TL;DR
MAGIC-Flow is a novel multiscale normalizing flow model that unifies generative and classification tasks, providing interpretability, controllability, and improved performance in challenging domains like medical imaging.
Contribution
It introduces MAGIC-Flow, a hierarchical invertible architecture that enables exact likelihood computation, visualization, and joint generation and classification within a single framework.
Findings
Creates realistic, diverse samples across datasets
Improves classification accuracy in noisy and modality-specific scenarios
Supports controllable sample synthesis and class-probability estimation
Abstract
Generative modeling has emerged as a powerful paradigm for representation learning, but its direct applicability to challenging fields like medical imaging remains limited: mere generation, without task alignment, fails to provide a robust foundation for clinical use. We propose MAGIC-Flow, a conditional multiscale normalizing flow architecture that performs generation and classification within a single modular framework. The model is built as a hierarchy of invertible and differentiable bijections, where the Jacobian determinant factorizes across sub-transformations. We show how this ensures exact likelihood computation and stable optimization, while invertibility enables explicit visualization of sample likelihoods, providing an interpretable lens into the model's reasoning. By conditioning on class labels, MAGIC-Flow supports controllable sample synthesis and principled…
Peer Reviews
Decision·Submitted to ICLR 2026
I find this paper to be technically strong and clearly presented. The authors address an important gap between generative and discriminative modeling in medical imaging by proposing an invertible architecture that unifies both within a single framework. This is an elegant idea, as existing flow-based methods typically require separate models or auxiliary classifiers. The paper provides a solid theoretical foundation showing that Jacobian factorization extend naturally to the conditional setting
While the paper and the idea of unifying generation and classification in a conditional flow is appealing, I find the degree of novelty to be moderate. The architecture largely builds upon established components from RealNVP and Glow, and while the conditional and task-specific extensions are thoughtfully designed, they remain incremental rather than groundbreaking. The main conceptual contribution is more in integration and adaptation than in introducing fundamentally new flow mechanisms. From
1. Unified and practical design: The method uses the same invertible network for both generation and classification, leveraging exact likelihood computation for both. This is a conceptually clean and theoretically sound approach. 2. The method's performance is demonstrated on a challenging, domain-specific problem in medical imaging.
1. Novelty concerns: “Unified generation + classification” sounds incremental relative to prior conditional flows. The paper claims the first conditional multiscale flow that supports both generation and classification on a shared invertible backbone. Yet the Related Work already notes conditional normalizing flows (cNFs), cINNs, and CAFLOW as frameworks for modeling p(x|y)and using label-conditioned transformations, which naturally enable likelihood-based decisions. The manuscript does not clea
1. The paper proposes a normalizing flow model that unifies image generation and classification within a single invertible architecture, demonstrating flexibility across tasks. 2. The use of likelihood attribution maps offers a theoretically grounded, faithful interpretability mechanism that is intrinsic to the model, rather than relying on gradient approximations. 3. The model achieves stable training and competitive quantitative results on scanner- and modality-conditioned generation tasks c
1. The paper mainly compares with GAN-based methods and lacks comparison with more recent state-of-the-art approaches, such as 3D medical latent diffusion models. 2. Evaluation should be extended to more recent and representative datasets (e.g., BraTS2021) to better demonstrate generalizability. 3. The paper presents qualitative visualization of likelihood attribution maps, which are common in flow-matching works; however, further clarification is needed on how these maps relate to classificat
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
