Open-MAGVIT2: Democratizing Autoregressive Visual Generation

VQGAN remains essential in autoregressive visual generation, despite limitations in codebook size and utilization that underestimate its capabilities. MAGVIT2 addresses these issues with a lookup-free technique and a large codebook ($2^{18}$), showing promising results in image and video generation, and playing a key role in VideoPoet. However, we currently lack access to this tokenizer. 💔

In our codebase, we have re-implemented the MAGVIT2 tokenizer in PyTorch, closely replicating the original results. We hope our efforts will foster innovation and creativity in the field of autoregressive visual generation. 💚

📰 News

• [2024.06.17] 🔥🔥🔥 We release the training code of the image tokenizer and checkpoints for different resolutions, achieving state-of-the-art performance (0.39 rFID for 8x downsampling) compared to VQGAN, MaskGIT, and recent TiTok, LlamaGen, and OmniTokenizer.

🎤 TODOs

• Better image tokenizer with scale-up training.
• Finalize the training of the autoregressive model.
• Video tokenizer and the corresponding autoregressive model.

🤗 Open-MAGVIT2 is still at an early stage and under active development. Stay tuned for the update!

📖 Implementations

Figure 1. The framework of the Open-MAGVIT2 tokenizer, composed of an encoder, a lookup-free quantizer (LFQ), and a decoder.

🛠️ Installation

• Env: We have tested on Python 3.8.8 and CUDA 11.7 (other versions may also be fine).
• Dependencies: pip install -r requirements
• Datasets

imagenet
└── train/
    ├── n01440764
        ├── n01440764_10026.JPEG
        ├── n01440764_10027.JPEG
        ├── ...
    ├── n01443537
    ├── ...
└── val/
    ├── ...

Stage I: Training of Visual Tokenizer

We follow the design of Generator in MAGVIT-2 but use PatchGAN instead of StyleGAN as Discriminator for GAN training. We use the combination of Loss utilized in MAGVIT-2 and VQGAN for better training stability and reconstruction quality. All the training details can be found in the config files. Note that, we train our model using 32 $\times$ V100.

🍺 Quantitative Comparison

Table 1. Reconstruction performance of different tokenizers on $256 \times 256$ ImageNet 50k validation set. Open-MAGVIT2 achieves SOTA results on different downsampling rates.

Method	Token Type	#Tokens	Train Data	Codebook Size	rFID	PSNR	Codebook Utilization	Checkpoint
VQGAN	2D	16 $\times$ 16	256 $\times$ 256 ImageNet	1024	7.94	19.4	-	-
MaskGIT	2D	16 $\times$ 16	256 $\times$ 256 ImageNet	1024	2.28	-	-	-
LlamaGen	2D	16 $\times$ 16	256 $\times$ 256 ImageNet	16384	2.19	20.79	97%	-
🔥Open-MAGVIT2	2D	16 $\times$ 16	256 $\times$ 256 ImageNet	262144	1.53	21.53	100%	IN256_Base
ViT-VQGAN	2D	32 $\times$ 32	256 $\times$ 256 ImageNet	8192	1.28	-	-	-
VQGAN	2D	32 $\times$ 32	OpenImages	16384	1.19	23.38	-	-
OmniTokenizer-VQ	2D	32 $\times$ 32	256 $\times$ 256 ImageNet	8192	1.11	-	-	-
LlamaGen	2D	32 $\times$ 32	256 $\times$ 256 ImageNet	16384	0.59	24.45	-	-
🔥Open-MAGVIT2*	2D	32 $\times$ 32	128 $\times$ 128 ImageNet	262144	0.39	25.78	100%	IN128_Base
TiTok-L	1D	32	256 $\times$ 256 ImageNet	4096	2.21	-	-	-
TiTok-B	1D	64	256 $\times$ 256 ImageNet	4096	1.70	-	-	-
TiTok-S	1D	128	256 $\times$ 256 ImageNet	4096	1.71	-	-	-

(*) denotes that the results are from the direct inference using the model trained with $128 \times 128$ resolution without fine-tuning.

Table 2. Compare with the original MAGVIT2 by training and testing with both $128 \times 128$ resolution as used in its original paper. ImageNet 50k validation set is used for testing.

Method	Token Type	#Tokens	Data	LFQ	Large Codebook	Up/Down Sampler	rFID	URL
MAGVIT2	2D	$16 \times 16$	128 $\times$ 128 ImageNet	√	√	√	1.21	-
Open-MAGVIT2	2D	$16 \times 16$	128 $\times$ 128 ImageNet	√	√	√	1.56	IN128_Base

👀 Reconstruction Visualization

Figure 2. Visualization of the Open-MAGVIT2 tokenizer trained at $256 \times 256$ resolution and tested at $256 \times 256$ resolution (imagenet_256_Base version). (a) indicates the original images while (b) specifies the reconstruction images.

Figure 3. Visualization of the Open-MAGVIT2 tokenizer trained at $128 \times 128$ resolution and tested at $512 \times 512$ resolution (imagenet_128_Base version). (a) indicates the original images while (b) specifies the reconstruction images.

🚀 Training Scripts

• $128\times 128$ Tokenizer Training

bash run_B_128.sh

• $256\times 256$ Tokenizer Training

bash run_B_256.sh

Stage II: Training of Autoregressive Generation

MAGVIT2 utilizes Non-AutoRegressive transformer for image generation. Instead, we would like to exploit the potential of Autogressive Visual Generation with the relatively large codebook. We are currently exploring Stage II training.

❤️ Acknowledgement

We thank Lijun Yu for his encouraging discussions. We refer a lot from VQGAN and MAGVIT. Thanks for their wonderful work.

✏️ Citation

If you found the codebase helpful, please cite it.

@software{Luo_Open-MAGVIT2_2024,
author = {Luo, Zhuoyan and Shi, Fengyuan and Ge, Yixiao},
month = jun,
title = {{Open-MAGVIT2}},
url = {https://github.com/TencentARC/Open-MAGVIT2},
version = {1.0},
year = {2024}
}

@inproceedings{
yu2024language,
title={Language Model Beats Diffusion - Tokenizer is key to visual generation},
author={Lijun Yu and Jose Lezama and Nitesh Bharadwaj Gundavarapu and Luca Versari and Kihyuk Sohn and David Minnen and Yong Cheng and Agrim Gupta and Xiuye Gu and Alexander G Hauptmann and Boqing Gong and Ming-Hsuan Yang and Irfan Essa and David A Ross and Lu Jiang},
booktitle={The Twelfth International Conference on Learning Representations},
year={2024},
url={https://openreview.net/forum?id=gzqrANCF4g}
}