git clone https://github.com/liming-ai/ControlNet_Plus_Plus.git
pip3 install -r requirements.txt
pip3 install -U openmim
mim install mmengine
mim install "mmcv==2.1.0"
pip3 install "mmsegmentation>=1.0.0"
pip3 install mmdet

All the organized data has been put on Huggingface and will be automatically downloaded during training or evaluation. You can preview it in advance to check the data samples and disk space occupied with following links.

By default, our training is based on 8 A100-80G GPUs. If your computational resources are insufficient for training, you may need to reduce the batch size and increase gradient accumulation at the same time, and we have not observed any performance degradation. Reducing the training resolution will result in performance degradation.

ControlNet V1.1 Seg is trained on both ADE20K and COCOStuff, and these two datasets have different masks. To this end, we first perform normal model fine-tuning on each dataset, and then perform reward fine-tuning.

# Please refer to the reward script for details
bash train/reward_ade20k.sh
bash train/reward_cocostuff.sh

We can directly perform reward fine-tuning.

bash train/reward_canny.sh
bash train/reward_depth.sh
bash train/reward_hed.sh
bash train/reward_linedrawing.sh

Please refer to the core code here, in summary:

# Predict the single-step denoised latents
pred_original_sample = [
    noise_scheduler.step(noise, t, noisy_latent).pred_original_sample.to(weight_dtype) \
        for (noise, t, noisy_latent) in zip(model_pred, timesteps, noisy_latents)
]
pred_original_sample = torch.stack(pred_original_sample)

# Map the denoised latents into RGB images
pred_original_sample = 1 / vae.config.scaling_factor * pred_original_sample
image = vae.decode(pred_original_sample.to(weight_dtype)).sample
image = (image / 2 + 0.5).clamp(0, 1)

# The normalization depends on different reward models.
if args.task_name == 'depth':
    image = torchvision.transforms.functional.resize(image, (384, 384))
    image = normalize(image, (0.5, 0.5, 0.5), (0.5, 0.5, 0.5))
elif args.task_name in ['canny', 'lineart', 'hed']:
    pass
else:
    image = normalize(image, (0.485, 0.456, 0.406), (0.229, 0.224, 0.225))

# reward model inference
if args.task_name == 'canny':
    outputs = reward_model(image.to(accelerator.device), low_threshold, high_threshold)
else:
    outputs = reward_model(image.to(accelerator.device))

# Determine which samples in the current batch need to calculate reward loss
timestep_mask = (args.min_timestep_rewarding <= timesteps.reshape(-1, 1)) & (timesteps.reshape(-1, 1) <= args.max_timestep_rewarding)

# Calculate reward loss
reward_loss = get_reward_loss(outputs, labels, args.task_name, reduction='none')

# Calculate final loss
reward_loss = reward_loss.reshape_as(timestep_mask)
reward_loss = (timestep_mask * reward_loss).sum() / (timestep_mask.sum() + 1e-10)
loss = pretrain_loss + reward_loss * args.grad_scale

Please download the model weights and put them into each subset of checkpoints:

Please make sure the folder directory is consistent with the test script, then you can eval each model by:

bash eval/eval_ade20k.sh
bash eval/eval_cocostuff.sh
bash eval/eval_canny.sh
bash eval/eval_depth.sh
bash eval/eval_hed.sh
bash eval/eval_linedrawing.sh

The segmentation mIoU results of ControlNet and ControlNet++ in the arXiv v1 version of the paper were tested using images and labels saved in .jpg format, which resulted in errors.

We retested and reported the results using images and labels saved in .png format, please refer to our latest arXiv and ECCV Camera Ready releases.

Other comparison methods (Gligen/T2I-Adapter/UniControl/UniControlNet) and other evaluation metrics (FID/CLIP-score) were not affected by this error.

Please refer to the Inference Branch or try our online Huggingface demo

This project is licensed under the Apache License 2.0 - see the LICENSE file for details.

We sincerely thank the Huggingface, ControlNet and ImageReward communities for their open source code and contributions. Our project would not be possible without these amazing works.

If our work assists your research, feel free to give us a star ⭐ or cite us using:

@inproceedings{controlnet_plus_plus,
    author    = {Ming Li, Taojiannan Yang, Huafeng Kuang, Jie Wu, Zhaoning Wang, Xuefeng Xiao, Chen Chen},
    title     = {ControlNet++: Improving Conditional Controls with Efficient Consistency Feedback},
    booktitle = {European Conference on Computer Vision (ECCV)},
    year      = {2024},
}