Model Card for ColorizeNet

This model is a ControlNet training to perform image colorization from black and white images.

Model Details

Model Description

ColorizeNet is an image colorization model based on ControlNet, trained using the pre-trained Stable Diffusion model version 2.1 proposed by Stability AI.

Finetuned from model : [https://huggingface.co./stabilityai/stable-diffusion-2-1]

Model Sources [optional]

Repository: [https://github.com/rensortino/ColorizeNet]

Usage

Training Data

The model has been trained on COCO, using all the images in the dataset and converting them to grayscale to use them to condition the ControlNet

[https://huggingface.co./datasets/detection-datasets/coco]

Run the model

Instantiate the model and load its configuration and weights

import random

import cv2
import einops
import numpy as np
import torch
from pytorch_lightning import seed_everything

from utils.data import HWC3, apply_color, resize_image
from utils.ddim import DDIMSampler
from utils.model import create_model, load_state_dict

model = create_model('./models/cldm_v21.yaml').cpu()
model.load_state_dict(load_state_dict(
    'lightning_logs/version_6/checkpoints/colorizenet-sd21.ckpt', location='cuda'))
model = model.cuda()
ddim_sampler = DDIMSampler(model)

Read the image to be colorized

input_image = cv2.imread("sample_data/sample1_bw.jpg")
input_image = HWC3(input_image)
img = resize_image(input_image, resolution=512)
H, W, C = img.shape

num_samples = 1
control = torch.from_numpy(img.copy()).float().cuda() / 255.0
control = torch.stack([control for _ in range(num_samples)], dim=0)
control = einops.rearrange(control, 'b h w c -> b c h w').clone()

Prepare the input and parameters of the model

seed = 1294574436
seed_everything(seed)
prompt = "Colorize this image"
n_prompt = ""
guess_mode = False
strength = 1.0
eta = 0.0
ddim_steps = 20
scale = 9.0

cond = {"c_concat": [control], "c_crossattn": [
    model.get_learned_conditioning([prompt] * num_samples)]}
un_cond = {"c_concat": None if guess_mode else [control], "c_crossattn": [
    model.get_learned_conditioning([n_prompt] * num_samples)]}
shape = (4, H // 8, W // 8)

model.control_scales = [strength * (0.825 ** float(12 - i)) for i in range(13)] if guess_mode else (
    [strength] * 13)

Sample and post-process the results

samples, intermediates = ddim_sampler.sample(ddim_steps, num_samples,
                                             shape, cond, verbose=False, eta=eta,
                                             unconditional_guidance_scale=scale,
                                             unconditional_conditioning=un_cond)

x_samples = model.decode_first_stage(samples)
x_samples = (einops.rearrange(x_samples, 'b c h w -> b h w c')
             * 127.5 + 127.5).cpu().numpy().clip(0, 255).astype(np.uint8)

results = [x_samples[i] for i in range(num_samples)]
colored_results = [apply_color(img, result) for result in results]