Every latent diffusion model depends on a visual tokenizer, the autoencoder that squeezes images into the compact latents the generator works with. VTP, a research project from Huazhong University of Science and Technology and MiniMax, asks what happens when you pre-train that tokenizer at scale the way you would a vision foundation model, and its answer is that generation gets better as the tokenizer grows.
| Repository | MiniMax-AI/VTP |
| Maintainer | MiniMax-AI (with Huazhong University of Science and Technology) |
| Primary language | Python |
| GitHub stars | 508 (as of 15 September 2026) |
| Last push | 15 April 2026 |
| License status | MIT (repository software) |
What it does
According to the README, the team trained visual tokenizers from scratch by combining contrastive, self-supervised and reconstruction objectives. The project’s stated takeaways are:
- At the same compute budget for diffusion transformer training, scaling VTP gives better generation.
- Traditional autoencoders “CANNOT be scaled up” for diffusion models, in the authors’ words.
- Understanding, not just reconstruction, is the key driver of how well the tokenizer scales.
The resulting model is used three ways in the README’s example code: as an autoencoder (reported reconstruction rFID 0.36), as a CLIP-style image-text model (reported zero-shot accuracy 78.2), and as a self-supervised feature extractor (reported linear-probing accuracy 85.7). These are the authors’ numbers; see the technical report for the evaluation details.
Who it’s for
- Researchers building or training latent diffusion models who want an alternative tokenizer.
- Readers following where image-generation quality gains are coming from beyond the generator itself.
Getting started
The README’s quick start is pip install -r requirements.txt, then loading a checkpoint such as VTP-Large-f16d64 with VTPModel.from_pretrained. Its example preprocesses images to 256×256 and runs under CUDA autocast. Checkpoints are listed through the project’s Hugging Face collection. No hardware requirements are documented.
License status
Checked October 17, 2026: The repository’s MIT License permits use, modification and distribution of the software subject to its notice requirements. Check the model card and license supplied with each checkpoint separately; this software license alone does not establish the terms of every weight file or dataset.
Limitations and open questions
- The README labels the project “Work still in Progress”.
- It is internally inconsistent about weights: the news section says pretrained weights were released on 16 December 2025, while the checkpoints section still says “Weights will be released very soon.”
- This is a component for model builders, not an end-user image generator.
- We have not reproduced any of the reported results.
Sources
- GitHub repository and README
- License file
- Technical report: Towards Scalable Pre-training of Visual Tokenizers for Generation (arXiv 2512.13687)
- Hugging Face collection
Related: AI repositories · ReVersion, research on steering diffusion models.
This profile was drafted with AI assistance from the project’s public repository and reviewed by The Signal’s editors before publication. See our AI Disclosure.
