Towards Scalable Pre-training of Visual Tokenizers for Generation

Jingfeng Yao¹, Yuda Song², Yucong Zhou², Xinggang Wang^1,*

¹Huazhong University of Science and Technology ²MiniMax
^*Corresponding author: [email protected]

Work still in Progress.

News

• [2025.12.16] We have released our technical report and pretrained weights.

Takeaways

By integrating contrastive, self-supervised, and reconstruction learning, we have trained numerous visual tokenizers from scratch. We are seeking to unveil the novel scalability interlinking understanding, generation, and reconstruction.

• Same FLOPs in DiT Training, VTP scaling helps better generation.

• Traditional auto-encoders CANNOT be scaled up for diffusion generative models.

• Understanding is the key driver for improving the learnability scaling.

• Parameter, data and training scalability can be seen while representation learning involved.

Get Checkpoints

Checkpoints

Weights will be released very soon.

🚀 Click Here to Quick Start

pip install -r requirements.txt

import torch
from PIL import Image
from torchvision import transforms

from vtp.models.vtp_hf import VTPConfig, VTPModel
from vtp.tokenizers import get_tokenizer

model = VTPModel.from_pretrained("/path/to/MiniMaxAI/VTP-Large-f16d64")
model.eval()

# print model parameters
def count_params(m): return sum(p.numel() for p in m.parameters()) / 1e6
print(f"Vision Encoder: {count_params(model.trunk):.1f}M")
print(f"Pixel Decoder:  {count_params(model.pixel_decoder):.1f}M")
print(f"Text Encoder:   {count_params(model.text_transformer):.1f}M")

preprocess = transforms.Compose([
    transforms.Resize((256, 256)),
    transforms.ToTensor(),
    transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
])
image = preprocess(Image.open("figures/dog.png")).unsqueeze(0)

# ---------------------------------------------------------------------------------------
# use it as an auto-encoder; rFID=0.36
# ---------------------------------------------------------------------------------------
denormalize = transforms.Normalize(
    mean=[-0.485/0.229, -0.456/0.224, -0.406/0.225],
    std=[1/0.229, 1/0.224, 1/0.225]
)
with torch.no_grad(), torch.autocast("cuda"):
    latents = model.get_reconstruction_latents(image)  # encode
    recon = model.get_latents_decoded_images(latents)  # decode
recon_image = denormalize(recon[0]).clamp(0, 1).permute(1, 2, 0).cpu().numpy()
Image.fromarray((recon_image * 255).astype("uint8")).save("output/reconstructed.png")


# ---------------------------------------------------------------------------------------
# use it as a clip; zero-shot 78.2
# ---------------------------------------------------------------------------------------
tokenizer = get_tokenizer('ViT-B-32', context_length=model.config.text_context_length)
text = tokenizer(["a diagram", "a dog", "a cat", "a person"])
with torch.no_grad(), torch.autocast("cuda"):
    image_features = model.get_clip_image_feature(image, normalize=True)
    text_features = model.get_clip_text_feature(text, normalize=True)
    text_probs = (100.0 * image_features @ text_features.T).softmax(dim=-1)
print("Label probs:", [f"{p:.4f}" for p in text_probs[0].tolist()])

# ---------------------------------------------------------------------------------------
# use it as an ssl feature extractor; linear probing 85.7
# ---------------------------------------------------------------------------------------
with torch.no_grad(), torch.autocast("cuda"):
    # get last layer features (cls token + patch tokens)
    features = model.get_last_layer_feature(image)
    cls_token = features['cls_token']      # (B, 1024)
    patch_tokens = features['patch_tokens']  # (B, 256, 1024) for 256x256 image

    # or get intermediate layer features for linear probing
    intermediate = model.get_intermediate_layers_feature(
        image, n=4, return_class_token=True
    )  # returns 4 x (patch_tokens, cls_token), each cls_token is (B, 1024)
    for i in range(1, 5):
        print('Last %d layers:' % i)
        print('Patch tokens shape:', intermediate[-i][0].shape)
        print('Cls token shape:', intermediate[-i][1].shape)

Performance

Model	Understanding		Reconstruction	Generation
	Zero-shot Acc.	Linear Probing	rFID	LightningDiT-XL 80ep nocfg FID-50K
OpenCLIP	74.0	-	-	-
CLIP	75.5	-	-	-
SigLIP	80.5	-	-	-
MAE	-	85.9	-	-
DINOv2	-	86.7	-	-
UniTok	70.8	-	0.41	-
VILA-U	73.3	-	1.80	-
VA-VAE-f16d32	-	-	0.28	4.29
VA-VAE-f16d64	-	-	0.15	-
RAE-f16d768	-	84.5	0.57	4.28
VTP-S-f16d64 (ours)	66.7	77.5	0.98	5.46
VTP-B-f16d64 (ours)	73.2	81.0	0.74	3.88
VTP-L-f16d64 (ours)	78.2	85.7	0.36	2.81

Introduction

The quality of the latent space in visual tokenizers (e.g., VAEs) is crucial for modern generative models. However, the standard reconstruction-based training paradigm produces a latent space that is biased towards low-level information, leading to a foundation flaw: better pixel-level accuracy does not lead to higher-quality generation. This implies that pouring extensive compute into visual tokenizer pre-training translates poorly to improved performance in generation.

We identify this as the "pre-training scaling problem" and suggest a necessary shift: to be effective for generation, a latent space must concisely represent high-level semantics. We present visual tokenizer pre-training, VTP, a unified visual tokenizer pre-training framework, pioneering the joint optimization of image-text contrastive, self-supervised, and reconstruction losses. Our large-scale study reveals two principal findings: (1) understanding is a key driver of generation, and (2) much better scaling properties, where generative performance scales effectively with compute, parameters, and data allocated to the pretraining of the visual tokenizer. After large-scale pre-training, our tokenizer delivers a competitive profile (78.2 zero-shot accuracy, 0.36 rFID) and 3× faster convergence on generation compared to advanced distillation methods. More importantly, it scales effectively: without modifying standard DiT training specs, solely investing more FLOPS in pretraining VTP achieves 65.8% FID improvement in downstream generation, while conventional autoencoder stagnates very early at 1/10 FLOPS.

Evaluation

Installation

conda create -n vtp python=3.10
conda activate vtp
git submodule update --init --recursive
pip install -r requirements.txt

Zero-shot Classification

Modify the corresponding paths in scripts/test_zero_shot_hf.sh. Run:

bash scripts/test_zero_shot_hf.sh 

Linear Probing Classification

Modify the corresponding paths in scripts/test_linear_probing_hf.sh. Run:

bash scripts/test_linear_probing_hf.sh

ImageNet Reconstruction

Modify the corresponding paths in scripts/test_reconstruction_hf.sh. Run:

bash scripts/test_reconstruction_hf.sh

ImageNet Generation

We use LightningDiT codes to evaluate our generation performance.

Feature extraction:

bash generation/scripts/extract_features_vtp.sh generation/configs/train_vtp_l_dit_xl.yaml

LightningDiT training:

bash generation/scripts/train_lightningdit_vtp.sh generation/configs/train_vtp_l_dit_xl.yaml

LightningDiT sampling:

bash generation/scripts/inference_lightningdit_vtp.sh generation/configs/train_vtp_l_dit_xl.yaml

Acknowledgements

Our pre-training codes are built upon OpenCLIP and DINOv2. Our final model variant uses DINOv3 architecture.

We use LightningDiT for generation evaluation.

Thanks for their great codes.

Citation

@article{vtp,
  title={Towards Scalable Pre-training of Visual Tokenizers for Generation},
  author={Yao, Jingfeng and Song, Yuda and Zhou, Yucong and Wang, Xinggang},
  journal={arXiv preprint arXiv:2512.13687},
  year={2025}
}

Contact Us

Contact us at [email protected].