, mapping, f"{ex['prompt']}"
|
246 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
247 |
with gr.Blocks(title="FLUX Quantization Challenge", theme=gr.themes.Soft()) as demo:
|
248 |
gr.Markdown("# FLUX Model Quantization Challenge")
|
249 |
-
gr.
|
250 |
-
|
251 |
-
|
252 |
-
|
253 |
-
|
254 |
-
|
255 |
-
|
256 |
-
|
257 |
-
|
258 |
-
|
259 |
-
|
260 |
-
|
261 |
-
|
262 |
-
|
263 |
-
|
264 |
-
|
265 |
-
|
266 |
-
|
267 |
-
|
268 |
-
|
269 |
-
|
270 |
-
|
271 |
-
|
272 |
-
|
273 |
-
|
274 |
-
|
275 |
-
|
276 |
-
|
277 |
-
|
278 |
-
|
279 |
-
|
280 |
-
|
281 |
-
|
282 |
-
|
283 |
-
|
284 |
-
|
285 |
-
|
286 |
-
|
287 |
-
|
288 |
-
|
289 |
-
|
290 |
-
|
291 |
-
|
292 |
-
|
293 |
-
|
294 |
-
|
295 |
-
|
296 |
-
|
297 |
-
|
298 |
-
|
299 |
-
|
300 |
-
|
301 |
-
|
302 |
-
|
303 |
-
|
304 |
-
|
305 |
-
|
306 |
-
|
307 |
-
|
308 |
-
|
309 |
-
|
310 |
-
|
311 |
-
|
312 |
-
|
313 |
-
|
314 |
-
|
315 |
-
|
316 |
-
|
317 |
-
|
318 |
-
|
319 |
-
|
320 |
-
|
321 |
-
|
322 |
-
|
323 |
-
|
324 |
-
|
325 |
-
|
326 |
-
|
327 |
-
|
328 |
-
|
329 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
330 |
|
331 |
if __name__ == "__main__":
|
332 |
demo.launch(share=True)
|
|
|
8 |
from PIL import Image
|
9 |
import os
|
10 |
import time
|
11 |
+
import json
|
12 |
+
from fasteners import InterProcessLock
|
13 |
import spaces
|
14 |
|
15 |
+
AGG_FILE = Path(__file__).parent / "agg_stats.json"
|
16 |
+
LOCK_FILE = AGG_FILE.with_suffix(".lock")
|
17 |
+
|
18 |
+
def _load_agg_stats() -> dict:
|
19 |
+
if AGG_FILE.exists():
|
20 |
+
with open(AGG_FILE, "r") as f:
|
21 |
+
try:
|
22 |
+
return json.load(f)
|
23 |
+
except json.JSONDecodeError:
|
24 |
+
print(f"Warning: {AGG_FILE} is corrupted. Starting with empty stats.")
|
25 |
+
return {"8-bit": {"attempts": 0, "correct": 0}, "4-bit": {"attempts": 0, "correct": 0}}
|
26 |
+
return {"8-bit": {"attempts": 0, "correct": 0},
|
27 |
+
"4-bit": {"attempts": 0, "correct": 0}}
|
28 |
+
|
29 |
+
def _save_agg_stats(stats: dict) -> None:
|
30 |
+
with InterProcessLock(str(LOCK_FILE)):
|
31 |
+
with open(AGG_FILE, "w") as f:
|
32 |
+
json.dump(stats, f, indent=2)
|
33 |
+
|
34 |
+
USER_STATS_FILE = Path(__file__).parent / "user_stats.json"
|
35 |
+
USER_STATS_LOCK_FILE = USER_STATS_FILE.with_suffix(".lock")
|
36 |
+
|
37 |
+
def _load_user_stats() -> dict:
|
38 |
+
if USER_STATS_FILE.exists():
|
39 |
+
with open(USER_STATS_FILE, "r") as f:
|
40 |
+
try:
|
41 |
+
return json.load(f)
|
42 |
+
except json.JSONDecodeError:
|
43 |
+
print(f"Warning: {USER_STATS_FILE} is corrupted. Starting with empty user stats.")
|
44 |
+
return {}
|
45 |
+
return {}
|
46 |
+
|
47 |
+
def _save_user_stats(stats: dict) -> None:
|
48 |
+
with InterProcessLock(str(USER_STATS_LOCK_FILE)):
|
49 |
+
with open(USER_STATS_FILE, "w") as f:
|
50 |
+
json.dump(stats, f, indent=2)
|
51 |
+
|
52 |
DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
|
53 |
print(f"Using device: {DEVICE}")
|
54 |
|
|
|
57 |
DEFAULT_GUIDANCE_SCALE = 3.5
|
58 |
DEFAULT_NUM_INFERENCE_STEPS = 15
|
59 |
DEFAULT_MAX_SEQUENCE_LENGTH = 512
|
|
|
60 |
HF_TOKEN = os.environ.get("HF_ACCESS_TOKEN")
|
61 |
|
62 |
CACHED_PIPES = {}
|
63 |
def load_bf16_pipeline():
|
|
|
64 |
print("Loading BF16 pipeline...")
|
65 |
MODEL_ID = "black-forest-labs/FLUX.1-dev"
|
66 |
if MODEL_ID in CACHED_PIPES:
|
|
|
73 |
token=HF_TOKEN
|
74 |
)
|
75 |
pipe.to(DEVICE)
|
|
|
76 |
end_time = time.time()
|
77 |
mem_reserved = torch.cuda.memory_reserved(0)/1024**3 if DEVICE == "cuda" else 0
|
78 |
print(f"BF16 Pipeline loaded in {end_time - start_time:.2f}s. Memory reserved: {mem_reserved:.2f} GB")
|
|
|
80 |
return pipe
|
81 |
except Exception as e:
|
82 |
print(f"Error loading BF16 pipeline: {e}")
|
83 |
+
raise
|
84 |
|
85 |
def load_bnb_8bit_pipeline():
|
|
|
86 |
print("Loading 8-bit BNB pipeline...")
|
87 |
MODEL_ID = "derekl35/FLUX.1-dev-bnb-8bit"
|
88 |
if MODEL_ID in CACHED_PIPES:
|
|
|
105 |
raise
|
106 |
|
107 |
def load_bnb_4bit_pipeline():
|
|
|
108 |
print("Loading 4-bit BNB pipeline...")
|
109 |
MODEL_ID = "derekl35/FLUX.1-dev-nf4"
|
110 |
if MODEL_ID in CACHED_PIPES:
|
|
|
123 |
CACHED_PIPES[MODEL_ID] = pipe
|
124 |
return pipe
|
125 |
except Exception as e:
|
126 |
+
print(f"Error loading 4-bit BNB pipeline: {e}")
|
127 |
raise
|
128 |
|
129 |
@spaces.GPU(duration=240)
|
130 |
def generate_images(prompt, quantization_choice, progress=gr.Progress(track_tqdm=True)):
|
|
|
131 |
if not prompt:
|
132 |
+
return None, {}, gr.update(value="Please enter a prompt.", interactive=False), gr.update(choices=[], value=None), gr.update(interactive=True), gr.update(interactive=True)
|
133 |
|
134 |
if not quantization_choice:
|
135 |
+
return None, {}, gr.update(value="Please select a quantization method.", interactive=False), gr.update(choices=[], value=None), gr.update(interactive=True), gr.update(interactive=True)
|
|
|
136 |
|
|
|
137 |
if quantization_choice == "8-bit":
|
138 |
quantized_load_func = load_bnb_8bit_pipeline
|
139 |
quantized_label = "Quantized (8-bit)"
|
|
|
141 |
quantized_load_func = load_bnb_4bit_pipeline
|
142 |
quantized_label = "Quantized (4-bit)"
|
143 |
else:
|
144 |
+
return None, {}, gr.update(value="Invalid quantization choice.", interactive=False), gr.update(choices=[], value=None), gr.update(interactive=True), gr.update(interactive=True)
|
|
|
145 |
|
146 |
model_configs = [
|
147 |
("Original", load_bf16_pipeline),
|
148 |
+
(quantized_label, quantized_load_func),
|
149 |
]
|
150 |
|
151 |
results = []
|
|
|
158 |
"max_sequence_length": DEFAULT_MAX_SEQUENCE_LENGTH,
|
159 |
}
|
160 |
|
|
|
|
|
161 |
seed = random.getrandbits(64)
|
162 |
print(f"Using seed: {seed}")
|
163 |
|
|
|
175 |
gen_start_time = time.time()
|
176 |
image_list = current_pipe(**pipe_kwargs, generator=torch.manual_seed(seed)).images
|
177 |
image = image_list[0]
|
|
|
178 |
gen_end_time = time.time()
|
179 |
results.append({"label": label, "image": image})
|
180 |
print(f"--- Finished Generation with {label} Model in {gen_end_time - gen_start_time:.2f} seconds ---")
|
|
|
183 |
|
184 |
except Exception as e:
|
185 |
print(f"Error during {label} model processing: {e}")
|
186 |
+
return None, {}, gr.update(value=f"Error processing {label} model: {e}", interactive=False), gr.update(choices=[], value=None), gr.update(interactive=True), gr.update(interactive=True)
|
|
|
187 |
|
|
|
188 |
|
189 |
if len(results) != len(model_configs):
|
190 |
+
return None, {}, gr.update(value="Failed to generate images for all model types.", interactive=False), gr.update(choices=[], value=None), gr.update(interactive=True), gr.update(interactive=True)
|
|
|
|
|
191 |
|
|
|
192 |
shuffled_results = results.copy()
|
193 |
random.shuffle(shuffled_results)
|
|
|
|
|
194 |
shuffled_data_for_gallery = [(res["image"], f"Image {i+1}") for i, res in enumerate(shuffled_results)]
|
|
|
|
|
195 |
correct_mapping = {i: res["label"] for i, res in enumerate(shuffled_results)}
|
196 |
print("Correct mapping (hidden):", correct_mapping)
|
197 |
|
198 |
+
return shuffled_data_for_gallery, correct_mapping, "Generation complete! Make your guess.", None, gr.update(interactive=True), gr.update(interactive=True)
|
199 |
|
|
|
|
|
200 |
|
|
|
|
|
201 |
def check_guess(user_guess, correct_mapping_state):
|
|
|
|
|
202 |
if not isinstance(correct_mapping_state, dict) or not correct_mapping_state:
|
203 |
return "Please generate images first (state is empty or invalid)."
|
|
|
204 |
if user_guess is None:
|
205 |
return "Please select which image you think is quantized."
|
206 |
|
|
|
207 |
quantized_image_index = -1
|
208 |
quantized_label_actual = ""
|
209 |
for index, label in correct_mapping_state.items():
|
210 |
+
if "Quantized" in label:
|
211 |
quantized_image_index = index
|
212 |
+
quantized_label_actual = label
|
213 |
break
|
|
|
214 |
if quantized_image_index == -1:
|
|
|
215 |
return "Error: Could not find the quantized image in the mapping data."
|
216 |
|
217 |
+
correct_guess_label = f"Image {quantized_image_index + 1}"
|
|
|
|
|
218 |
if user_guess == correct_guess_label:
|
219 |
feedback = f"Correct! {correct_guess_label} used the {quantized_label_actual} model."
|
220 |
else:
|
221 |
feedback = f"Incorrect. The quantized image ({quantized_label_actual}) was {correct_guess_label}."
|
|
|
222 |
return feedback
|
223 |
|
224 |
EXAMPLE_DIR = Path(__file__).parent / "examples"
|
|
|
226 |
{
|
227 |
"prompt": "A photorealistic portrait of an astronaut on Mars",
|
228 |
"files": ["astronauts_seed_6456306350371904162.png", "astronauts_bnb_8bit.png"],
|
229 |
+
"quantized_idx": 1,
|
230 |
+
"quant_method": "bnb 8-bit",
|
231 |
},
|
232 |
{
|
233 |
"prompt": "Water-color painting of a cat wearing sunglasses",
|
234 |
"files": ["watercolor_cat_bnb_8bit.png", "watercolor_cat_seed_14269059182221286790.png"],
|
235 |
"quantized_idx": 0,
|
236 |
+
"quant_method": "bnb 8-bit",
|
237 |
},
|
238 |
# {
|
239 |
# "prompt": "Neo-tokyo cyberpunk cityscape at night, rain-soaked streets, 8-K",
|
|
|
243 |
]
|
244 |
|
245 |
def load_example(idx):
|
|
|
246 |
ex = EXAMPLES[idx]
|
247 |
imgs = [Image.open(EXAMPLE_DIR / f) for f in ex["files"]]
|
248 |
gallery_items = [(img, f"Image {i+1}") for i, img in enumerate(imgs)]
|
249 |
+
mapping = {i: (f"Quantized {ex['quant_method']}" if i == ex["quantized_idx"] else "Original")
|
250 |
for i in range(2)}
|
251 |
return gallery_items, mapping, f"{ex['prompt']}"
|
252 |
|
253 |
+
def _accuracy_string(correct: int, attempts: int) -> tuple[str, float]:
|
254 |
+
if attempts:
|
255 |
+
pct = 100 * correct / attempts
|
256 |
+
return f"{pct:.1f}%", pct
|
257 |
+
return "N/A", -1.0
|
258 |
+
|
259 |
+
def _add_medals(user_rows):
|
260 |
+
MEDALS = {0: "🥇 ", 1: "🥈 ", 2: "🥉 "}
|
261 |
+
return [
|
262 |
+
[MEDALS.get(i, "") + row[0], *row[1:]]
|
263 |
+
for i, row in enumerate(user_rows)
|
264 |
+
]
|
265 |
+
|
266 |
+
def update_leaderboards_data():
|
267 |
+
agg = _load_agg_stats()
|
268 |
+
quant_rows = []
|
269 |
+
for method, stats in agg.items():
|
270 |
+
acc_str, acc_val = _accuracy_string(stats["correct"], stats["attempts"])
|
271 |
+
quant_rows.append([
|
272 |
+
method,
|
273 |
+
stats["correct"],
|
274 |
+
stats["attempts"],
|
275 |
+
acc_str
|
276 |
+
])
|
277 |
+
quant_rows.sort(key=lambda r: r[1]/r[2] if r[2] != 0 else 1e9)
|
278 |
+
|
279 |
+
user_stats = _load_user_stats()
|
280 |
+
user_rows = []
|
281 |
+
for user, st in user_stats.items():
|
282 |
+
acc_str, acc_val = _accuracy_string(st["total_correct"], st["total_attempts"])
|
283 |
+
user_rows.append([user, st["total_correct"], st["total_attempts"], acc_str])
|
284 |
+
user_rows.sort(key=lambda r: (-float(r[3].rstrip('%')) if r[3] != "N/A" else float('-inf'), -r[2]))
|
285 |
+
user_rows = _add_medals(user_rows)
|
286 |
+
|
287 |
+
return quant_rows, user_rows
|
288 |
+
|
289 |
with gr.Blocks(title="FLUX Quantization Challenge", theme=gr.themes.Soft()) as demo:
|
290 |
gr.Markdown("# FLUX Model Quantization Challenge")
|
291 |
+
with gr.Tabs():
|
292 |
+
with gr.TabItem("Challenge"):
|
293 |
+
gr.Markdown(
|
294 |
+
"Compare the original FLUX.1-dev (BF16) model against a quantized version (4-bit or 8-bit). "
|
295 |
+
"Enter a prompt, choose the quantization method, and generate two images. "
|
296 |
+
"The images will be shuffled, can you spot which one was quantized?"
|
297 |
+
)
|
298 |
+
|
299 |
+
gr.Markdown("### Examples")
|
300 |
+
ex_selector = gr.Radio(
|
301 |
+
choices=[f"Example {i+1}" for i in range(len(EXAMPLES))],
|
302 |
+
label="Choose an example prompt",
|
303 |
+
interactive=True,
|
304 |
+
)
|
305 |
+
gr.Markdown("### …or create your own comparison")
|
306 |
+
with gr.Row():
|
307 |
+
prompt_input = gr.Textbox(label="Enter Prompt", scale=3)
|
308 |
+
quantization_choice_radio = gr.Radio(
|
309 |
+
choices=["8-bit", "4-bit"],
|
310 |
+
label="Select Quantization",
|
311 |
+
value="8-bit",
|
312 |
+
scale=1
|
313 |
+
)
|
314 |
+
generate_button = gr.Button("Generate & Compare", variant="primary", scale=1)
|
315 |
+
|
316 |
+
output_gallery = gr.Gallery(
|
317 |
+
label="Generated Images",
|
318 |
+
columns=2,
|
319 |
+
height=606,
|
320 |
+
object_fit="contain",
|
321 |
+
allow_preview=True,
|
322 |
+
show_label=True,
|
323 |
+
)
|
324 |
+
|
325 |
+
gr.Markdown("### Which image used the selected quantization method?")
|
326 |
+
with gr.Row():
|
327 |
+
image1_btn = gr.Button("Image 1")
|
328 |
+
image2_btn = gr.Button("Image 2")
|
329 |
+
|
330 |
+
feedback_box = gr.Textbox(label="Feedback", interactive=False, lines=1)
|
331 |
+
|
332 |
+
with gr.Row():
|
333 |
+
session_score_box = gr.Textbox(label="Your accuracy this session", interactive=False)
|
334 |
+
|
335 |
+
with gr.Row(equal_height=False):
|
336 |
+
username_input = gr.Textbox(
|
337 |
+
label="Enter Your Name for Leaderboard",
|
338 |
+
placeholder="YourName",
|
339 |
+
visible=False,
|
340 |
+
interactive=True,
|
341 |
+
scale=2
|
342 |
+
)
|
343 |
+
add_score_button = gr.Button(
|
344 |
+
"Add My Score to Leaderboard",
|
345 |
+
visible=False,
|
346 |
+
variant="secondary",
|
347 |
+
scale=1
|
348 |
+
)
|
349 |
+
add_score_feedback = gr.Textbox(
|
350 |
+
label="Leaderboard Update",
|
351 |
+
visible=False,
|
352 |
+
interactive=False,
|
353 |
+
lines=1
|
354 |
+
)
|
355 |
+
|
356 |
+
correct_mapping_state = gr.State({})
|
357 |
+
session_stats_state = gr.State(
|
358 |
+
{"8-bit": {"attempts": 0, "correct": 0},
|
359 |
+
"4-bit": {"attempts": 0, "correct": 0}}
|
360 |
+
)
|
361 |
+
is_example_state = gr.State(False)
|
362 |
+
has_added_score_state = gr.State(False)
|
363 |
+
|
364 |
+
def _load_example(sel):
|
365 |
+
idx = int(sel.split()[-1]) - 1
|
366 |
+
gallery_items, mapping, prompt = load_example(idx)
|
367 |
+
quant_data, user_data = update_leaderboards_data()
|
368 |
+
return gallery_items, mapping, prompt, True, quant_data, user_data
|
369 |
+
|
370 |
+
ex_selector.change(
|
371 |
+
fn=_load_example,
|
372 |
+
inputs=ex_selector,
|
373 |
+
outputs=[output_gallery, correct_mapping_state, prompt_input, is_example_state, quant_df, user_df],
|
374 |
+
).then(
|
375 |
+
lambda: (gr.update(interactive=True), gr.update(interactive=True)),
|
376 |
+
outputs=[image1_btn, image2_btn],
|
377 |
+
)
|
378 |
+
|
379 |
+
generate_button.click(
|
380 |
+
fn=generate_images,
|
381 |
+
inputs=[prompt_input, quantization_choice_radio],
|
382 |
+
outputs=[output_gallery, correct_mapping_state]
|
383 |
+
).then(
|
384 |
+
lambda: (False, # for is_example_state
|
385 |
+
False, # for has_added_score_state
|
386 |
+
gr.update(visible=False, value="", interactive=True), # username_input reset
|
387 |
+
gr.update(visible=False), # add_score_button reset
|
388 |
+
gr.update(visible=False, value="")), # add_score_feedback reset
|
389 |
+
outputs=[is_example_state,
|
390 |
+
has_added_score_state,
|
391 |
+
username_input,
|
392 |
+
add_score_button,
|
393 |
+
add_score_feedback]
|
394 |
+
).then(
|
395 |
+
lambda: (gr.update(interactive=True),
|
396 |
+
gr.update(interactive=True),
|
397 |
+
""),
|
398 |
+
outputs=[image1_btn, image2_btn, feedback_box],
|
399 |
+
)
|
400 |
+
|
401 |
+
def choose(choice_string, mapping, session_stats, is_example, has_added_score_curr):
|
402 |
+
feedback = check_guess(choice_string, mapping)
|
403 |
+
|
404 |
+
quant_label = next(label for label in mapping.values() if "Quantized" in label)
|
405 |
+
quant_key = "8-bit" if "8-bit" in quant_label else "4-bit"
|
406 |
+
|
407 |
+
got_it_right = "Correct!" in feedback
|
408 |
+
|
409 |
+
sess = session_stats.copy()
|
410 |
+
if not is_example and not has_added_score_curr:
|
411 |
+
sess[quant_key]["attempts"] += 1
|
412 |
+
if got_it_right:
|
413 |
+
sess[quant_key]["correct"] += 1
|
414 |
+
session_stats = sess
|
415 |
+
|
416 |
+
AGG_STATS = _load_agg_stats()
|
417 |
+
AGG_STATS[quant_key]["attempts"] += 1
|
418 |
+
if got_it_right:
|
419 |
+
AGG_STATS[quant_key]["correct"] += 1
|
420 |
+
_save_agg_stats(AGG_STATS)
|
421 |
+
|
422 |
+
def _fmt(d):
|
423 |
+
a, c = d["attempts"], d["correct"]
|
424 |
+
pct = 100 * c / a if a else 0
|
425 |
+
return f"{c} / {a} ({pct:.1f}%)"
|
426 |
+
|
427 |
+
session_msg = ", ".join(
|
428 |
+
f"{k}: {_fmt(v)}" for k, v in sess.items()
|
429 |
+
)
|
430 |
+
current_agg_stats = _load_agg_stats()
|
431 |
+
global_msg = ", ".join(
|
432 |
+
f"{k}: {_fmt(v)}" for k, v in current_agg_stats.items()
|
433 |
+
)
|
434 |
+
|
435 |
+
username_input_update = gr.update(visible=False, interactive=True)
|
436 |
+
add_score_button_update = gr.update(visible=False)
|
437 |
+
# Keep existing feedback if score was already added and feedback is visible
|
438 |
+
current_feedback_text = add_score_feedback.value if hasattr(add_score_feedback, 'value') and add_score_feedback.value else ""
|
439 |
+
add_score_feedback_update = gr.update(visible=has_added_score_curr, value=current_feedback_text)
|
440 |
+
|
441 |
+
session_total_attempts = sum(stats["attempts"] for stats in sess.values())
|
442 |
+
|
443 |
+
if not is_example and not has_added_score_curr:
|
444 |
+
if session_total_attempts >= 1 : # Show button if more than 1 attempt
|
445 |
+
username_input_update = gr.update(visible=True, interactive=True)
|
446 |
+
add_score_button_update = gr.update(visible=True, interactive=True)
|
447 |
+
add_score_feedback_update = gr.update(visible=False, value="")
|
448 |
+
else: # Less than 1 attempts, keep hidden
|
449 |
+
username_input_update = gr.update(visible=False, value=username_input.value if hasattr(username_input, 'value') else "")
|
450 |
+
add_score_button_update = gr.update(visible=False)
|
451 |
+
add_score_feedback_update = gr.update(visible=False, value="")
|
452 |
+
elif has_added_score_curr:
|
453 |
+
username_input_update = gr.update(visible=True, interactive=False, value=username_input.value if hasattr(username_input, 'value') else "")
|
454 |
+
add_score_button_update = gr.update(visible=True, interactive=False)
|
455 |
+
add_score_feedback_update = gr.update(visible=True)
|
456 |
+
|
457 |
+
# disable the buttons so the user can't vote twice
|
458 |
+
quant_data, user_data = update_leaderboards_data() # Get updated leaderboard data
|
459 |
+
return (feedback,
|
460 |
+
gr.update(interactive=False),
|
461 |
+
gr.update(interactive=False),
|
462 |
+
session_msg,
|
463 |
+
session_stats,
|
464 |
+
quant_data,
|
465 |
+
user_data,
|
466 |
+
username_input_update,
|
467 |
+
add_score_button_update,
|
468 |
+
add_score_feedback_update)
|
469 |
+
|
470 |
+
|
471 |
+
image1_btn.click(
|
472 |
+
fn=lambda mapping, sess, is_ex, has_added: choose("Image 1", mapping, sess, is_ex, has_added),
|
473 |
+
inputs=[correct_mapping_state, session_stats_state, is_example_state, has_added_score_state],
|
474 |
+
outputs=[feedback_box, image1_btn, image2_btn,
|
475 |
+
session_score_box, session_stats_state,
|
476 |
+
quant_df, user_df,
|
477 |
+
username_input, add_score_button, add_score_feedback],
|
478 |
+
)
|
479 |
+
image2_btn.click(
|
480 |
+
fn=lambda mapping, sess, is_ex, has_added: choose("Image 2", mapping, sess, is_ex, has_added),
|
481 |
+
inputs=[correct_mapping_state, session_stats_state, is_example_state, has_added_score_state],
|
482 |
+
outputs=[feedback_box, image1_btn, image2_btn,
|
483 |
+
session_score_box, session_stats_state,
|
484 |
+
quant_df, user_df,
|
485 |
+
username_input, add_score_button, add_score_feedback],
|
486 |
+
)
|
487 |
+
|
488 |
+
def handle_add_score_to_leaderboard(username_str, current_session_stats_dict):
|
489 |
+
if not username_str or not username_str.strip():
|
490 |
+
return ("Username is required.", # Feedback for add_score_feedback
|
491 |
+
gr.update(interactive=True), # username_input
|
492 |
+
gr.update(interactive=True), # add_score_button
|
493 |
+
False, # has_added_score_state
|
494 |
+
None, None) # quant_df, user_df
|
495 |
+
|
496 |
+
user_stats = _load_user_stats()
|
497 |
+
user_key = username_str.strip()
|
498 |
+
|
499 |
+
session_total_correct = sum(stats["correct"] for stats in current_session_stats_dict.values())
|
500 |
+
session_total_attempts = sum(stats["attempts"] for stats in current_session_stats_dict.values())
|
501 |
+
|
502 |
+
if session_total_attempts == 0:
|
503 |
+
return ("No attempts made in this session to add to leaderboard.",
|
504 |
+
gr.update(interactive=True),
|
505 |
+
gr.update(interactive=True),
|
506 |
+
False, None, None)
|
507 |
+
|
508 |
+
if user_key in user_stats:
|
509 |
+
user_stats[user_key]["total_correct"] += session_total_correct
|
510 |
+
user_stats[user_key]["total_attempts"] += session_total_attempts
|
511 |
+
else:
|
512 |
+
user_stats[user_key] = {
|
513 |
+
"total_correct": session_total_correct,
|
514 |
+
"total_attempts": session_total_attempts
|
515 |
+
}
|
516 |
+
_save_user_stats(user_stats)
|
517 |
+
|
518 |
+
new_quant_data, new_user_data = update_leaderboards_data()
|
519 |
+
feedback_msg = f"Score for '{user_key}' submitted to leaderboard!"
|
520 |
+
return (feedback_msg, # To add_score_feedback
|
521 |
+
gr.update(interactive=False), # username_input
|
522 |
+
gr.update(interactive=False), # add_score_button
|
523 |
+
True, # has_added_score_state (set to true)
|
524 |
+
new_quant_data, # To quant_df
|
525 |
+
new_user_data) # To user_df
|
526 |
+
|
527 |
+
add_score_button.click(
|
528 |
+
fn=handle_add_score_to_leaderboard,
|
529 |
+
inputs=[username_input, session_stats_state],
|
530 |
+
outputs=[add_score_feedback, username_input, add_score_button, has_added_score_state, quant_df, user_df]
|
531 |
+
)
|
532 |
+
with gr.TabItem("Leaderboard"):
|
533 |
+
gr.Markdown("## Quantization Method Leaderboard *(Lower % ⇒ harder to detect)*")
|
534 |
+
quant_df = gr.DataFrame(
|
535 |
+
headers=["Method", "Correct Guesses", "Total Attempts", "Detectability %"],
|
536 |
+
interactive=False, col_count=(4, "fixed")
|
537 |
+
)
|
538 |
+
gr.Markdown("## User Leaderboard *(Higher % ⇒ better spotter)*")
|
539 |
+
user_df = gr.DataFrame(
|
540 |
+
headers=["User", "Correct Guesses", "Total Attempts", "Accuracy %"],
|
541 |
+
interactive=False, col_count=(4, "fixed")
|
542 |
+
)
|
543 |
+
demo.load(update_leaderboards_data, outputs=[quant_df, user_df])
|
544 |
|
545 |
if __name__ == "__main__":
|
546 |
demo.launch(share=True)
|