This demo is licensed under a Creative Commons Attribution-ShareAlike 4.0 International License.
' DESCRIPTION += '\nNote: it takes about 5 minutes for 250 iterations to generate the final GIF. For faster inference without waiting in queue, you can
"
warnings.filterwarnings("ignore")
pydiffvg.set_print_timing(False)
gamma = 1.0
def set_config(semantic_concept, word, script, prompt_suffix, font_name, num_steps, seed, is_seed_rand, dist_loss_weight, pixel_dist_kernel_blur, pixel_dist_sigma, angeles_w):
cfg_d = edict()
cfg_d.config = "code/config/base.yaml"
cfg_d.experiment = "default"
with open(cfg_d.config, 'r') as f:
cfg_full = yaml.load(f, Loader=yaml.FullLoader)
cfg_key = cfg_d.experiment
cfgs = [cfg_d]
while cfg_key:
cfgs.append(cfg_full[cfg_key])
cfg_key = cfgs[-1].get('parent_config', 'baseline')
cfg = edict()
for options in reversed(cfgs):
update(cfg, options)
del cfgs
cfg.semantic_concept = semantic_concept
cfg.prompt_suffix = prompt_suffix
cfg.word = word
cfg.optimized_letter = word
cfg.script = script.lower()
if cfg.script == "simplified chinese":
cfg.script = "chinese"
script_path = f"code/data/fonts/{cfg.script}"
cfg.font = [x for x in os.listdir(script_path) if "ttf" in x][0][:-4]
if is_seed_rand == "Random Seed":
cfg.seed = np.random.randint(10000)
else:
cfg.seed = int(seed)
cfg.num_iter = num_steps
cfg.batch_size = 1
cfg.loss.tone.dist_loss_weight = int(dist_loss_weight)
cfg.loss.tone.pixel_dist_kernel_blur = int(pixel_dist_kernel_blur)
cfg.loss.tone.pixel_dist_sigma = int(pixel_dist_sigma)
cfg.loss.conformal.angeles_w = angeles_w
cfg.caption = f"a {cfg.semantic_concept}. {cfg.prompt_suffix}"
cfg.log_dir = f"{cfg.script}"
if cfg.optimized_letter in cfg.word:
cfg.optimized_letter = cfg.optimized_letter
else:
raise gr.Error(f'letter should be in word')
# if ' ' in cfg.word:
# cfg.optimized_letter = cfg.optimized_letter.replace(' ', '_')
cfg.letter = f"{cfg.font}_{cfg.optimized_letter}_scaled"
cfg.target = f"code/data/init/{cfg.letter}"
if ' ' in cfg.target:
cfg.target = cfg.target.replace(' ', '_')
# set experiment dir
signature = f"{cfg.word}_{cfg.semantic_concept}_{cfg.seed}"
cfg.experiment_dir = osp.join(cfg.log_dir, cfg.font, signature)
configfile = osp.join(cfg.experiment_dir, 'config.yaml')
# create experiment dir and save config
check_and_create_dir(configfile)
with open(osp.join(configfile), 'w') as f:
yaml.dump(edict_2_dict(cfg), f)
if cfg.seed is not None:
random.seed(cfg.seed)
npr.seed(cfg.seed)
torch.manual_seed(cfg.seed)
torch.backends.cudnn.benchmark = False
else:
assert False
return cfg
def init_shapes(svg_path, trainable: Mapping[str, bool]):
svg = f'{svg_path}.svg'
canvas_width, canvas_height, shapes_init, shape_groups_init = pydiffvg.svg_to_scene(svg)
parameters = edict()
# path points
if trainable.point:
parameters.point = []
for path in shapes_init:
path.points.requires_grad = True
parameters.point.append(path.points)
return shapes_init, shape_groups_init, parameters
def run_main_ex(word, semantic_concept, script, num_steps, seed):
prompt_suffix = "minimal flat 2d vector. lineal color. trending on artstation"
is_seed_rand = "Use Set Value"
return list(next(run_main_app(semantic_concept, word, script, prompt_suffix, font_name, num_steps, seed, is_seed_rand, 100, 201, 30, 0.5, 0)))
def run_main_app(semantic_concept, word, script, prompt_suffix, font_name, num_steps, seed, is_seed_rand, dist_loss_weight, pixel_dist_kernel_blur, pixel_dist_sigma, angeles_w, example=0):
cfg = set_config(semantic_concept, word, script, prompt_suffix, font_name, num_steps, seed, is_seed_rand, dist_loss_weight, pixel_dist_kernel_blur, pixel_dist_sigma, angeles_w)
pydiffvg.set_use_gpu(torch.cuda.is_available())
print("preprocessing")
preprocess(cfg.font, cfg.word, cfg.optimized_letter, cfg.script, cfg.level_of_cc)
filename_init = os.path.join("code/data/init/", f"{cfg.font}_{cfg.word}_scaled.svg").replace(" ", "_")
if not example:
yield gr.update(value=filename_init,visible=True),gr.update(visible=True, label='Initializing'),gr.update(visible=False),gr.update(value=cfg.caption,visible=True),gr.update(value=cfg.seed,visible=True)
sds_loss = SDSLoss(cfg, device, model)
h, w = cfg.render_size, cfg.render_size
data_augs = get_data_augs(cfg.cut_size)
render = pydiffvg.RenderFunction.apply
# initialize shape
print('initializing shape')
shapes, shape_groups, parameters = init_shapes(svg_path=cfg.target, trainable=cfg.trainable)
scene_args = pydiffvg.RenderFunction.serialize_scene(w, h, shapes, shape_groups)
img_init = render(w, h, 2, 2, 0, None, *scene_args)
img_init = img_init[:, :, 3:4] * img_init[:, :, :3] + \
torch.ones(img_init.shape[0], img_init.shape[1], 3, device=device) * (1 - img_init[:, :, 3:4])
img_init = img_init[:, :, :3]
tone_loss = ToneLoss(cfg)
tone_loss.set_image_init(img_init)
num_iter = cfg.num_iter
pg = [{'params': parameters["point"], 'lr': cfg.lr_base["point"]}]
optim = torch.optim.Adam(pg, betas=(0.9, 0.9), eps=1e-6)
conformal_loss = ConformalLoss(parameters, device, cfg.optimized_letter, shape_groups)
lr_lambda = lambda step: learning_rate_decay(step, cfg.lr.lr_init, cfg.lr.lr_final, num_iter,
lr_delay_steps=cfg.lr.lr_delay_steps,
lr_delay_mult=cfg.lr.lr_delay_mult) / cfg.lr.lr_init
scheduler = LambdaLR(optim, lr_lambda=lr_lambda, last_epoch=-1) # lr.base * lrlambda_f
print("start training")
# training loop
t_range = tqdm(range(num_iter))
gif_frames = []
skip = 5
for step in t_range:
optim.zero_grad()
# render image
scene_args = pydiffvg.RenderFunction.serialize_scene(w, h, shapes, shape_groups)
img = render(w, h, 2, 2, step, None, *scene_args)
# compose image with white background
img = img[:, :, 3:4] * img[:, :, :3] + torch.ones(img.shape[0], img.shape[1], 3, device=device) * (
1 - img[:, :, 3:4])
img = img[:, :, :3]
if step % skip == 0:
gif_frames += [img.detach().cpu().numpy()*255]
filename = os.path.join(cfg.experiment_dir, "video-svg", f"iter{step:04d}.svg")
check_and_create_dir(filename)
save_svg.save_svg(filename, w, h, shapes, shape_groups)
if not example:
yield gr.update(visible=True),gr.update(value=filename, label=f'iters: {step} / {num_iter}', visible=True),gr.update(visible=False),gr.update(value=cfg.caption,visible=True),gr.update(value=cfg.seed,visible=True)
x = img.unsqueeze(0).permute(0, 3, 1, 2) # HWC -> NCHW
x = x.repeat(cfg.batch_size, 1, 1, 1)
x_aug = data_augs.forward(x)
# compute diffusion loss per pixel
loss = sds_loss(x_aug)
tone_loss_res = tone_loss(x, step)
loss = loss + tone_loss_res
loss_angles = conformal_loss()
loss_angles = cfg.loss.conformal.angeles_w * loss_angles
loss = loss + loss_angles
loss.backward()
optim.step()
scheduler.step()
filename = os.path.join(
cfg.experiment_dir, "output-svg", "output.svg")
check_and_create_dir(filename)
save_svg.save_svg(
filename, w, h, shapes, shape_groups)
# combine_word(cfg.word, cfg.optimized_letter, cfg.font, cfg.experiment_dir, device)
filename = os.path.join(cfg.experiment_dir, "final.gif")
# writer = imageio.get_writer(filename, fps=20)
# for frame in gif_frames: writer.append_data(frame)
# writer.close()
imageio.mimsave(filename, np.array(gif_frames).astype(np.uint8))
yield gr.update(value=filename_init,visible=True),gr.update(visible=False),gr.update(value=filename,visible=True),gr.update(value=cfg.caption,visible=True),gr.update(value=cfg.seed,visible=True)
with gr.Blocks() as demo:
gr.HTML(TITLE)
gr.Markdown(DESCRIPTION)
with gr.Row():
with gr.Column():
word = gr.Text(
label='Text',
max_lines=1,
placeholder=
'Enter text. For example: قطة|猫|γάτα|кошка|பூனை|Cat'
)
semantic_concept = gr.Text(
label='Concept',
max_lines=1,
placeholder=
'Enter a semantic concept that you want your text to morph into (in English). For example: cat'
)
script = gr.Dropdown(
["Arabic", "Simplified Chinese", "Cyrillic", "Greek", "Latin", "Tamil"],
value="Arabic",
label="Font Script"
)
prompt_suffix = gr.Text(
label='Prompt Suffix',
max_lines=1,
value="minimal flat 2d vector. lineal color. trending on artstation"
)
with gr.Row():
with gr.Accordion("Advanced Parameters", open=False, visible=True):
with gr.Row():
is_seed_rand = gr.Radio(["Random Seed", "Use Set Value"], label="Use Random Seed", value="Random Seed")
seed = gr.Number(
label='Seed (Set Value)',
value=42
)
angeles_w = gr.Number(
label='ACAP Deformation Loss Weight',
value=0.5
)
dist_loss_weight = gr.Number(
label='Tone Loss: dist_loss_weight',
value=100
)
pixel_dist_kernel_blur = gr.Number(
label='Tone Loss: pixel_dist_kernel_blur',
value=201
)
pixel_dist_sigma = gr.Number(
label='Tone Loss: pixel_dist_sigma',
value=30
)
num_steps = gr.Slider(label='Optimization Iterations',
minimum=0,
maximum=500,
step=10,
value=250)
font_name = gr.Text(value=None,visible=False,label="Font Name")
def on_select(evt: gr.SelectData):
return evt.value
font_name.value = "ArefRuqaa"
run = gr.Button('Generate')
with gr.Column():
with gr.Row():
prompt = gr.Text(
label='Prompt',
visible=False,
max_lines=1,
interactive=False,
)
seed_value = gr.Text(
label='Seed Used',
visible=False,
max_lines=1,
interactive=False,
)
result0 = gr.Image(type="filepath", label="Initial Word").style(height=250)
result1 = gr.Image(type="filepath", label="Optimization Process").style(height=300)
result2 = gr.Image(type="filepath", label="Final GIF",visible=False).style(height=300)
with gr.Row():
# examples
examples = [
["قطة", "Cat", "Arabic", 250, 42],
["猫", "Cat", "Simplified Chinese", 250, 42],
["γάτα", "Cat", "Greek", 250, 42],
["кошка", "Cat", "Cyrillic", 250, 42],
["பூனை", "Cat", "Tamil", 250, 42],
]
demo.queue(max_size=10, concurrency_count=2)
gr.Examples(examples=examples,
inputs=[
word,
semantic_concept,
script,
num_steps,
seed
],
outputs=[
result0,
result1,
result2
],
fn=run_main_ex,
cache_examples=True)
# inputs
inputs = [
semantic_concept,
word,
script,
prompt_suffix,
font_name,
num_steps,
seed,
is_seed_rand,
dist_loss_weight,
pixel_dist_kernel_blur,
pixel_dist_sigma,
angeles_w
]
outputs = [
result0,
result1,
result2,
prompt,
seed_value
]
run.click(fn=run_main_app, inputs=inputs, outputs=outputs, queue=True)
demo.launch(share=False)