Phrugsa Limbunlom

DeltaFlow is a PyTorch library I authored for radiograph generative pretraining. It provides composable primitives for two ideas that are usually bundled together in ad-hoc research code like flow matching, a simulation-free objective that regresses a velocity field onto the conditional velocity of a probability path (Lipman et al., 2023), and delta alignment, a multi-scale loss that aligns the guidance-difference feature across augmented views instead of the raw, anatomy-entangled features.

\[ \Delta h = h_{\text{cond}} - h_{\text{uncond}}, \qquad \mathcal{L} = \lambda_{\text{flow}}\,\mathcal{L}_{\text{FM}} + \lambda_{\text{align}}\,\mathcal{L}_{\Delta} \]

The name captures the mechanism: Delta is the guidance-difference feature that cancels shared anatomy, and Flow is the flow-matching engine that consumes it. The library targets 2D radiography (chest X-ray, cephalometric, hand), but the core modules including interpolants, samplers, losses, and models.projector are domain-agnostic and reusable for any flow-matching workflow.

Where discriminative models learn a decision boundary \(P(Y\mid X)\), generative modelling learns the data distribution \(P(X)\), or the joint \(P(X,Y)=P(X\mid Y)\,P(Y)\), in order to synthesise new samples. Classifier-free guidance trains a single model to handle both unconditional and conditional denoising, then interpolates the two score estimates at inference with a guidance scale \(w\):

\[ \tilde{\varepsilon} = \varepsilon_{\text{uncond}} + w\,\bigl(\varepsilon_{\text{cond}} - \varepsilon_{\text{uncond}}\bigr) \]

Higher \(w\) pushes samples deeper into the conditional distribution, giving stronger condition alignment at the cost of diversity. The animation traces reverse diffusion on an 8-mode Gaussian mixture: from pure noise at \(t=1.0\), the unconditional model wanders while each class-conditioned model collapses toward its target mode.