Self conditioning boosts deep visual reasoning in flow matching models
Self-Conditioning Enables Deep Sequential Reasoning in Flow Matching
Computer Vision and Pattern RecognitionMachine Learning
Summary
It can be hard for computer models to solve puzzles by creating pictures step by step. The researchers found that the usual method of flow matching doesn’t really become smarter just by taking more steps. They discovered that using a trick called self-conditioning—where the model remembers and builds on its previous guess—lets the model think more deeply and improve its answers with more steps. This finding helps explain how to design better AI systems for solving visual problems.
What this means in practice
- •For computer vision engineers: Improve visual reasoning AI by integrating self-conditioning into flow matching samplers to achieve more accurate and deeper puzzle-solving.
- •For machine learning platform developers: Develop training tools that incorporate self-conditioning to enable multi-step reasoning in generative image models for complex visual tasks.
Authors
Zeyang Zhang, Linxiang Gao, Meiqi Gu, Xingyan Chen, Minrui Luo, Kaifeng Lyu, Tianxing He
Abstract
Recent work has explored visual reasoning through image generation, where models solve visual-reasoning problems by directly generating solution images. Intuitively, iterative samplers such as flow matching are well suited to this task, since each sampling step may resemble one reasoning step. However, in this work, from a circuit complexity perspective, we show that this intuition is misleading. We prove that as long as the model outputs satisfy certain mild regularity conditions, the trajectory of the vanilla flow-matching sampler can be simulated by a constant number of sequential rounds of parallel model evaluations. Consequently, additional sampling steps cannot increase the expressive power of a constant-depth backbone model under the standard complexity-theoretic model of transformers. We then reveal that self-conditioning, which conditions the model on the previous step's estimate of the final image, plays a previously unrecognized role in removing this bottleneck. Controlled experiments on diverse visual-reasoning tasks show that flow matching with self-conditioning achieves higher accuracy and greater reasoning depth than vanilla flow matching, and its performance improves with more sampling steps. Theoretically, we prove that with self-conditioning, a constant-depth backbone can achieve expressive power equivalent to that of space-bounded Turing machines. Our study suggests a design principle for future visual reasoning research: carry more information between sampling steps, so that additional steps translate into deeper reasoning.