Selftok: Discrete Visual Tokens of Autoregression, by Diffusion, and for Reasoning

Selftok Team

Media Technology Institute, Huawei Singapore

Our preliminary work has been accepted as an oral presentation at CVPR 2025. You can find more details at the following link.

Abstract

We completely discard the conventional spatial prior in image representation and introduce a novel discrete visual tokenizer: Self-Consistency Tokenizer (Selftok). At its design core, we compose an autoregressive (AR) prior—mirroring the causal structure of language—into visual tokens by using the reverse diffusion process of image generation. The AR property makes Selftok fundamentally distinct from traditional spatial tokens in the following two key ways:

Selftok offers an elegant and minimalist approach to unify diffusion and AR for vision-language models: By representing images with Selftok tokens, we can train vision-language models (VLMs) using a purely discrete autoregressive architecture—like that in LLMs—without requiring additional modules or training objectives.
We theoretically show that the AR prior satisfies the Bellman equation, whereas the spatial prior does not. Therefore, Selftok supports reinforcement learning (RL) for visual generation with effectiveness comparable to that achieved in LLMs.

Besides the AR property, Selftok is also a SOTA tokenizer that achieves both high-quality reconstruction and high compression bit rate. After representing the training images as Selftok tokens, as a pure AR model, our VLM achieves both SOTA visual comprehension and generation performances. Impressively, without using any text-image training pairs, a simple policy gradient RL working in the visual tokens can significantly boost the visual generation benchmark, surpassing all the existing models by a large margin.

Therefore, we believe that Selftok effectively addresses the long-standing challenge that visual tokens cannot support effective RL. When combined with the well-established strengths of RL in LLMs, this brings us one step closer to realizing a truly multimodal LLM.

Introduction

Why Discrete?

Comparison of pure discrete autoregressive model (dAR) and hybrid dAR/cAR model — Comparison of (a) pure discrete autoregressive model (dAR) and (b) hybrid model that combines dAR and continuous autoregressive model (cAR). $\texttt{<BOS>}/\texttt{<BOV>}$ indicates the start of a sentence/image. $w_i/v_i$ denotes the $i$ -th language/visual token. Both models predict the next token given all previous ones, *e.g.*, $[\texttt{<BOS>}, \dots, \texttt{<$v_3$>}] \to \texttt{<$v_4$>}$ .

Why Not Spatial?

Illustration of spurious dependencies in spatial tokens and anti-causal links — (a) As an image can be viewed as the effect of spatial pixels (or patches), observing any part of it introduces spurious dependencies among spatial tokens, making them non-AR. (b) Due to the anti-causal links (red) for spatial tokens, learning the locally optimal policy $\pi^*_4$ for a later token (*e.g.*, $v_4$ ) can propagate backward and interfere with earlier tokens that were already optimized (*e.g.*, $v_1,v_2,v_3$ ). In contrast, AR tokens without such links do not have this issue.

Progressive Reconstruction & Interpolation

Illustration of progressive reconstruction and interpolation — Progressive reconstruction (left to right): Reconstructions by progressively masking out a shorter sequence of tokens before inputting to the decoder. Interpolation (left to right): Reconstructions by gradually replacing tokens of the left image with those of the right one. All methods except Selftok exhibit strong spatial characteristics *i.e.*, tokens $\Leftrightarrow$ patches.

Selftok: Self-Consistency Tokenizer

Architecture

Main Results

Reconstruction performance of different tokenizers on - resolution ImageNet 50k validation set. ^† Results from the original paper.
Tokenizer	Type	PSNR↑	SSIM↑	LPIPS↓
LlamaGen	2D
MaskBiT^†	2D			-
Cosmos	2D
VQGAN-LC^†	2D
OpenMagViT-V2	2D
ViT-VQGAN^†	2D	-	-	-
LlamaGen	2D
Cosmos	2D
VAR	2D
TiTok-L-32	1D
TiTok-B-64	1D
TiTok-S-128	1D
FlexTok	1D
FlowMo-Lo^†	1D
FlowMo-Hi^†	1D
Selftok (Ours)	1D
Selftok (Ours)	1D

Evaluation of text-to-image generation ability on GenEval benchmark. Janus-Pro-7B† represents the result of our evaluation. Janus-Pro-7B-Zero represents a model that has undergone the same visual RL process as Selftok-Pre-Zero and Selftok-Zero. † represents the result of our evaluation. ‡ GPT-4o results are from OpenAI’s technical report.
Type	Method	Single Obj.	Two Obj.	Counting	Colors	Position	Color Attr.	Overall
Diffusion Only	PixArt-α
	SDXL
	FLUX.1-dev
	DALL-E 3
	SD3-Medium
	CogView4-6B
	HiDream-I1
Hybrid Model	SEED-X
	Transfusion
	D-DiT
	Show-o
	GPT-4o‡
Pure dAR	Emu3-Gen>
	TokenFlow-XL
	ILLUME+
	Infinity
	Janus-Pro-7B
	Janus-Pro-7B†
	Janus-Pro-7B-Zero
	Selftok-Pre
	Sefltok-Pre-Zero
	Selftok-SFT
	Selftok-Zero

Performances on DPG-Bench. The methods in this table are all generation-specific models except Show-o, Janus-Pro, and Selftok. † represents the result of our evaluation.
Type	Method	Global	Entity	Attribute	Relation	Other	Overall
Diffusion Only	PixArt-α
	SDXL
	DALL-E 3
	SD3-Medium
	FLUX.1-dev
	CogView4-6B
	HiDream-I1
Hybrid Model	Show-o
Pure dAR	Emu3-Gen
	Janus
	Infinity
	Janus-Pro-7B
	Janus-Pro-7B†
	Janus-Pro-7B-Zero
	Selftok-Pre
	Selftok-SFT
	Selftok-Zero

Examples

Text-to-Image Generation

Text-to-Image generation results by Selftok using the text prompts of DPG-Bench

Selftok-based visual RL

Single-turn Image Editing

Multi-turn Image Editing