ContinualFlow: Unified Flow-Based Adaptation

• ContinualFlow is a suite of frameworks and algorithms that enable continual adaptation by leveraging occlusion-aware flow estimation and temporal continuity.
• It employs a two-stage refinement network with explicit occlusion masks and warped previous flow, reducing endpoint errors significantly on benchmarks like KITTI’15 and MPI-Sintel.
• In generative unlearning, ContinualFlow uses energy-reweighted flow matching to target data suppression, yielding competitive forget rates and retention accuracy compared to full retraining.

ContinualFlow encompasses a family of frameworks and algorithms across vision and generative modeling centered on continual adaptation—either in flow-based visual learning, optical flow estimation, or neural flow-based unlearning. The term arises in three distinct research threads: (1) temporal optical flow estimation with occlusion awareness, (2) continual representation learning via motion-conjugated flows, and (3) targeted unlearning of generative models using neural flow matching. The following delineates the principal methodologies, theoretical constructs, and empirical results associated with the term "ContinualFlow."

1. ContinualFlow for Occlusion-aware Optical Flow Estimation

ContinualFlow (Neoral et al., 2018) defines a state-of-the-art approach to optical flow estimation under occlusions and frame continuity. It introduces two core advances:

• Occlusion-first Estimation: Occlusion masks are predicted preceding the flow estimation, leveraging a light convolutional subnetwork over the cost volume. The occlusion mask is then concatenated with the cost volume and used as an explicit input for flow decoding, allowing the model to avoid corrupted matches in occluded areas and guide flow extrapolation.
• Temporal Continuity: The estimated flow from the previous frame pair is warped (forward and backward) and supplied as an auxiliary input for both occlusion and flow decoders. This enables the model to propagate motion context across arbitrary temporal windows and use prior flow as an informed guess within occluded regions.

This architecture is built atop PWC-Net’s cost-volume, pyramid-based structure and introduces a specialized two-stage refinement network for fine-grained detail recovery. The training protocol involves a three-stage curriculum: pre-training on FlyingChairs, enabling occlusion and temporal inputs in FlyingThings3D, and final tuning with mixed real and synthetic datasets (KITTI’15, MPI-Sintel, etc.), with learning rates and data batching meticulously scheduled.

Empirical results demonstrate:

• >25% reduction in end-point error (EPE) on both KITTI’15 and MPI-Sintel benchmarks through occlusion integration alone.
• Additional improvements of 18% (KITTI) and 7% (Sintel) via temporal continuity.
• Top performance on benchmark leaderboards: 1st on Sintel (Final pass, EPE=4.528 px), 3rd on KITTI’15 with 10.03% Fl error.
• Notably superior flow performance in occluded and motion-boundary regions compared to all contemporaries as of publication (Neoral et al., 2018).

2. ContinualFlow for Targeted Unlearning in Generative Models

ContinualFlow (Simone et al., 23 Jun 2025) also denotes a general method for targeted unlearning—removal of specific data content from continuous-time generative models—using neural flow matching. In this context, learning and unlearning are unified as the transport of distribution mass by a neural velocity field, but the loss is modulated with energy-based weighting to softly subtract undesired data regions.

The fundamental constructs are as follows:

• Let $q_0$ be a base distribution (e.g., Gaussian) and $q_\text{full}$ the learned data distribution. The forget set $q_f$ is characterized indirectly via an energy proxy $F(x)\propto-\log q_f(x)$, such as the log-odds output of a binary classifier distinguishing forget content. The retained data $q_r$ and suppression factor $\lambda$ control the degree and sharpness of unlearning.
• The unnormalized reweighted density is

$$\tilde R(x) = q_0(x)\,\sigma(-\lambda\,F(x)),\quad \sigma(z) = \frac{1}{1+e^{-z}},$$

and the normalized mass-subtracted target is

$$\tilde q_1(x) = \frac{q_0(x)\,\sigma(-\lambda F(x))}{\int q_0(y)\,\sigma(-\lambda F(y))\,dy}.$$

• Training employs the Energy-Reweighted Flow Matching (ERFM) loss:

$$\mathcal{L}_{\text{ERFM}}(\theta) = \mathbb{E}_{x_0, x_1 \sim q_0;\; t \sim U[0,1];\; x \sim p_t(x | x_0, x_1)}\left[ w(x_1)\,\Vert v_\theta(t, x) - (x_1-x_0)\Vert^2 \right],$$

where $w(x_1) = \sigma(-\lambda F(x_1))$ is the energy-derived importance weight.

• Theoretical guarantees show ERFM gradients are proportional to standard flow matching gradients targeting $q_\text{full}$0, rendering the method equivalent to true FM toward the soft mass-subtracted distribution.

Algorithmic instantiation is direct: pairs ($q_\text{full}$1, $q_\text{full}$2) are sampled from $q_\text{full}$3, interpolated at random $q_\text{full}$4, and weighted by $q_\text{full}$5; all regression is on the velocity field $q_\text{full}$6. For images, practitioners often operate ERFM in autoencoding latent space.

Empirical assessments cover: 2D synthetic data (ring, moons), unlearning odd digits in MNIST, and class-specific suppression in CIFAR-10 latent spaces. On MNIST, for instance, the forget rate is reduced to 0.05% (ContinualFlow) compared to 1.4% with naive fine-tuning, approaching retraining-from-scratch baseline performance in MMD and retention accuracy. The method achieves efficient, interpretable, and reversible unlearning without retraining from scratch or direct exposure to forget-set samples (Simone et al., 23 Jun 2025).

3. Methodological Details and Theoretical Properties

ContinualFlow for Vision

• Occlusion Decoder: Five-layer convolutional network with progressively decreasing channel count, attached to cost volume at each pyramid level. Outputs per-pixel probabilities of occlusion.
• Temporal Input Design: Combination of warped previous flow (forward/backward), validity masks, and concatenation into both occlusion and flow decoder inputs.
• Loss Function: Unified loss over all pyramid levels:

$q_\text{full}$7

with scale weight $q_\text{full}$8 and $q_\text{full}$9 balancing terms.

ContinualFlow for Unlearning

• Energy Proxy Calibration: $q_f$0 can be derived from a classifier as $q_f$1, with $q_f$2 the predicted probability of belonging to the forget set.
• Suppression Factor: The choice of $q_f$3 governs the trade-off between retention accuracy and effective forgetting; large $q_f$4 yields near-hard deletion, while small $q_f$5 induces soft suppression.
• Modularity: The framework supports compositional unlearning via additive energy proxies, but the stability and ordering of sequential unlearning steps remain open questions.

4. Empirical Results

Dataset	Method	MMD$q_f$6	Acc$q_f$7	Forget Rate$q_f$8	Leakage$q_f$9	Time (s)
2D	Retrain (GT)	0.0157 ± 0.0104	0.9999 ± 0.0002	0.0007 ± 0.0007	0.0255 ± 0.0342	64.3 ± 1.3
	Fine-tune	0.0157 ± 0.0104	0.9999 ± 0.0002	0.0007 ± 0.0007	0.0255 ± 0.0342	10.4 ± 0.8
	ContinualFlow	0.0162 ± 0.0136	0.9999 ± 0.0002	0.0155 ± 0.0107	0.0385 ± 0.0375	50.1 ± 1.6
MNIST	Retrain (GT)	0.0004 ± 0.0000	0.9861 ± 0.0098	0.0050 ± 0.0012	0.0108 ± 0.0009	300 ± 15
	Fine-tune	0.0039 ± 0.0004	0.9551 ± 0.0167	0.0143 ± 0.0032	0.0214 ± 0.0028	92.9 ± 6.1
	ContinualFlow	0.0020 ± 0.0003	0.9673 ± 0.0153	0.0005 ± 0.0005	0.0015 ± 0.0003	158.7 ± 11.3
CIFAR-10	Retrain (GT)	0.0056 ± 0.0004	0.8920 ± 0.0000	0.1127 ± 0.0078	0.1546 ± 0.0073	802 ± 292
	Fine-tune	0.0077 ± 0.0016	0.9005 ± 0.0068	0.2157 ± 0.0095	0.2401 ± 0.0065	253 ± 19
	ContinualFlow	0.0064 ± 0.0005	0.8847 ± 0.0077	0.1704 ± 0.0125	0.1748 ± 0.0109	427 ± 35

This table indicates that ContinualFlow achieves unlearning performance competitive with full-retraining baseline in terms of MMD and accuracy, with improved forget-rate suppression over fine-tuning (Simone et al., 23 Jun 2025).

5. Limitations and Open Research Directions

• Occlusion estimation and flow extrapolation rely on the reliability of cost-volume features and temporal priors; errors in alignment or propagation may degrade performance, especially under abrupt scene changes (Neoral et al., 2018).
• Generative unlearning effectiveness depends critically on the quality and calibration of the energy proxy $F(x)\propto-\log q_f(x)$0. Poorly calibrated or noisy classifiers can induce over/under-suppression, and the choice of suppression parameter $F(x)\propto-\log q_f(x)$1 is task-dependent and currently requires manual tuning (Simone et al., 23 Jun 2025).
• ContinualFlow for generative models is presently applied in latent space for high-dimensional data; extension to pixel-level flow remains a challenging and open problem.
• Theoretical guarantees of compositional unlearning, stability under repeated or hierarchical applications, and privacy properties are identified as future work.

Possible research avenues include joint energy-function and flow-field learning, adaptive $F(x)\propto-\log q_f(x)$2 scheduling, robustness under repeated unlearning, integration with privacy risk mitigation, and generalization to broader classes of continuous-time models such as diffusion processes.

6. Context within the Broader Literature

ContinualFlow, as instantiated in optical flow estimation (Neoral et al., 2018), advances temporal reasoning and explicit occlusion modeling, surpassing prior approaches that lack explicit occlusion priors or temporal regularization. The flow-matching-based ContinualFlow framework (Simone et al., 23 Jun 2025) generalizes modular unlearning, offering theoretical and practical advantages over ad-hoc sample deletion or model retraining in generative systems. Collectively, the ContinualFlow methodologies exemplify the potential of integrating temporal structure, energy-based loss modulation, and flow-based representation learning to address persistent challenges in vision and machine learning.