⚡ Selftok: Discrete Visual Tokens of Autoregression, by Diffusion, and for Reasoning
Selftok Team, Media Technology Institute, Huawei
✨ Highlights
Propose Self-Consistency Tokenizer (Selftok), a SOTA tokenizer that achieves both high-quality reconstruction and high compression bit rate.
Selftok offers an elegant and minimalist approach to unify diffusion and AR for vision-language models (VLM).
Our VLM achieves both SOTA visual comprehension and generation performances.
📰 News
[2025.05.18] The weights of tokenizer for Selftok are available on HuggingFace.
[2025.05.15] We have released the code of tokenizer for Selftok! The weights will be released soon.
[2025.05.12] We have released the paper of Selftok (arXiv)!
[2025.04.04] Our preliminary work DDT-LLaMA (project page) has been accepted as an Oral Presentation at CVPR 2025!
📄 Introduction
We completely discard the conventional spatial prior in image representation and introduce a novel discrete visual tokenizer: Self-Consistency Tokenizer (Selftok). At its design core, we compose an autoregressive (AR) prior—mirroring the causal structure of language—into visual tokens by using the reverse diffusion process of image generation. The AR property makes Selftok fundamentally distinct from traditional spatial tokens in the following two key ways:
- Selftok offers an elegant and minimalist approach to unify diffusion and AR for vision-language models: By representing images with Selftok tokens, we can train vision-language models (VLMs) using a purely discrete autoregressive architecture—like that in LLMs—without requiring additional modules or training objectives.
- We theoretically show that the AR prior satisfies the Bellman equation, whereas the spatial prior does not. Therefore, Selftok supports reinforcement learning (RL) for visual generation with effectiveness comparable to that achieved in LLMs.
Besides the AR property, Selftok is also a SOTA tokenizer that achieves both high-quality reconstruction and high compression bit rate. After representing the training images as Selftok tokens, as a pure AR model, our VLM achieves both SOTA visual comprehension and generation performances. Impressively, without using any text-image training pairs, a simple policy gradient RL working in the visual tokens can significantly boost the visual generation benchmark, surpassing all the existing models by a large margin.
Therefore, we believe that Selftok effectively addresses the long-standing challenge that visual tokens cannot support effective RL. When combined with the well-established strengths of RL in LLMs, this brings us one step closer to realizing a truly multimodal LLM.
📝 Results
- SoTA Reconstruction Performance on ImageNet 256x256
🎯 How to Use
🛠️ Installation
conda create -n selftok python=3.10 # or your preferred version
conda activate selftok
# For Ascend environment
pip install -r requirements.txt
# For GPU environment
pip install -r requirements_gpu.txt
🧠 Tokenizer Inference with Pre-trained Models
- Download Pretrained Weights
| Tokenizer | Image Resolution | # Tokens | PSNR |
|---|---|---|---|
| Selftok w/o Renderer (HuggingFace) | 256×256 | 512 | 21.86 |
| Selftok w/ Renderer (HuggingFace) | 256×256 | 512 | 24.14 |
| Selftok w/o Renderer (HuggingFace) | 256×256 | 1024 | 23.06 |
| Selftok w/ Renderer (HuggingFace) | 256×256 | 1024 | 26.30 |
Pipeline Overview
The inference pipeline includes three key stages:
- Tokenization – Convert images into discrete token sequences.
- Diffusion Decoding – Reconstruct images using a 50-step diffusion model.
- One-step Decoding – Quickly reconstruct images using a fast renderer.
bash
git clone https://github.com/selftok-team/SelftokTokenizer.git
cd ./SelftokTokenizer
1. Tokenization
This script demonstrates how to convert images into token sequences using a pretrained Selftok model.
import argparse
from mimogpt.engine.utils import parse_args_from_yaml
from torchvision import transforms
from PIL import Image
import torch
import numpy as np
from mimogpt.infer.SelftokPipeline import SelftokPipeline
from mimogpt.infer.SelftokPipeline import NormalizeToTensor
from torchvision.utils import save_image
parser = argparse.ArgumentParser()
parser.add_argument("--yml-path", type=str, default="path/to/your/config.yml")
parser.add_argument("--pretrained", type=str, default="path/to/your/ckpt.pth")
parser.add_argument("--data_size", type=int, default=256)
cfg = parse_args_from_yaml(args.yml_path)
model = SelftokPipeline(cfg=cfg, ckpt_path=args.pretrained, datasize=args.data_size, device='cuda')
img_transform = transforms.Compose([
transforms.Resize(args.data_size),
transforms.CenterCrop(args.data_size),
NormalizeToTensor(),
])
image_paths = ['path/to/image1.png', 'path/to/image2.png']
images = [img_transform(Image.open(p)) for p in image_paths]
images = torch.stack(images).to('cuda')
tokens = model.encoding(images, device='cuda')
np.save('token.npy', tokens.detach().cpu().numpy())
2. Diffusion Decoding
Reconstruct images from token sequences using the full diffusion model (50 steps):
import argparse
from mimogpt.engine.utils import parse_args_from_yaml
from torchvision import transforms
from PIL import Image
import torch
import numpy as np
from mimogpt.infer.SelftokPipeline import SelftokPipeline
from mimogpt.infer.SelftokPipeline import NormalizeToTensor
from torchvision.utils import save_image
parser = argparse.ArgumentParser()
parser.add_argument("--yml-path", type=str, default="path/to/your/config.yml")
parser.add_argument("--pretrained", type=str, default="path/to/your/ckpt.pth")
parser.add_argument("--data_size", type=int, default=256)
cfg = parse_args_from_yaml(args.yml_path)
model = SelftokPipeline(cfg=cfg, ckpt_path=args.pretrained, datasize=args.data_size, device='cuda')
tokens = np.load('token.npy')
images = model.decoding(tokens, device='cuda')
for b, img in enumerate(images):
save_image(img, f"re_{b}.png")
3. One-step Renderer Decoding
Reconstruct images using a fast, one-step renderer:
import argparse
from mimogpt.engine.utils import parse_args_from_yaml
from torchvision import transforms
from PIL import Image
import torch
import numpy as np
from mimogpt.infer.SelftokPipeline import SelftokPipeline
from mimogpt.infer.SelftokPipeline import NormalizeToTensor
from torchvision.utils import save_image
parser = argparse.ArgumentParser()
parser.add_argument("--yml-path", type=str, default="path/to/your/config.yml")
parser.add_argument("--pretrained", type=str, default="path/to/your/ckpt.pth")
parser.add_argument("--data_size", type=int, default=256)
cfg = parse_args_from_yaml(args.yml_path)
model = SelftokPipeline(cfg=cfg, ckpt_path=args.pretrained, datasize=args.data_size, device='cuda')
tokens = np.load('token.npy')
images = model.decoding_with_renderer(tokens, device='cuda')
for b, img in enumerate(images):
save_image(img, f"render_{b}.png")
Notes
- Replace all
path/to/...with actual paths on your system or object storage. - The scripts assume CUDA is available; modify
device='cuda'to'cpu'if running on CPU. - The scripts support both Ascend and GPU. If inference with GPU, replace
mimogpt.infer.SelftokPipelinewithmimogpt.infer.SelftokPipeline_GPU. - If you use Selftok Tokenizer for AR training, note that we decoder the image token sequence reversely!
🎮 Train Your Own Models
The training code is currently under preparation and will be released shortly. Please stay tuned for updates.
📝 Citation
If you find our work useful, please cite our related paper:
# Arxiv
@article{wang2025discretevisualtokensautoregression,
title={Discrete Visual Tokens of Autoregression, by Diffusion, and for Reasoning},
author={Bohan Wang and Zhongqi Yue and Fengda Zhang and Shuo Chen and Li'an Bi and Junzhe Zhang and Xue Song and Kennard Yanting Chan and Jiachun Pan and Weijia Wu and Mingze Zhou and Wang Lin and Kaihang Pan and Saining Zhang and Liyu Jia and Wentao Hu and Wei Zhao and Hanwang Zhang},
year={2025},
eprint={2505.07538},
archivePrefix={arXiv},
primaryClass={cs.CV},
url={https://arxiv.org/abs/2505.07538},
}
# CVPR 2025
@article{pan2025generative,
title={Generative Multimodal Pretraining with Discrete Diffusion Timestep Tokens},
author={Pan, Kaihang and Lin, Wang and Yue, Zhongqi and Ao, Tenglong and Jia, Liyu and Zhao, Wei and Li, Juncheng and Tang, Siliang and Zhang, Hanwang},
journal={arXiv preprint arXiv:2504.14666},
year={2025}
}
Disclaimer
This open-source project is not an official Huawei product. Huawei is not responsible for providing support or maintenance for this project.