Can it colorize big resolution images?

[FurkanGozukara]

i plan to put some time into making a tutorial for this

but before spending time i want to learn this

it is

parser.add_argument("--image-size", default=64, dest="img_size", type=int)

so can we colorize like 1024x1024 images as well or only 64px?

i am not interested in training

@ErwannMillon

[saulchiu]

i plan to put some time into making a tutorial for this

but before spending time i want to learn this

it is

parser.add_argument("--image-size", default=64, dest="img_size", type=int)

so can we colorize like 1024x1024 images as well or only 64px?

i am not interested in training

@ErwannMillon

Hi, sorry to bother you, have you solve this problem now? Recently I face the same problem too: when training for colorization for dataset with 256*256 resolusion, the loss is hard to decrease.

[XTY-starstar]

我打算花一些时间为此制作一个教程
but before spending time i want to learn this
it is

parser.add_argument("--image-size", default=64, dest="img_size", type=int)

so can we colorize like 1024x1024 images as well or only 64px?
i am not interested in training
@ErwannMillon

Hi, sorry to bother you, have you solve this problem now? Recently I face the same problem too: when training for colorization for dataset with 256*256 resolusion, the loss is hard to decrease.

Hello bro, sorry to bother you, have you solve this problem now? Recently I face the same problem tooooooo

[saulchiu]

我打算花一些时间为此制作一个教程
but before spending time i want to learn this
it is

parser.add_argument("--image-size", default=64, dest="img_size", type=int)

so can we colorize like 1024x1024 images as well or only 64px?
i am not interested in training
@ErwannMillon

Hi, sorry to bother you, have you solve this problem now? Recently I face the same problem too: when training for colorization for dataset with 256*256 resolusion, the loss is hard to decrease.

Hello bro, sorry to bother you, have you solve this problem now? Recently I face the same problem tooooooo

nope, i change another way to do colorization. Here is the process:

    x_lab = rgb_tensor_to_lab_tensor(x_rgb)
    x_l, _ = split_lab_channels(x_lab)
    l_rgb = x_l.repeat(1, 3, 1, 1).to(x_rgb)
    x_t = l_rgb.clone()
    col_list.append(l_rgb.cpu().clone())

    distance = int(t / loop)
    decreasing_list = [t - i * distance for i in range(loop)]
    print(decreasing_list)
    for i in tqdm(range(0, loop)):
        x_t = diffusion.q_sample(x_t, (torch.ones(size=(l_rgb.shape[0],), device=device) * decreasing_list[i]).to(torch.int64))
        for j in reversed(range(0, decreasing_list[i])):
            x_t_m_1 = diffusion.p_sample(x_t, torch.tensor([j], device=device))
            x_t = x_t_m_1
        x_t = rgb_tensor_to_lab_tensor(x_t)
        _, x_ab = split_lab_channels(x_t)
        x_t = torch.cat((x_l, x_ab), dim=1)
        x_t = lab_tensor_to_rgb_tensor(x_t).to(device)
        col_list.append(x_t.cpu().clone())

and the visulization:
original ---- gray ---- colorized

hope it can help u!

[XTY-starstar]

Hello bro, sorry to bother you, have you solve this problem now? Recently I face the same problem tooooooo

nope, i change another way to do colorization. Here is the process:

x_lab = rgb_tensor_to_lab_tensor(x_rgb)
x_l, _ = split_lab_channels(x_lab)
l_rgb = x_l.repeat(1, 3, 1, 1).to(x_rgb)
x_t = l_rgb.clone()
col_list.append(l_rgb.cpu().clone())

distance = int(t / loop)
decreasing_list = [t - i * distance for i in range(loop)]
print(decreasing_list)
for i in tqdm(range(0, loop)):
    x_t = diffusion.q_sample(x_t, (torch.ones(size=(l_rgb.shape[0],), device=device) * decreasing_list[i]).to(torch.int64))
    for j in reversed(range(0, decreasing_list[i])):
        x_t_m_1 = diffusion.p_sample(x_t, torch.tensor([j], device=device))
        x_t = x_t_m_1
    x_t = rgb_tensor_to_lab_tensor(x_t)
    _, x_ab = split_lab_channels(x_t)
    x_t = torch.cat((x_l, x_ab), dim=1)
    x_t = lab_tensor_to_rgb_tensor(x_t).to(device)
    col_list.append(x_t.cpu().clone())

and the visulization: original ---- gray ---- colorized hope it can help u!

Impressive work! Could you please provide a more detailed version of the code, including q_sample、p_sample and so on.
@chengyiqiu1121

[saulchiu]

Hello bro, sorry to bother you, have you solve this problem now? Recently I face the same problem tooooooo

nope, i change another way to do colorization. Here is the process:

x_lab = rgb_tensor_to_lab_tensor(x_rgb)
x_l, _ = split_lab_channels(x_lab)
l_rgb = x_l.repeat(1, 3, 1, 1).to(x_rgb)
x_t = l_rgb.clone()
col_list.append(l_rgb.cpu().clone())

distance = int(t / loop)
decreasing_list = [t - i * distance for i in range(loop)]
print(decreasing_list)
for i in tqdm(range(0, loop)):
    x_t = diffusion.q_sample(x_t, (torch.ones(size=(l_rgb.shape[0],), device=device) * decreasing_list[i]).to(torch.int64))
    for j in reversed(range(0, decreasing_list[i])):
        x_t_m_1 = diffusion.p_sample(x_t, torch.tensor([j], device=device))
        x_t = x_t_m_1
    x_t = rgb_tensor_to_lab_tensor(x_t)
    _, x_ab = split_lab_channels(x_t)
    x_t = torch.cat((x_l, x_ab), dim=1)
    x_t = lab_tensor_to_rgb_tensor(x_t).to(device)
    col_list.append(x_t.cpu().clone())

and the visulization: original ---- gray ---- colorized hope it can help u!

Impressive work! Could you please provide a more detailed version of the code, including q_sample、p_sample and so on. @chengyiqiu1121
Thanks! The APIs (q_sample and p_sample...) are the same as this opensource repository, you can refer to this repo to get more informations. Besides, i will open my repository in the future.