Balanced Conic Rectified Flow

• Balanced Conic Rectified Flow is a high-dimensional generative modeling approach that balances synthetic and real data using conic rectification to reduce drift and computational demands.
• It introduces a conic rectification mechanism with spherical linear interpolation (Slerp) to generate straighter, more stable trajectories and enhance model fidelity.
• Empirical results on CIFAR-10 show significant improvements in FID, reconstruction error, and efficiency, achieving high-quality generation with far fewer synthetic samples.

Balanced Conic Rectified Flow (BCRF) is a methodology for training generative ordinary differential equation (ODE) models that improves upon existing rectified flow frameworks by efficiently leveraging both synthetic and real data. BCRF introduces a conic rectification mechanism and a balanced alternation strategy between synthetic and real-pair reflow, resulting in straighter generative paths, lower computational requirements, and superior fidelity in high-dimensional generative modeling tasks. The approach addresses core limitations of previous rectified flows, including drift from the data distribution and inefficiency induced by overuse of synthetic pairs, by introducing geometric supervision anchored around real data inversions and spherically perturbed conic neighborhoods (Seong et al., 29 Oct 2025).

1. Generative ODE Framework and Rectified Flow

Generative flows model data synthesis as the solution to an ODE transporting an initial distribution $p_0$ (usually Gaussian noise) to a complex target distribution $p_1$. The system is defined as: $$\frac{dx_t}{dt} = v_\theta(x_t, t), \quad x_0 \sim p_0, \; x_1 \sim p_1$$ where $v_\theta$ is a neural network parameterizing the velocity field. In 1-rectified flow ($k=1$), training proceeds by minimizing: $$\mathcal{L}_{\rm fake} = \mathbb{E}_{X_0 \sim p_0,\; X_1 \sim p_1,\, t \sim \mathrm{Exp}([0,1])} \left\| X_1 - X_0 - v_\theta(t X_1 + (1-t) X_0,\, t)\right\|^2$$ This objective enforces that the learned velocity field reproduces the straight-line paths joining synthetic sample pairs. Iterating this training using reflow ($k$ times) yields increasingly rectified flows, but at the cost of requiring large numbers of generated data pairs.

2. Conic Rectification and Real-Pair Injections

A primary issue in traditional reflow is the accumulation of drift from the target distribution $p_1$ and poor generalization to the structure of real data. To anchor the learning process, BCRF introduces real sample inversions $Z_{0,R} = v_\theta^{-1}(X_1)$ for $X_1 \sim p_1$. Around each inversion, random Gaussian perturbations $\epsilon$ are applied, and a spherical linear interpolation (Slerp) is conducted: $$\mathrm{Slerp}(Z_{0,R}, \epsilon, \zeta) = \frac{\sin((1-\zeta)\phi)}{\sin\phi}Z_{0,R} + \frac{\sin(\zeta \phi)}{\sin\phi}\epsilon, \quad \phi = \arccos \langle Z_{0,R}, \epsilon \rangle$$ The resulting conic path at time $t$ is: $$\mathrm{Conic}(X_1, \epsilon, \zeta, t) = t X_1 + (1-t)\mathrm{Slerp}(Z_{0,R}, \epsilon, \zeta)$$ This forms a conic (Editor’s term: “supervision cone”) of trajectories reaching towards $X_1$, providing stability and local geometric anchoring for the flow.

3. Balanced Reflow Objectives and Training Algorithm

BCRF alternates between two training objectives within each minibatch:

• Fake-pair reflow:

$$\mathcal{L}_{\rm fake} = \mathbb{E}_{(Z_{0,F}, Z_{1,F}), t} \left\| Z_{1,F} - Z_{0,F} - v_\theta(t Z_{1,F} + (1-t) Z_{0,F}, t)\right\|^2$$

where $Z_{0,F}$ is drawn from $p_0$ and $Z_{1,F} = v_\theta(Z_{0,F})$.

• Conic real-pair reflow:

$$\mathcal{L}_{\rm real} = \int_0^1 \mathbb{E}_{X_1, \epsilon, \zeta} \left\| X_1 - \mathrm{Slerp}(Z_{0,R}, \epsilon, \zeta) - v_\theta(\mathrm{Conic}(X_1, \epsilon, \zeta, t),\, t) \right\|^2 dt$$

The balanced training objective alternates between $\chi_{\rm fake}$ and $\chi_{\rm real}$, with corresponding sampling ratios $U_{\rm real}\!:\!U_{\rm fake}$ and hyperparameters such as the Slerp schedule $\zeta(t)$ and the exponential distribution over $t$.

The iterative reflow procedure involves initializing $v_\theta$ on fake pairs, then alternating between real and fake reflow steps, as specified in pseudocode within the source.

4. Sample Efficiency and Empirical Results

A defining feature of BCRF is its substantial reduction in computational requirements due to enhanced sample efficiency. BCRF achieves superior or comparable generation quality with an order of magnitude fewer synthetic examples than prior approaches.

CIFAR-10 Results (2-Rectified Flow)

Model	FID	IS	Curv.	IVD	ReconΔ	p-ReconΔ
Original	12.21	8.08	0.00284	0.29508	0.03367	0.03548
BCRF (full)	5.98	8.79	0.00230	0.25333	0.01940	0.02238
Fixed real-pair	6.69	8.59	0.00231	0.24244	0.02023	0.02289
No Slerp	6.60	8.57	0.00232	0.24088	0.02338	0.02606

• In the one-step Euler generation regime, BCRF with distillation achieves IS = 8.79 and FID = 5.98, halving the FID compared to the original (12.21) while using 300 K fake + 60 K real pairs, versus 4 M fake-only pairs.
• For full ODE integration (RK45), BCRF improves FID from 3.36 to 3.24 with slightly fewer function evaluations.
• Curvature and initial velocity delta decrease by approximately 20%, indicating straighter and more consistent learned flows.
• The reconstruction error gap (real vs.\ fake) present in baseline rectified flows collapses when using BCRF.

5. Theoretical Considerations and Empirical Validation

BCRF introduces no new formal theorems or convergence proofs. Empirical findings demonstrate two key issues in plain rectified flows: (1) the accumulation of KL divergence from $p_1$, and (2) a growing gap between the reconstruction error of real and fake samples. The conic real-pair reflow mechanism shrinks these errors and maintains geometric fidelity to the real data manifold. The scheduling of hyperparameters, especially $\zeta_{\max}$, is currently determined by heuristics aimed at maximizing the real–fake reconstruction gap after warm-up. Investigations into optimization grounded in manifold dimension or information-theoretic criteria remain open problems.

6. Applications and Future Directions

BCRF is broadly compatible with diffusion-based rectified flows such as EDM, SD3, and InstaFlow, as well as models employing discriminator-guided flows. It is suitable for a wide range of applications, including image-to-image translation, conditional generation, and inverse problems using manifold-aware reflow. Future work includes formalizing convergence via optimal transport or ODE stability theory, further optimizing the selection of the Slerp perturbation schedule, integrating geometric or adversarial regularization strategies within the conic supervision, and extending the method to additional domains where geometric fidelity and sample efficiency are paramount (Seong et al., 29 Oct 2025).