Spaces:
Runtime error
Runtime error
erwold
commited on
Commit
·
0ded2d6
1
Parent(s):
2645f74
Initial Commit
Browse files
app.py
CHANGED
|
@@ -7,9 +7,20 @@ from diffusers import AutoencoderKL, FlowMatchEulerDiscreteScheduler
|
|
| 7 |
from flux.transformer_flux import FluxTransformer2DModel
|
| 8 |
from flux.pipeline_flux_chameleon import FluxPipeline
|
| 9 |
import torch.nn as nn
|
|
|
|
| 10 |
|
| 11 |
MODEL_ID = "Djrango/Qwen2vl-Flux"
|
| 12 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 13 |
class Qwen2Connector(nn.Module):
|
| 14 |
def __init__(self, input_dim=3584, output_dim=4096):
|
| 15 |
super().__init__()
|
|
@@ -88,6 +99,23 @@ class FluxInterface:
|
|
| 88 |
text_encoder=text_encoder,
|
| 89 |
tokenizer=tokenizer,
|
| 90 |
)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 91 |
|
| 92 |
# [Previous methods remain unchanged...]
|
| 93 |
def process_image(self, image):
|
|
@@ -109,8 +137,8 @@ class FluxInterface:
|
|
| 109 |
image_hidden_state = self.models['connector'](image_hidden_state)
|
| 110 |
|
| 111 |
return image_hidden_state, image_grid_thw
|
| 112 |
-
|
| 113 |
-
def
|
| 114 |
"""Compute T5 embeddings for text prompt"""
|
| 115 |
if prompt == "":
|
| 116 |
return None
|
|
@@ -129,13 +157,36 @@ class FluxInterface:
|
|
| 129 |
|
| 130 |
return prompt_embeds
|
| 131 |
|
| 132 |
-
def
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 133 |
try:
|
| 134 |
if seed is not None:
|
| 135 |
torch.manual_seed(seed)
|
| 136 |
|
| 137 |
self.load_models()
|
| 138 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 139 |
# Process input image
|
| 140 |
input_image = self.resize_image(input_image)
|
| 141 |
qwen2_hidden_state, image_grid_thw = self.process_image(input_image)
|
|
@@ -151,6 +202,8 @@ class FluxInterface:
|
|
| 151 |
t5_prompt_embeds=t5_prompt_embeds.repeat(num_images, 1, 1) if t5_prompt_embeds is not None else None,
|
| 152 |
num_inference_steps=num_inference_steps,
|
| 153 |
guidance_scale=guidance_scale,
|
|
|
|
|
|
|
| 154 |
).images
|
| 155 |
|
| 156 |
return output_images
|
|
@@ -212,7 +265,7 @@ with gr.Blocks(
|
|
| 212 |
with gr.Group():
|
| 213 |
prompt = gr.Textbox(
|
| 214 |
label="Text Prompt (Optional)",
|
| 215 |
-
placeholder="
|
| 216 |
lines=3
|
| 217 |
)
|
| 218 |
|
|
@@ -249,6 +302,12 @@ with gr.Blocks(
|
|
| 249 |
precision=0,
|
| 250 |
info="Set for reproducible results"
|
| 251 |
)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 252 |
|
| 253 |
submit_btn = gr.Button(
|
| 254 |
"🎨 Generate Variations",
|
|
@@ -288,7 +347,8 @@ with gr.Blocks(
|
|
| 288 |
guidance,
|
| 289 |
steps,
|
| 290 |
num_images,
|
| 291 |
-
seed
|
|
|
|
| 292 |
],
|
| 293 |
outputs=output_gallery,
|
| 294 |
show_progress="minimal"
|
|
|
|
| 7 |
from flux.transformer_flux import FluxTransformer2DModel
|
| 8 |
from flux.pipeline_flux_chameleon import FluxPipeline
|
| 9 |
import torch.nn as nn
|
| 10 |
+
import math
|
| 11 |
|
| 12 |
MODEL_ID = "Djrango/Qwen2vl-Flux"
|
| 13 |
|
| 14 |
+
# Add aspect ratio options
|
| 15 |
+
ASPECT_RATIOS = {
|
| 16 |
+
"1:1": (1024, 1024),
|
| 17 |
+
"16:9": (1344, 768),
|
| 18 |
+
"9:16": (768, 1344),
|
| 19 |
+
"2.4:1": (1536, 640),
|
| 20 |
+
"3:4": (896, 1152),
|
| 21 |
+
"4:3": (1152, 896),
|
| 22 |
+
}
|
| 23 |
+
|
| 24 |
class Qwen2Connector(nn.Module):
|
| 25 |
def __init__(self, input_dim=3584, output_dim=4096):
|
| 26 |
super().__init__()
|
|
|
|
| 99 |
text_encoder=text_encoder,
|
| 100 |
tokenizer=tokenizer,
|
| 101 |
)
|
| 102 |
+
|
| 103 |
+
def resize_image(self, img, max_pixels=1050000):
|
| 104 |
+
if not isinstance(img, Image.Image):
|
| 105 |
+
img = Image.fromarray(img)
|
| 106 |
+
|
| 107 |
+
width, height = img.size
|
| 108 |
+
num_pixels = width * height
|
| 109 |
+
|
| 110 |
+
if num_pixels > max_pixels:
|
| 111 |
+
scale = math.sqrt(max_pixels / num_pixels)
|
| 112 |
+
new_width = int(width * scale)
|
| 113 |
+
new_height = int(height * scale)
|
| 114 |
+
new_width = new_width - (new_width % 8)
|
| 115 |
+
new_height = new_height - (new_height % 8)
|
| 116 |
+
img = img.resize((new_width, new_height), Image.LANCZOS)
|
| 117 |
+
|
| 118 |
+
return img
|
| 119 |
|
| 120 |
# [Previous methods remain unchanged...]
|
| 121 |
def process_image(self, image):
|
|
|
|
| 137 |
image_hidden_state = self.models['connector'](image_hidden_state)
|
| 138 |
|
| 139 |
return image_hidden_state, image_grid_thw
|
| 140 |
+
|
| 141 |
+
def compute_t5_text_embeddings(self, prompt):
|
| 142 |
"""Compute T5 embeddings for text prompt"""
|
| 143 |
if prompt == "":
|
| 144 |
return None
|
|
|
|
| 157 |
|
| 158 |
return prompt_embeds
|
| 159 |
|
| 160 |
+
def compute_text_embeddings(self, prompt=""):
|
| 161 |
+
with torch.no_grad():
|
| 162 |
+
text_inputs = self.models['tokenizer'](
|
| 163 |
+
prompt,
|
| 164 |
+
padding="max_length",
|
| 165 |
+
max_length=77,
|
| 166 |
+
truncation=True,
|
| 167 |
+
return_tensors="pt"
|
| 168 |
+
).to(self.device)
|
| 169 |
+
|
| 170 |
+
prompt_embeds = self.models['text_encoder'](
|
| 171 |
+
text_inputs.input_ids,
|
| 172 |
+
output_hidden_states=False
|
| 173 |
+
)
|
| 174 |
+
pooled_prompt_embeds = prompt_embeds.pooler_output.to(self.dtype)
|
| 175 |
+
|
| 176 |
+
return pooled_prompt_embeds
|
| 177 |
+
|
| 178 |
+
def generate(self, input_image, prompt="", guidance_scale=3.5, num_inference_steps=28, num_images=2, seed=None, aspect_ratio="1:1"):
|
| 179 |
try:
|
| 180 |
if seed is not None:
|
| 181 |
torch.manual_seed(seed)
|
| 182 |
|
| 183 |
self.load_models()
|
| 184 |
|
| 185 |
+
# Get dimensions from aspect ratio
|
| 186 |
+
if aspect_ratio not in ASPECT_RATIOS:
|
| 187 |
+
raise ValueError(f"Invalid aspect ratio. Choose from {list(ASPECT_RATIOS.keys())}")
|
| 188 |
+
width, height = ASPECT_RATIOS[aspect_ratio]
|
| 189 |
+
|
| 190 |
# Process input image
|
| 191 |
input_image = self.resize_image(input_image)
|
| 192 |
qwen2_hidden_state, image_grid_thw = self.process_image(input_image)
|
|
|
|
| 202 |
t5_prompt_embeds=t5_prompt_embeds.repeat(num_images, 1, 1) if t5_prompt_embeds is not None else None,
|
| 203 |
num_inference_steps=num_inference_steps,
|
| 204 |
guidance_scale=guidance_scale,
|
| 205 |
+
height=height,
|
| 206 |
+
width=width,
|
| 207 |
).images
|
| 208 |
|
| 209 |
return output_images
|
|
|
|
| 265 |
with gr.Group():
|
| 266 |
prompt = gr.Textbox(
|
| 267 |
label="Text Prompt (Optional)",
|
| 268 |
+
placeholder="As Long As Possible...",
|
| 269 |
lines=3
|
| 270 |
)
|
| 271 |
|
|
|
|
| 302 |
precision=0,
|
| 303 |
info="Set for reproducible results"
|
| 304 |
)
|
| 305 |
+
aspect_ratio = gr.Radio(
|
| 306 |
+
label="Aspect Ratio",
|
| 307 |
+
choices=["1:1", "16:9", "9:16", "2.4:1", "3:4", "4:3"],
|
| 308 |
+
value="1:1",
|
| 309 |
+
info="Choose aspect ratio for generated images"
|
| 310 |
+
)
|
| 311 |
|
| 312 |
submit_btn = gr.Button(
|
| 313 |
"🎨 Generate Variations",
|
|
|
|
| 347 |
guidance,
|
| 348 |
steps,
|
| 349 |
num_images,
|
| 350 |
+
seed,
|
| 351 |
+
aspect_ratio
|
| 352 |
],
|
| 353 |
outputs=output_gallery,
|
| 354 |
show_progress="minimal"
|