Representation Forcing for Bottleneck-Free Unified Multimodal Models
This work addresses the structural bottleneck of external VAEs in unified multimodal models, offering a path towards more end-to-end architectures for the machine learning community.
This paper introduces Representation Forcing (RF), a technique that enables unified multimodal models (UMMs) to perform image generation directly in pixel space without relying on external VAEs. RF achieves this by forcing the decoder to autoregressively predict visual representations as intermediate tokens, which then guide pixel diffusion. This approach allows their pixel-space model to match state-of-the-art VAE-based unified models in image generation and generally outperform VAE-based variants in image understanding.
Unified multimodal models (UMMs) aim to handle perception and generation in a single model. Yet existing UMMs still rely on a frozen, separately pretrained VAE for image generation, imposing a structural bottleneck. Naively removing it introduces a quality gap, as the model must learn both high-level structure and low-level details from raw pixels. In this paper, we propose Representation Forcing (RF), a technique that closes this gap by making representation prediction a native capability of the model. Concretely, RF forces the decoder to autoregressively predict visual representations as intermediate tokens before pixels; these tokens then stay in context to guide pixel diffusion within the same backbone. By turning representations from perception outputs into generation targets, RF eliminates the need for any external generative latent space. We find that RF benefits both understanding and generation. On image generation, our pixel-space model with RF matches state-of-the-art VAE-based unified models. On image understanding, pixel-space RF generally outperforms its VAE-based variant. Together, these results offer an effective step toward end-to-end, bottleneck-free UMMs.