MiniMax-AI/VTP

MiniMax-AI/VTP

Description: Towards Scalable Pre-training of Visual Tokenizers for Generation

Language: Python

License: NOASSERTION

Stars: 490

Forks: 14

Open issues: 8

Created: 2025-12-11T10:32:23Z

Pushed: 2026-04-15T07:09:46Z

Default branch: main

Fork: no

Archived: no

README:

News

• [2026.03.09] We have updated our technical report with more experimental results.

• [2025.12.16] We have released our technical report 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.

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

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 foundational flaw: better pixel-level reconstruction 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 VTP, a unified visual tokenizer pre-training framework, pioneering the joint optimization of image-text contrastive, self-supervised, and reconstruction losses. Our study reveals that perception-oriented tokenizer pre-training unlocks a new scaling law for generation, where generative performance scales effectively with compute, parameters, and data allocated to the pre-training of the visual tokenizer. Our large-scale pre-training experiments demonstrate the following results: (1) Without modifying DiT training specs and FLOPs, solely scaling VTP pre-training consistently achieves gains in both ImageNet class-conditional and LAION text-to-image generation, while conventional autoencoders stagnate very early at 1/10 of the FLOPs. (2) VTP achieves 0.36 rFID while simultaneously delivering 78.2% zero-shot accuracy and 85.7% linear probing accuracy, surpassing prior unified tokenizers such as VILA-U and UniTok. (3) Furthermore, the VTP-based diffusion model exhibits exceptionally fast convergence---reaching 2.03 gFID in only 80 epochs without guidance tricks, outperforming previous methods like VA-VAE and RAE---and ultimately scales to achieve a remarkable 1.11 gFID on ImageNet 256×256 generation.

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…