ComfyUI-StyleTransferPlus

Advance Non-Diffusion-based Style Transfer in ComfyUI

Click name to jump to workflow

1. Neural Neighbor. Paper: Neural Neighbor Style Transfer
2. CAST. Paper: Domain Enhanced Arbitrary Image Style Transfer via Contrastive Learning
3. EFDM. Paper: Exact Feature Distribution Matching for Arbitrary Style Transfer and Domain Generalization
4. MicroAST. Paper: Towards Super-Fast Ultra-Resolution Arbitrary Style Transfer
5. UniST. Paper: Two Birds, One Stone: A Unified Framework for Joint Learning of Image and Video Style Transfers
6. AesPA-Net. Paper: Aesthetic Pattern-Aware Style Transfer Networks
7. TSSAT. Paper: Two-Stage Statistics-Aware Transformation for Artistic Style Transfer
8. AesFA. Paper: An Aesthetic Feature-Aware Arbitrary Neural Style Transfer

Extra nodes

More will come soon

Workflows

All nodes support batched input (i.e video) but is generally not recommended. One should generate 1 or 2 style frames (start and end), then use ComfyUI-EbSynth to propagate the style to the entire video.

Neural Neighbor (2020-2022) - Slowest	CAST (2022-2023) - Fast
EFDM (2022) - Fast	MicroAST (2023) - Fast
UniST (2023) - Medium - Only squares	AesPA-Net (2023) - Medium - Only squares
TSSAT (2023) - Slow	AesFA (2024) - Fast - Only squares

Neural Neighbor

Arguments:

• size: Either 512 or 1024, to be used for scaling the images. 1024 takes ~6-12 GB of VRAM. Defaults to 512
• scale_long: Scale by the longest (or shortest side) to size. Defaults to True
• flip: Augment style image with rotations. Slows down algorithm and increases memory requirement. Generally improves content preservation but hurts stylization slightly. Defaults to False
• content_loss: Defaults to True

  Use experimental content loss. The most common failure mode of our method is that colors will shift within an object creating highly visible artifacts, if that happens this flag can usually fix it, but currently it has some drawbacks which is why it isn't enabled by default [...]. One advantage of using this flag though is that [content_weight] can typically be set all the way to 0.0 and the content will remain recognizable.

• colorize: Whether to apply color correction. Defaults to True
• content_weight: weight = 1.0 corresponds to maximum content preservation, weight = 0.0 is maximum stylization (Defaults to 0.75)
• max_iter: How long each optimization step should take. size=512 does 4 scaling steps * max_iter. size=1024 does 5. The longer, the better and sharper the result.

See more information on the original repository. No model is needed.

workflow_neural_neighbor.json

CAST

Download the CAST and UCAST models in the Test section (Google Drive) (Unzip).

Download the vgg_normalized.pth model in the Train section

models/
│ .gitkeep
│ vgg_normalised.pth
─CAST_model
│   latest_net_AE.pth
│   latest_net_Dec_A.pth    # Safe to delete
│   latest_net_Dec_B.pth
│   test_opt.txt            # Safe to delete
│
└─UCAST_model
    latest_net_AE.pth 
    latest_net_Dec_A.pth    # Safe to delete
    latest_net_Dec_B.pth

workflow_cast.json

EFDM

Download the hm_decoder_iter_160000.pth in https://github.com/YBZh/EFDM/tree/main/ArbitraryStyleTransfer/models and put it in models/hm_decoder_iter_160000.pth

Download vgg_normalized.pth as instructed in CAST

Arguments:

• style_interp_weights: Weight for each style (only applicable for multiple styles). If empty, each style will have same weight (1/num_styles). Weights are automatically normalized so you can just enter anything
• model_arch: ["adain", "adamean", "adastd", "efdm", "hm"], defaults to "efdm"
• style_strength: From 0.00 to 1.00. The higher the stronger the style. Defaults to 1.00
• do_crop: Whether to resize and then crop the content and style images to size x size or just resize. Defaults to False
• preserve_color: Whether to preserve the color of the content image. Defaults to False
• size: Size of the height of the images after resizing. The aspect ratio of content is kept, while the styles will be resized to match content's height and width. The higher the size, the better the result, but also the more VRAM it will take. You may try use_cpu to circumvent OOM. Tested with size=2048 on 12GB VRAM.
• use_cpu: Whether to use CPU or GPU for this. CPU takes very long! (default workflow = ~21 seconds). Defaults to False.

workflow_efdm.json

MicroAST

Download models from https://github.com/EndyWon/MicroAST/tree/main/models and place them in the models folder

models/microast/content_encoder_iter_160000.pth.tar
models/microast/decoder_iter_160000.pth.tar
models/microast/modulator_iter_160000.pth.tar
models/microast/style_encoder_iter_160000.pth.tar

Arguments: Similar to EFDM

workflow_microast.json

UniST

Download UniST_model.pt in Testing and vgg_r41.pth, dec_r41.pth in Training section and place them in

models/unist/UniST_model.pt
models/unist/dec_r41.pth
models/unist/vgg_r41.pth

The model only works with square images, so inputs are resized to size x size. do_crop will resize height, then crop to square.

The Video node is "more native" than the Image node for batched images (video) inputs. The model works with batch=3 (3 consecutive frames), so we split the video into such.

UniST	UniST Video
workflow_unist_image.json	workflow_unist_video.json
	(Not a workflow-embedded image)

AesPA-Net

Download models from Usage section.

For the VGG model, download vgg_normalised_conv5_1.pth from https://github.com/pietrocarbo/deep-transfer/tree/master/models/autoencoder_vgg19/vgg19_5

models/aespa/dec_model.pth
models/aespa/transformer_model.pth
models/aespa/vgg_normalised_conv5_1.pth

The original authors provided a vgg_normalised_conv5_1.t7 model, which can only be opened with torchfile, and pyTorch has dropped native support for it. It also doesn't work reliably on Windows.

workflow_aespa.json

TSSAT

Download the TSSAT-model.zip in the Model Testing section, and unzip it

models/tssat/decoder_iter_160000.pth

# Outside folder, can be ignored from the zip if already downloaded from above
models/vgg_normalised.pth             

workflow_tssat.json

AesFA

Download main.pth in the Getting Started and place it in models/aesfa/main.pth.

workflow_aesfa_w_blend.json

Authors' Notes:

[...] style blending, i.e., using the low-frequency and high-frequency style information from different images. The style-transferred outputs finely keep the color information from the low-frequency image and change the texture information based on the high frequency image.

The vertical line-shape artifacts alongside the images are often observed. We reason that these appear because the content features are being convolved directly with the predicted aesthetic feature-aware kernels and biases in our model. In addition, the upsampling operation could be the ones that create artifacts.

Extra nodes

Coral Color Transfer

Credits

@inproceedings{zhang2020cast,
  author = {Zhang, Yuxin and Tang, Fan and Dong, Weiming and Huang, Haibin and Ma, Chongyang and Lee, Tong-Yee and Xu, Changsheng},
  title = {Domain Enhanced Arbitrary Image Style Transfer via Contrastive Learning},
  booktitle = {ACM SIGGRAPH},
  year = {2022}
}

@article{zhang2023unified,
  title={A Unified Arbitrary Style Transfer Framework via Adaptive Contrastive Learning},
  author={Zhang, Yuxin and Tang, Fan and Dong, Weiming and Huang, Haibin and Ma, Chongyang and Lee, Tong-Yee and Xu, Changsheng},
  journal={ACM Transactions on Graphics},
  year={2023},
  publisher={ACM New York, NY}
}

@inproceedings{zhang2021exact,
  title={Exact Feature Distribution Matching for Arbitrary Style Transfer and Domain Generalization},
  author={Zhang, Yabin and Li, Minghan and Li, Ruihuang and Jia, Kui and Zhang, Lei},
  booktitle={CVPR},
  year={2022}
}

@inproceedings{wang2023microast,
  title={MicroAST: Towards Super-Fast Ultra-Resolution Arbitrary Style Transfer},
  author={Wang, Zhizhong and Zhao, Lei and Zuo, Zhiwen and Li, Ailin and Chen, Haibo and Xing, Wei and Lu, Dongming},
  booktitle={Proceedings of the AAAI Conference on Artificial Intelligence},
  year={2023}
}

@InProceedings{Gu_2023_ICCV,
    author    = {Gu, Bohai and Fan, Heng and Zhang, Libo},
    title     = {Two Birds, One Stone: A Unified Framework for Joint Learning of Image and Video Style Transfers},
    booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
    month     = {October},
    year      = {2023},
    pages     = {23545-23554}
}

@article{Hong2023AesPANetAP,
  title={AesPA-Net: Aesthetic Pattern-Aware Style Transfer Networks},
  author={Kibeom Hong and Seogkyu Jeon and Junsoo Lee and Namhyuk Ahn and Kunhee Kim and Pilhyeon Lee and Daesik Kim and Youngjung Uh and Hyeran Byun},
  journal={ArXiv},
  year={2023},
  volume={abs/2307.09724},
  url={https://api.semanticscholar.org/CorpusID:259982728}
}