Paper Sharing | CHEERS: A Simpler, More Efficient Route to Unified Multimodal Models
Over the past few years, multimodal models have made great strides along two separate paths: understanding, with steady gains in image question answering and visual reasoning, and generation, with continuous breakthroughs in visual quality and instruction following. A natural question follows: can a single model do both well? That is exactly what Unified Multimodal Models (UMMs) try to answer. But it is far from easy. Understanding relies on stable, abstract semantic representations, while generation must preserve ample local detail and high-frequency texture — the two have an inherent tension in visual representation and optimization objectives.
Today we introduce an open-source work from Tsinghua University, Xi'an Jiaotong University, and the University of Chinese Academy of Sciences: CHEERS, an architecture path toward unified multimodal understanding and generation. While keeping the system simple, it unifies understanding and generation in a single end-to-end framework for joint optimization and inherits as much existing knowledge from open-source pre-trained models as possible. CHEERS achieves:
An end-to-end framework for unified multimodal understanding and generation
A simple architecture with joint optimization of understanding and generation
Zero extra pre-training cost, fully inheriting knowledge from open-source pre-trained models (only a lightweight VAE and a Cascaded Flow Matching Head are added to a conventional VLM)
SOTA or leading performance at the same scale with less training data
4× token compression, combining unified modeling with high efficiency
Three Technical Routes for Unified Multimodal Models
From the perspective of visual representation, existing work falls into three typical approaches:
Separate: Understanding and generation live in relatively separate visual spaces, each optimized without interfering with the other. Task performance is usually decent, but unification is limited.
Single semantic space: One shared representation supports both understanding and generation, but structural detail often becomes a bottleneck.
Heterogeneous fusion: Semantic and pixel-level information are combined, but interference and tug-of-war often follow.
CHEERS positions itself between these routes with a more restrained answer: it does not try to compress everything into one representation, nor does it go for a fully separate dual system. Instead, it rearranges the responsibilities of semantics and detail within a unified framework.
CHEERS Architecture: Unification Through a Lightweight Increment
2.1 Overall architecture: a lightweight upgrade from an understanding model to a unified model
CHEERS starts from a clear premise: rather than introducing a huge, complex composite system for the sake of unification, it keeps existing open-source pre-trained capabilities and makes the smallest possible architectural addition to upgrade an "understanding model" into a "unified understanding + generation model." Specifically, CHEERS builds a unified multimodal large-model framework that brings multimodal understanding and image generation into one end-to-end pipeline through a unified visual tokenizer, an LLM backbone, and a Cascaded Flow Matching Head.
The design's most important value lies in two points:
Joint optimization of understanding and generation: The same LLM backbone serves text autoregression and multimodal understanding as well as conditional modeling for image generation, so unification is no longer "module assembly" but truly end-to-end joint modeling.
Full inheritance of pre-trained knowledge: CHEERS does not need another expensive round of large-scale pre-training just to "unify"; it fully retains and uses existing pre-trained capabilities, significantly lowering the cost of building a unified multimodal model. This matters especially for the open-source community, because a route that can be continuously reproduced, extended, and used must not only be strong but also practical enough.
2.2 Semantics and detail: decoupled, not severed; cascaded generation
It is worth stressing that "decoupling" itself is not the point of this work. In fact, several recent unified multimodal efforts have observed the same lesson: understanding leans toward stable semantics, generation depends more on detail fidelity. The field is gradually seeing the problem clearly. In CHEERS, visual information is organized into two complementary components: semantic tokens for multimodal understanding and generation conditioning, and detail residuals that fill in high-frequency texture and local fidelity during generation.
Correspondingly, in the generation stage CHEERS uses a "semantics first, detail later" cascade: it first generates a global semantic layout, then gradually injects detail through semantic gating to correct and enhance local texture. The authors also observed that even without explicit supervision of the high-frequency detail injection strength, the model naturally increases its use of high-frequency detail late in generation. This is much like how humans paint — structure first, then content, then texture — and shows that CHEERS is not mechanically stacking modules but modeling in a way that better matches the respective rhythms of understanding and generation.
Small Data, Big Performance: What Matters More Is a Unified Feature Space
Experiments show that CHEERS delivers strong overall performance among unified multimodal models of the same scale. On several mainstream understanding and generation benchmarks, it is competitive or even leading:
0.78 on the GenEval generation benchmark
Solid results on the MMBench, MMStar, AI2D, MathVista, and other understanding benchmarks
CHEERS also achieves 4× token compression, offering a more efficient unified modeling approach for high-resolution visual understanding and generation.
▎ More noteworthy: efficient inheritance of pre-trained knowledge
Beyond the performance numbers, there is another highlight: CHEERS reaches these results with significantly less data than some comparable methods. Its total training set is 83M samples, roughly 2× or more less data than some peers, while still achieving SOTA or leading performance at the same scale. This shows that CHEERS's advantage is not merely that it "can be trained," but that it inherits and uses existing pre-trained knowledge more efficiently. In a sense, this matters more than a higher score on any single benchmark: in unified multimodal research today, what is truly scarce is no longer data scale, but how to organize existing knowledge more efficiently.
Conclusion
The authors see CHEERS's significance not only in its results but also in the directions it offers for further thought in unified multimodal research.
1. What a unified model really needs to unify may not be a single visual representation, but an information interface stable and efficient enough. If different tasks inherently need different visual information, then rather than forcing all capabilities to share one representation, it may be more important to let different information enter the unified backbone in the right way.
2. Understanding and generation do not necessarily drag each other down; the key is whether the architecture is designed well. The paper shows that under a unified architecture, jointly training the generation objective does not significantly hurt understanding and can even bring gains in fine-grained perception.
3. Efficient unified multimodality does not have to mean heavier, bigger, and more complex. With simple system changes, CHEERS upgrades a conventional understanding-oriented VLM into a unified model, offering a more realistic reference path for many future open-source efforts.
4. Efficiency is itself part of the unified multimodal problem. CHEERS's 4× token compression is not just an engineering optimization; it means high-resolution understanding and generation can both fit into one system under a more realistic compute budget, which is crucial for unified modeling of longer contexts and more complex visual inputs in the future.
The authors say they want more than a stronger model: they want a route that more researchers and developers can keep advancing. That is why they have open-sourced the training, inference, and evaluation code and model weights, and recently open-sourced the fine-tuning data.
➤ Related Links
Project:
🔗 https://github.com/AI9Stars/Cheers
Model:
🔗 https://huggingface.co/ai9stars/Cheers
Paper:
🔗 https://arxiv.org/abs/2603.12793