Chuyang Zhao, Yifei Song, Hongfa Wang, Jianlong Yuan, Yuan Zhang, Siming Fu, Zhineng Chen, Huilin Deng, Haoyang Huang, Nan Duan
5 min
Abstract
We propose Perceptual Flow Matching (PFM), a simple yet effective framework for few-step generation in flow-matching models. Rather than performing velocity regression in the conventional VAE latent space, PFM supervises flow matching in a perceptual feature space using pretrained perceptual models. This simple change substantially improves the few-step generation capability of flow-matching models, reducing the number of sampling steps from 35-50 to 4-8 while preserving generation quality. Unlike existing acceleration and distillation approaches, PFM requires neither teacher models nor auxiliary score networks and can be integrated into standard flow-matching training pipelines with minimal modifications. Extensive experiments on image generation, video generation, and image editing tasks demonstrate that PFM consistently produces high-quality results while producing fewer artifacts than existing distillation-based methods. We further show that perceptual supervision shifts the regression minimizer from mean-seeking to mode-seeking, biasing predictions toward on-manifold modes that remain accurate under coarse few-step integration. Our results reveal that standard flow-matching training can naturally yield high-quality few-step generators when supervised in an appropriate representation space. We hope this insight inspires future research into representation-aware objectives for efficient generative modeling.
Sam: Does this require rebuilding the whole training pipeline from scratch?
Alex: That's one of the more practical points the paper makes. It doesn't. PFM uses the same underlying structure as standard flow matching—the same architecture, the same training loop. The only change is what the model is being scored against. It's a targeted modification rather than a complete redesign.
Sam: There's another bottleneck I've heard about with these models. Usually, to get a good result, you have to run the model twice—once following your prompt, and once ignoring it—and then combine those outputs. That doubles the computation at the end.
Alex: Right, and PFM addresses that too, through what the authors call "baking." The idea is to move that extra work from the end of the process into the training phase. During training, the model is shown both the guided and unguided versions of a result, and it learns to produce the guided output directly, without needing to run both passes separately later.
Sam: Like the difference between a musician who needs sheet music every time they perform, versus one who's practised enough to play from memory.
Alex: Exactly. And there's an interesting consequence here. Because PFM is already pushing the model toward sharp, committed predictions, the authors found that the need for that extra guidance pass is significantly reduced in the first place. The two improvements reinforce each other.
Sam: So you're solving the same underlying problem from two directions at once—better training signals and less redundant work at the end.
Alex: That's a fair summary. And it points to something worth understanding about why this works geometrically. Think of all possible sharp, realistic images as a landscape of valid destinations. A blurry, averaged image doesn't sit on that landscape—it falls into a valley between valid points. Standard training doesn't penalise that heavily. But the perceptual loss treats those off-landscape positions as very costly, so the model learns to stay on solid ground.
Sam: So the "manifold"—the collection of all valid sharp images—isn't just a metaphor. It's actually shaping where the model is allowed to go during training.
Alex: Precisely. Anything outside that collection looks expensive to the model, so it learns to avoid those blurry in-between states. The geometry of the prediction space turns out to matter just as much as the data itself.
Sam: It's a meaningful shift in perspective. The model isn't being told to work harder—it's being given a better definition of what success looks like.
Alex: And that's the core contribution the paper is making. By changing the supervision space—the lens through which the model evaluates its own output—you can achieve faster, sharper results without adding architectural complexity. Sometimes the most effective improvement isn't a bigger system. It's a clearer goal.
Sam: That's a satisfying place to land. Thanks for walking through it.
Alex: Thanks for listening to ResearchPod.