| <div align="center"> | |
| <img src="figures/logo.png" alt="Logo" width="200"/> | |
| <h2> Towards Scalable Pre-training of Visual Tokenizers for Generation </h2> | |
| [Jingfeng Yao](https://github.com/JingfengYao)<sup>1</sup>, [Yuda Song](https://github.com/IDKiro)<sup>2</sup>, Yucong Zhou<sup>2</sup>, [Xinggang Wang](https://xwcv.github.io/)<sup>1,*</sup> | |
| <sup>1</sup>Huazhong University of Science and Technology | |
| <sup>2</sup>MiniMax | |
| <sup>*</sup>Corresponding author: xgwang@hust.edu.cn | |
| ***Work still in Progress.*** | |
| [](https://www.minimax.io/) | |
| [](https://www.minimax.io/news/minimax-hailuo-23) | |
| [](https://github.com/hustvl) | |
| [](https://huggingface.co/MiniMaxAI/VTP-Large-f16d64) | |
| [](https://github.com/MiniMax-AI/VTP) | |
| [](https://arxiv.org/abs/2512.13687) | |
| <img src="figures/abs.png" alt="Abstract Figure" width="900"/> | |
| </div> | |
| ## News | |
| - **[2025.12.16]** We have released our [technical report](https://arxiv.org/abs/2512.13687) and [pretrained weights](#get-checkpoints). | |
| ## 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.** | |
| <div align="center"> | |
| <img src="figures/scaling_v2.png" alt="Overview Figure" width="900"/> | |
| </div> | |
| ## Get Checkpoints | |
| | Checkpoints | | |
| |-------| | |
| | [](https://huggingface.co/MiniMaxAI/VTP-Small-f16d64) | | |
| | [](https://huggingface.co/MiniMaxAI/VTP-Base-f16d64) | | |
| | [](https://huggingface.co/MiniMaxAI/VTP-Large-f16d64) | | |
| Weights will be released very soon. | |
| <details> | |
| <summary><b style="font-size: 1.1em;">🚀 Click Here to Quick Start </b></summary> | |
| ``` | |
| pip install -r requirements.txt | |
| ``` | |
| ```python | |
| 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 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 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 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) | |
| ``` | |
| </details> | |
| ## Performance | |
| <table> | |
| <tr> | |
| <th rowspan="2">Model</th> | |
| <th colspan="2" style="text-align: center;">Understanding</th> | |
| <th colspan="1" style="text-align: center;">Reconstruction</th> | |
| <th colspan="1" style="text-align: center;">Generation</th> | |
| </tr> | |
| <tr> | |
| <th style="text-align: center;">Zero-shot Acc.</th> | |
| <th style="text-align: center;">Linear Probing</th> | |
| <th style="text-align: center;">rFID</th> | |
| <th style="text-align: center;">LightningDiT-XL 80ep<br>nocfg FID-50K</th> | |
| </tr> | |
| <tr><td><a href="https://github.com/mlfoundations/open_clip">OpenCLIP</a></td><td style="text-align: center;">74.0</td><td style="text-align: center;">-</td><td style="text-align: center;">-</td><td style="text-align: center;">-</td></tr> | |
| <tr><td><a href="https://github.com/openai/CLIP">CLIP</a></td><td style="text-align: center;">75.5</td><td style="text-align: center;">-</td><td style="text-align: center;">-</td><td style="text-align: center;">-</td></tr> | |
| <tr><td><a href="https://github.com/google-research/big_vision">SigLIP</a></td><td style="text-align: center;"><strong>80.5</strong></td><td style="text-align: center;">-</td><td style="text-align: center;">-</td><td style="text-align: center;">-</td></tr> | |
| <tr><td><a href="https://github.com/facebookresearch/mae">MAE</a></td><td style="text-align: center;">-</td><td style="text-align: center;">85.9</td><td style="text-align: center;">-</td><td style="text-align: center;">-</td></tr> | |
| <tr><td><a href="https://github.com/facebookresearch/dinov2">DINOv2</a></td><td style="text-align: center;">-</td><td style="text-align: center;"><strong>86.7</strong></td><td style="text-align: center;">-</td><td style="text-align: center;">-</td></tr> | |
| <tr><td><a href="https://github.com/FoundationVision/UniTok">UniTok</a></td><td style="text-align: center;">70.8</td><td style="text-align: center;">-</td><td style="text-align: center;">0.41</td><td style="text-align: center;">-</td></tr> | |
| <tr><td><a href="https://github.com/mit-han-lab/vila-u">VILA-U</a></td><td style="text-align: center;">73.3</td><td style="text-align: center;">-</td><td style="text-align: center;">1.80</td><td style="text-align: center;">-</td></tr> | |
| <tr><td><a href="https://github.com/hustvl/LightningDiT">VA-VAE-f16d32</a></td><td style="text-align: center;">-</td><td style="text-align: center;">-</td><td style="text-align: center;">0.28</td><td style="text-align: center;">4.29</td></tr> | |
| <tr><td><a href="https://github.com/hustvl/LightningDiT">VA-VAE-f16d64</a></td><td style="text-align: center;">-</td><td style="text-align: center;">-</td><td style="text-align: center;"><strong>0.15</strong></td><td style="text-align: center;">-</td></tr> | |
| <tr><td><a href="https://github.com/bytetriper/RAE">RAE-f16d768</a></td><td style="text-align: center;">-</td><td style="text-align: center;">84.5</td><td style="text-align: center;">0.57</td><td style="text-align: center;">4.28</td></tr> | |
| <tr><td><b>VTP-S-f16d64 (ours)</b></td><td style="text-align: center;">66.7</td><td style="text-align: center;">77.5</td><td style="text-align: center;">0.98</td><td style="text-align: center;">5.46</td></tr> | |
| <tr><td><b>VTP-B-f16d64 (ours)</b></td><td style="text-align: center;">73.2</td><td style="text-align: center;">81.0</td><td style="text-align: center;">0.74</td><td style="text-align: center;">3.88</td></tr> | |
| <tr><td><b>VTP-L-f16d64 (ours)</b></td><td style="text-align: center;">78.2</td><td style="text-align: center;">85.7</td><td style="text-align: center;">0.36</td><td style="text-align: center;"><strong>2.81</strong></td></tr> | |
| </table> | |
| ## 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. | |
| <div align="center"> | |
| <img src="figures/overview.png" alt="Overview Figure" width="900"/> | |
| </div> | |
| ## Evaluation | |
| #### Installation | |
| ```bash | |
| 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](https://github.com/hustvl/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](https://github.com/mlfoundations/open_clip) and [DINOv2](https://github.com/facebookresearch/dinov2). Our final model variant uses [DINOv3](https://github.com/facebookresearch/dinov3) architecture. | |
| We use [LightningDiT](https://github.com/hustvl/LightningDiT) for generation evaluation. | |
| Thanks for their great codes. | |
| ## Citation | |
| ```bibtex | |
| @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 model@minimax.io. |