The performance of Latent Diffusion Models (LDMs) is critically dependent on the quality of their visual tokenizer. While recent works have explored incorporating Vision Foundation Models (VFMs) via distillation, we identify a fundamental flaw in this approach: it inevitably weakens the robustness of alignment with the original VFM, causing the aligned latents to deviate semantically under distribution shifts. In this paper, we bypass distillation by proposing a more direct approach: Vision Foundation Model Variational Autoencoder (VFM-VAE). To resolve the inherent tension between the VFM's semantic focus and the need for pixel-level fidelity, we redesign the VFM-VAE decoder with Multi-Scale Latent Fusion and Progressive Resolution Reconstruction blocks, enabling high-quality reconstruction from spatially coarse VFM features. Furthermore, we provide a comprehensive analysis of representation dynamics during diffusion training, introducing the proposed SE-CKNNA metric as a more precise tool for this diagnosis. This analysis allows us to develop a joint tokenizer-diffusion alignment strategy that dramatically accelerates convergence. Our innovations in tokenizer design and training strategy lead to superior performance and efficiency: our system reaches a gFID (w/o CFG) of 2.20 in merely 80 epochs (a 10x speedup over prior tokenizers). With continued training to 640 epochs, it further attains a gFID (w/o CFG) of 1.62, establishing direct VFM integration as a superior paradigm for LDMs.

Specifically, we introduce a pure transformer architecture with 4D rotary position embeddings to process visual inputs of arbitrary resolutions and temporal durations. To ensure stable training, we introduce an adversarial-free training objective that combines perceptual and Gram matrix losses, achieving state-of-the-art reconstruction quality. These results shed light on the next-generation multimodal AI systems built upon unified visual tokenization. Large Language Models (LLMs) (Chowdhery et al., 2023; Achiam et al., 2023; Touvron et al., 2023; Team et al., 2023; Guo et al., 2025) have achieved unprecedented generalization, with single models handling coding, reasoning, translation, and numerous other tasks that previously required specialized systems. This versatility largely stems from transformer architectures and simple tokenizers, such as BPE (Sennrich et al., 2015), which convert all text types - code, documents, tables, and multiple languages - into a unified token space. This shared representation enables efficient scaling and seamless knowledge transfer across language tasks. In contrast, visual representations remain fragmented due to inherent complexities. Unlike text's discrete symbolic nature, visual tasks demand distinct levels of abstraction: generation requires tokenizers that preserve low-level visual details for reconstruction, while understanding requires encoders that extract high-level semantic features through text alignment. Moreover, visual data exists in disparate formats: 2D grids for images, temporal sequences for videos, and varied 3D representations (e.g., meshes, voxels, and Gaussian splats) (Mescheder et al., 2019; Achlioptas et al., 2018; Mildenhall et al., 2021; Kerbl et al., 2023). Without a shared representation, vision systems remain fundamentally limited, unable to achieve the generalization and transfer learning that characterizes modern language models. Despite recent progress, unified visual tokenizers face three fundamental challenges. First, existing approaches optimize for either reconstruction or understanding, but not both: visual encoders (Radford et al., 2021; Zhai et al., 2023; Bolya et al., 2025) achieve semantic alignment but lack Description of each author's contribution is available in Appendix A. Corresponding to Jiasen Lu. Second, architectural choices create different limitations: convolutional tokenizers exhibit diminishing returns when scaling model parameters (Xiong et al., 2025), while transformer tokenizers (Y u et al., 2021; Wang et al., 2024b; Hansen-Estruch et al., 2025) achieve better scaling but suffer from severe adversarial training instabilities.

Recently, multimodal large language models (MLLMs) have emerged as a key approach in achieving artificial general intelligence. In particular, vision-language MLLMs have been developed to generate not only text but also visual outputs from multimodal inputs. This advancement requires efficient image tokens that LLMs can process effectively both in input and output. However, existing image tokenization methods for MLLMs typically capture only global abstract concepts or uniformly segmented image patches, restricting MLLMs' capability to effectively understand or generate detailed visual content, particularly at the object level. To address this limitation, we propose an object-centric visual tokenizer based on Slot Attention specifically for MLLMs. In particular, based on the Q-Former encoder, diffusion decoder, and residual vector quantization, our proposed discretized slot tokens can encode local visual details while maintaining high-level semantics, and also align with textual data to be integrated seamlessly within a unified next-token prediction framework of LLMs. The resulting Slot-MLLM demonstrates significant performance improvements over baselines with previous visual tokenizers across various vision-language tasks that entail local detailed comprehension and generation. Notably, this work is the first demonstration of the feasibility of object-centric slot attention performed with MLLMs and in-the-wild natural images.

Autoregressive (AR) models have recently shown strong performance in image generation, where a critical component is the visual tokenizer (VT) that maps continuous pixel inputs to discrete token sequences. The quality of the VT largely defines the upper bound of AR model performance. However, current discrete VTs fall significantly behind continuous variational autoencoders (VAEs), leading to degraded image reconstructions and poor preservation of details and text. Existing benchmarks focus on end-to-end generation quality, without isolating VT performance. To address this gap, we introduce VTBench, a comprehensive benchmark that systematically evaluates VTs across three core tasks: Image Reconstruction, Detail Preservation, and Text Preservation, and covers a diverse range of evaluation scenarios. We systematically assess state-of-the-art VTs using a set of metrics to evaluate the quality of reconstructed images. Our findings reveal that continuous VAEs produce superior visual representations compared to discrete VTs, particularly in retaining spatial structure and semantic detail. In contrast, the degraded representations produced by discrete VTs often lead to distorted reconstructions, loss of fine-grained textures, and failures in preserving text and object integrity. Furthermore, we conduct experiments on GPT-4o image generation and discuss its potential AR nature, offering new insights into the role of visual tokenization. We release our benchmark and codebase publicly to support further research and call on the community to develop strong, general-purpose open-source VTs.

Latent diffusion models with Transformer architectures excel at generating high-fidelity images. However, recent studies reveal an optimization dilemma in this two-stage design: while increasing the per-token feature dimension in visual tokenizers improves reconstruction quality, it requires substantially larger diffusion models and more training iterations to achieve comparable generation performance. Consequently, existing systems often settle for sub-optimal solutions, either producing visual artifacts due to information loss within tokenizers or failing to converge fully due to expensive computation costs. We argue that this dilemma stems from the inherent difficulty in learning unconstrained high-dimensional latent spaces. To address this, we propose aligning the latent space with pre-trained vision foundation models when training the visual tokenizers. Our proposed VA-VAE (Vision foundation model Aligned Variational AutoEncoder) significantly expands the reconstruction-generation frontier of latent diffusion models, enabling faster convergence of Diffusion Transformers (DiT) in high-dimensional latent spaces. To exploit the full potential of VA-VAE, we build an enhanced DiT baseline with improved training strategies and architecture designs, termed LightningDiT. The integrated system achieves state-of-the-art (SOTA) performance on ImageNet 256x256 generation with an FID score of 1.35 while demonstrating remarkable training efficiency by reaching an FID score of 2.11 in just 64 epochs--representing an over 21 times convergence speedup compared to the original DiT. Models and codes are available at: https://github.com/hustvl/LightningDiT.

Existing vector quantization (VQ) methods struggle with scalability, largely attributed to the instability of the codebook that undergoes partial updates during training. The codebook is prone to collapse as utilization decreases, due to the progressively widening distribution gap between non-activated codes and visual features. To solve the problem, we propose Index Backpropagation Quantization (IBQ), a new VQ method for the joint optimization of all codebook embeddings and the visual encoder. Applying a straight-through estimator on the one-hot categorical distribution between the encoded feature and codebook, all codes are differentiable and maintain a consistent latent space with the visual encoder. IBQ enables scalable training of visual tokenizers and, for the first time, achieves a large-scale codebook ($2^{18}$) with high dimension ($256$) and high utilization. Experiments on the standard ImageNet benchmark demonstrate the scalability and superiority of IBQ, achieving competitive results on both reconstruction ($1.00$ rFID) and autoregressive visual generation ($2.05$ gFID). The code and models are available at https://github.com/TencentARC/SEED-Voken.

Figure 1: Generations from DnD-Transformers trained on class-conditional ImageNet256 256 (a.top) and unconditional arXiv images (a.bottom). Unconditional rich-text image generations by trained diffusion (b.1) and autoregressive model (b.2), This work tackles the information loss bottleneck of vector-quantization (VQ) autoregressive image generation by introducing a novel model architecture called the 2-Dimensional Autoregression (DnD) Transformer. The DnD-Transformer predicts more codes for an image by introducing a new autoregression direction, model depth, along with the sequence length direction. Compared to traditional 1D autoregression and previous work utilizing similar 2D image decomposition such as RQ-Transformer, the DnD-Transformer is an end-to-end model that can generate higher quality images with the same backbone model size and sequence length, opening a new optimization perspective for autoregressive image generation. Furthermore, our experiments reveal that the DnD-Transformer's potential extends beyond generating natural images. It can even generate images with rich text and graphical elements in a self-supervised manner, demonstrating an understanding of these combined modalities. This has not been previously demonstrated for popular vision generative models such as diffusion models, showing a spark of vision-language intelligence when trained solely on images. The field of autoregressive (AR) image generation is experiencing a resurgence of interest, largely driven by groundbreaking advancements in large language models (LLMs), exemplified by the release of ChatGPT (OpenAI, 2022). Because typical AR image generation methods also predict output in a next-token prediction manner, this resemblance has sparked significant efforts in two main areas: 1) transferring advanced, large-scale training techniques and expertise from LLMs to AR image generation models (Bai et al., 2023; Tian et al., 2024; Sun et al., 2024), and 2) developing truly multimodal foundation models capable of both understanding and generating multimodal information within a unified training framework (Lu et al., 2022; 2023; Team, 2024).

We present Open-MAGVIT2, a family of auto-regressive image generation models ranging from 300M to 1.5B. The Open-MAGVIT2 project produces an open-source replication of Google's MAGVIT-v2 tokenizer, a tokenizer with a super-large codebook (i.e., $2^{18}$ codes), and achieves the state-of-the-art reconstruction performance (1.17 rFID) on ImageNet $256 \times 256$. Furthermore, we explore its application in plain auto-regressive models and validate scalability properties. To assist auto-regressive models in predicting with a super-large vocabulary, we factorize it into two sub-vocabulary of different sizes by asymmetric token factorization, and further introduce "next sub-token prediction" to enhance sub-token interaction for better generation quality. We release all models and codes to foster innovation and creativity in the field of auto-regressive visual generation.

The tokenizer, as one of the fundamental components of large models, has long been overlooked or even misunderstood in visual tasks. One key factor of the great comprehension power of the large language model is that natural language tokenizers utilize meaningful words or subwords as the basic elements of language. In contrast, mainstream visual tokenizers, represented by patch-based methods such as Patch Embed, rely on meaningless rectangular patches as basic elements of vision, which cannot serve as effectively as words or subwords in language. Starting from the essence of the tokenizer, we defined semantically independent regions (SIRs) for vision. We designed a simple HOmogeneous visual tOKenizer: HOOK. HOOK mainly consists of two modules: the Object Perception Module (OPM) and the Object Vectorization Module (OVM). To achieve homogeneity, the OPM splits the image into 4*4 pixel seeds and then utilizes the attention mechanism to perceive SIRs. The OVM employs cross-attention to merge seeds within the same SIR. To achieve adaptability, the OVM defines a variable number of learnable vectors as cross-attention queries, allowing for the adjustment of token quantity. We conducted experiments on the NWPU-RESISC45, WHU-RS19 classification dataset, and GID5 segmentation dataset for sparse and dense tasks. The results demonstrate that the visual tokens obtained by HOOK correspond to individual objects, which demonstrates homogeneity. HOOK outperformed Patch Embed by 6\% and 10\% in the two tasks and achieved state-of-the-art performance compared to the baselines used for comparison. Compared to Patch Embed, which requires more than one hundred tokens for one image, HOOK requires only 6 and 8 tokens for sparse and dense tasks, respectively, resulting in efficiency improvements of 1.5 to 2.8 times. The code is available at https://github.com/GeoX-Lab/Hook.

In this paper, we propose \textbf{UniCode}, a novel approach within the domain of multimodal large language models (MLLMs) that learns a unified codebook to efficiently tokenize visual, text, and potentially other types of signals. This innovation addresses a critical limitation in existing MLLMs: their reliance on a text-only codebook, which restricts MLLM's ability to generate images and texts in a multimodal context. Towards this end, we propose a language-driven iterative training paradigm, coupled with an in-context pre-training task we term ``image decompression'', enabling our model to interpret compressed visual data and generate high-quality images.The unified codebook empowers our model to extend visual instruction tuning to non-linguistic generation tasks. Moreover, UniCode is adaptable to diverse stacked quantization approaches in order to compress visual signals into a more compact token representation. Despite using significantly fewer parameters and less data during training, Unicode demonstrates promising capabilities in visual reconstruction and generation. It also achieves performances comparable to leading MLLMs across a spectrum of VQA benchmarks.